Read Anytime Anywhere Easy Ebook Downloads at ebookmeta.

com

Intelligent Computing Proceedings of the 2021

Computing Conference Volume 3 Lecture Notes in
Networks and Systems 285 Kohei Arai (Editor)

https://ebookmeta.com/product/intelligent-computing-
proceedings-of-the-2021-computing-conference-
volume-3-lecture-notes-in-networks-and-systems-285-kohei-
arai-editor/

OR CLICK HERE

DOWLOAD EBOOK

Visit and Get More Ebook Downloads Instantly at https://ebookmeta.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Intelligent Computing Proceedings of the 2021 Computing

Conference Volume 2 Lecture Notes in Networks and Systems
284 Kohei Arai (Editor)
https://ebookmeta.com/product/intelligent-computing-proceedings-of-
the-2021-computing-conference-volume-2-lecture-notes-in-networks-and-
systems-284-kohei-arai-editor/
ebookmeta.com

Intelligent Computing: Proceedings of the 2021 Computing

Conference, Volume 1 1st Edition Kohei Arai

https://ebookmeta.com/product/intelligent-computing-proceedings-of-
the-2021-computing-conference-volume-1-1st-edition-kohei-arai/

ebookmeta.com

Intelligent Systems and Applications Proceedings of the

2021 Intelligent Systems Conference IntelliSys Volume 1
294 Lecture Notes in Networks and Systems Kohei Arai
(Editor)
https://ebookmeta.com/product/intelligent-systems-and-applications-
proceedings-of-the-2021-intelligent-systems-conference-intellisys-
volume-1-294-lecture-notes-in-networks-and-systems-kohei-arai-editor/
ebookmeta.com

All You Knead Is Love 3rd Edition Guerrero Tanya

https://ebookmeta.com/product/all-you-knead-is-love-3rd-edition-
guerrero-tanya/

ebookmeta.com
Digital Anthropology Second Edition Haidy Geismar (Editor)

https://ebookmeta.com/product/digital-anthropology-second-edition-
haidy-geismar-editor/

ebookmeta.com

Unmasking Theatre Design A Designer s Guide to Finding

Inspiration and Cultivating Creativity 1st Edition Lynne
Porter
https://ebookmeta.com/product/unmasking-theatre-design-a-designer-s-
guide-to-finding-inspiration-and-cultivating-creativity-1st-edition-
lynne-porter/
ebookmeta.com

Cherishing Her: A mc biker instalove romance (Skin Sins

Tattoo Shop Book 1) 1st Edition Jailaa West

https://ebookmeta.com/product/cherishing-her-a-mc-biker-instalove-
romance-skin-sins-tattoo-shop-book-1-1st-edition-jailaa-west/

ebookmeta.com

Patent Pending in 24 Hours 8th Edition Richard Stim

https://ebookmeta.com/product/patent-pending-in-24-hours-8th-edition-
richard-stim/

ebookmeta.com

Building Behavior The Educator s Guide to Evidence Based

Initiatives 1st Edition Jessica Hannigan

https://ebookmeta.com/product/building-behavior-the-educator-s-guide-
to-evidence-based-initiatives-1st-edition-jessica-hannigan/

ebookmeta.com
Theories of International Relations 6th Edition Scott
Burchill

https://ebookmeta.com/product/theories-of-international-relations-6th-
edition-scott-burchill/

ebookmeta.com
Lecture Notes in Networks and Systems 285

Kohei Arai Editor

Intelligent
Computing
Proceedings of the 2021 Computing
Conference, Volume 3
Lecture Notes in Networks and Systems

Volume 285

Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland

Advisory Editors
Fernando Gomide, Department of Computer Engineering and Automation—DCA,
School of Electrical and Computer Engineering—FEEC, University of Campinas—
UNICAMP, São Paulo, Brazil
Okyay Kaynak, Department of Electrical and Electronic Engineering,
Bogazici University, Istanbul, Turkey
Derong Liu, Department of Electrical and Computer Engineering, University
of Illinois at Chicago, Chicago, USA; Institute of Automation, Chinese Academy
of Sciences, Beijing, China
Witold Pedrycz, Department of Electrical and Computer Engineering,
University of Alberta, Alberta, Canada; Systems Research Institute,
Polish Academy of Sciences, Warsaw, Poland
Marios M. Polycarpou, Department of Electrical and Computer Engineering,
KIOS Research Center for Intelligent Systems and Networks, University of Cyprus,
Nicosia, Cyprus
Imre J. Rudas, Óbuda University, Budapest, Hungary
Jun Wang, Department of Computer Science, City University of Hong Kong,
Kowloon, Hong Kong
The series “Lecture Notes in Networks and Systems” publishes the latest
developments in Networks and Systems—quickly, informally and with high quality.
Original research reported in proceedings and post-proceedings represents the core
of LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as new
challenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks,
spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor
Networks, Control Systems, Energy Systems, Automotive Systems, Biological
Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems,
Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems,
Robotics, Social Systems, Economic Systems and other. Of particular value to both
the contributors and the readership are the short publication timeframe and the
world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.
The series covers the theory, applications, and perspectives on the state of the art
and future developments relevant to systems and networks, decision making, control,
complex processes and related areas, as embedded in the fields of interdisciplinary
and applied sciences, engineering, computer science, physics, economics, social, and
life sciences, as well as the paradigms and methodologies behind them.
Indexed by SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago.
All books published in the series are submitted for consideration in Web of
Science.

More information about this series at http://www.springer.com/series/15179

Kohei Arai
Editor

Intelligent Computing
Proceedings of the 2021 Computing
Conference, Volume 3

123
Editor
Kohei Arai
Faculty of Science and Engineering
Saga University
Saga, Japan

ISSN 2367-3370 ISSN 2367-3389 (electronic)

Lecture Notes in Networks and Systems
ISBN 978-3-030-80128-1 ISBN 978-3-030-80129-8 (eBook)
https://doi.org/10.1007/978-3-030-80129-8
© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Editor’s Preface

It is a great privilege for us to present the proceedings of the Computing Conference

2021, held virtually on July 15 and 16, 2021.
The conference is held every year to make it an ideal platform for researchers to
share views, experiences and information with their peers working all around the
world. This is done by offering plenty of networking opportunities to meet and
interact with the world-leading scientists, engineers and researchers as well as
industrial partners in all aspects of computer science and its applications.
The main conference brings a strong program of papers, posters, videos, all in
single-track sessions and invited talks to stimulate significant contemplation and
discussions. These talks were also anticipated to pique the interest of the entire
computing audience by their thought-provoking claims which were streamed live
during the conferences. Moreover, all authors had very professionally presented
their research papers which were viewed by a large international audience online.
The proceedings for this edition consist of 235 chapters selected out of a total of
638 submissions from 50+ countries. All submissions underwent a double-blind
peer-review process. The published proceedings has been divided into three vol-
umes covering a wide range of conference topics, such as technology trends,
computing, intelligent systems, machine vision, security, communication, elec-
tronics and e-learning to name a few.
Deep appreciation goes to the keynote speakers for sharing their knowledge and
expertise with us and to all the authors who have spent the time and effort to
contribute significantly to this conference. We are also indebted to the organizing
committee for their great efforts in ensuring the successful implementation of the
conference. In particular, we would like to thank the technical committee for their
constructive and enlightening reviews on the manuscripts in the limited timescale.
We hope that all the participants and the interested readers benefit scientifically
from this book and find it stimulating in the process.

v
vi Editor’s Preface

Hope to see you in 2022, in our next Computing Conference, with the same
amplitude, focus and determination.

Kohei Arai
Contents

Feature Analysis for Aphasic or Abnormal Language Caused

by Injury . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Marisol Roldán-Palacios and Aurelio López-López
Automatic Detection and Segmentation of Liver Tumors
in Computed Tomography Images: Methods and Limitations . . . . . . . . 17
Odai S. Salman and Ran Klein
HEMAYAH: A Proposed Mobile Application System for Tracking
and Preventing COVID-19 Outbreaks in Saudi Arabia . . . . . . . . . . . . . 36
Reem Alwashmi, Nesreen Alharbi, Aliaa Alabdali, and Arwa Mashat
Data Security in Health Systems: Case of Cameroon . . . . . . . . . . . . . . . 48
Igor Godefroy Kouam Kamdem and Marcellin Julius Antonio Nkenlifack
Predicting Levels of Depression and Anxiety in People
with Neurodegenerative Memory Complaints Presenting with
Confounding Symptoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Dalia Attas, Bahman Mirheidari, Daniel Blackburn, Annalena Venneri,
Traci Walker, Kirsty Harkness, Markus Reuber, Chris Blackmore,
and Heidi Christensen
Towards Collecting Big Data for Remote Photoplethysmography . . . . . 70
Konstantin Kalinin, Yuriy Mironenko, Mikhail Kopeliovich,
and Mikhail Petrushan
Towards Digital Twins Driven Breast Cancer Detection . . . . . . . . . . . . 87
Safa Meraghni, Khaled Benaggoune, Zeina Al Masry,
Labib Sadek Terrissa, Christine Devalland, and Noureddine Zerhouni
Towards the Localisation of Lesions in Diabetic Retinopathy . . . . . . . . 100
Samuel Ofosu Mensah, Bubacarr Bah, and Willie Brink

vii
viii Contents

Framework for a DLT Based COVID-19 Passport . . . . . . . . . . . . . . . . 108

Sarang Chaudhari, Michael Clear, Philip Bradish, and Hitesh Tewari
Deep Learning Causal Attributions of Breast Cancer . . . . . . . . . . . . . . 124
Daqing Chen, Laureta Hajderanj, Sarah Mallet, Pierre Camenen, Bo Li,
Hao Ren, and Erlong Zhao
Effect of Adaptive Histogram Equalization of Orthopedic
Radiograph Images on the Accuracy of Implant Delineation
Using Active Contours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Alicja Smolikowska, Paweł Kamiński, Rafał Obuchowicz,
and Adam Piórkowski
Identification Diseases Using Apriori Algorithm on DevOps . . . . . . . . . 145
Aya Mohamed Morsy and Mostafa Abdel Azim Mostafa
IoT for Diabetics: A User Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Signe Marie Cleveland and Moutaz Haddara
Case Study: UML Framework of Obesity Control Health
Information System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Majed Alkhusaili and Kalim Qureshi
An IoT Based Epilepsy Monitoring Model . . . . . . . . . . . . . . . . . . . . . . . 192
S. A. McHale and E. Pereira
Biostatistics in Biomedicine and Informatics . . . . . . . . . . . . . . . . . . . . . 208
Esther Pearson
Neural Network Compression Framework for Fast Model Inference . . . 213
Alexander Kozlov, Ivan Lazarevich, Vasily Shamporov,
Nikolay Lyalyushkin, and Yury Gorbachev
Small-World Propensity in Developmental Dyslexia After Visual
Training Intervention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Tihomir Taskov and Juliana Dushanova
An Assessment of Extreme Learning Machine Model for Estimation
of Flow Variables in Curved Irrigation Channels . . . . . . . . . . . . . . . . . 259
Hossein Bonakdari, Azadeh Gholami, Bahram Gharabaghi, Isa Ebtehaj,
and Ali Akbar Akhtari
Benchmarking of Fully Connected and Convolutional Neural
Networks on Face Recognition: Case of Makeup and Occlusions . . . . . . 270
Stanislav Selitskiy, Nikolaos Christou, and Natalya Selitskaya
Wavelet Neural Network Model with Time-Frequency Analysis
for Accurate Share Prices Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
Yaqing Luo
Contents ix

A Hybrid Method for Training Convolutional Neural Networks . . . . . . 298

Vasco Lopes and Paulo Fazendeiro
A Layered Framework for Dealing with Dilemmas in Artificial
Intelligence Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Jim Q. Chen
Improving Student Engagement and Active Learning with
Embedded Automated Self-assessment Quizzes: Case Study
in Computer System Architecture Design . . . . . . . . . . . . . . . . . . . . . . . . 327
Ryan M. Gibson and Gordon Morison
Cognitive Load of Ontology as a Means of Information
Representation in the Educational Process . . . . . . . . . . . . . . . . . . . . . . . 340
Maryna Popova and Rina Novogrudska
The Influence of Computer Games on High School
Students (Adolescents) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Dalibor Peić and Andrija Bernik
Pedagogical Approaches in Computational Thinking-Integrated
STEAM Learning Settings: A Literature Review . . . . . . . . . . . . . . . . . . 369
Ashok Kumar Veerasamy, Peter Larsson, Mikko-Ville Apiola,
Daryl D’Souza, and Mikko-Jussi Laakso
Development of Anaglyph 3D Functionality for Cost-Effective
Virtual Reality Anatomical Education . . . . . . . . . . . . . . . . . . . . . . . . . . 390
Alex J. Deakyne, Thomas Valenzuela, and Paul A. Iaizzo
Augmented Reality in Virtual Classroom for Higher Education
During COVID-19 Pandemic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Monica Maiti, M. Priyaadharshini, and B. Vinayaga Sundaram
The Teaching and Learning of the Untyped Lambda Calculus
Through Web-Based e-Learning Tools . . . . . . . . . . . . . . . . . . . . . . . . . . 419
Levis Zerpa
Development and Design of a Library Information System Intended
for Automation of Processes in Higher Education Institution . . . . . . . . . 437
Askar Boranbayev, Ruslan Baidyussenov, and Mikhail Mazhitov
Informally Teaching Black Youth STEM Concepts Virtually Using
Artificial Intelligence and Machine Learning . . . . . . . . . . . . . . . . . . . . . 446
Darron Lamkin, Robin Ghosh, Tutaleni I. Asino, and Tor A. Kwembe
Accrediting Artificial Intelligence Programs from the Omani
and the International ABET Perspectives . . . . . . . . . . . . . . . . . . . . . . . . 462
Osama A. Marzouk
x Contents

Embodied Learning: Capitalizing on Predictive Processing . . . . . . . . . . 475

Susana Sanchez
Exploring the Darkness of Gamification: You Want It Darker? . . . . . . 491
Tobias Nyström
The Future of Information and Communication
Technology Courses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
Dalize van Heerden and Leila Goosen
The Impact of Computer Games on Preschool Children’s
Cognitive Skills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
Dolores Kamenar Čokor and Andrija Bernik
Using STEM to Improve Computer Programming Teaching . . . . . . . . . 542
Manuel Rodrigues
Tips for Using Open Educational Resources to Reduce Textbook
Costs in Computing and IT Courses . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
Richard Halstead-Nussloch and Rebecca Rutherfoord
Cross-Cultural Design of Simplified Facial Expressions . . . . . . . . . . . . . 560
Meina Tawaki, Keiko Yamamoto, and Ichi Kanaya
Misinformation and Its Stakeholders in Europe:
A Web-Based Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
Emmanouil Koulas, Marios Anthopoulos, Sotiria Grammenou,
Christos Kaimakamis, Konstantinos Kousaris, Fotini-Rafailia Panavou,
Orestis Piskioulis, Syed Iftikhar Hussain Shah, and Vassilios Peristeras
Artificial Life: Investigations About a Universal Osmotic
Paradigm (UOP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
Bianca Tonino-Heiden, Bernhard Heiden, and Volodymyr Alieksieiev
Latency and Throughput Advantage of Leaf-Enforced Quality
of Service in Software-Defined Networking for Large Traffic Flows . . . 606
Josiah Eleazar T. Regencia and William Emmanuel S. Yu
The Impact of Online Social Networking (Social Media)
on Interpersonal Communication and Relationships . . . . . . . . . . . . . . . 624
Mui Joo Tang and Eang Teng Chan
Embedded Piezoelectric Array for Measuring Relative Distributed
Forces on Snow Skis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
Andres Rico, Carson Smuts, Jason Nawyn, and Kent Larson
Experimental Study of a Tethered Balloon Using 5G Antenna
to Enhance Internet Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Samirah A. Alhusayni, Shuruq K. Alsuwat, Shahd H. Altalhi,
Faris A. Almalki, and Hawwaa S. Alzahrani
Contents xi

Incremental Dendritic Cell Algorithm for Intrusion Detection

in Cyber-Physical Production Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 664
Rui Pinto, Gil Gonçalves, Jerker Delsing, and Eduardo Tovar
Performance Comparison Between Deep Learning-Based and
Conventional Cryptographic Distinguishers . . . . . . . . . . . . . . . . . . . . . . 681
Emanuele Bellini and Matteo Rossi
TFHE Parameter Setup for Effective and Error-Free Neural
Network Prediction on Encrypted Data . . . . . . . . . . . . . . . . . . . . . . . . . 702
Jakub Klemsa
Security in Distributed Ledger Technology: An Analysis
of Vulnerabilities and Attack Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
Efthimios-Enias Gojka, Niclas Kannengießer, Benjamin Sturm,
Jan Bartsch, and Ali Sunyaev
An Analysis of Twitter Security and Privacy Using
Memory Forensics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743
Ahmad Ghafarian and Darius Fiallo
A Security Requirement Engineering Case Study: Challenges
and Lessons Learned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 761
Souhaïl El Ghazi El Houssaïni, Ilham Maskani, and Jaouad Boutahar
Towards Automated Surveillance: A Review of Intelligent
Video Surveillance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 784
Romas Vijeikis, Vidas Raudonis, and Gintaras Dervinis
Homicide Prediction Using Sequential Features from Graph
Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
Juan Moreno, Sebastian Quintero, Alvaro Riascos, Luis Gustavo Nonato,
and Cristian Sanchez
Effectiveness and Adoption of NIST Managerial Practices for Cyber
Resilience in Italy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 818
Alessandro Annarelli, Serena Clemente, Fabio Nonino, and Giulia Palombi
Steganographic Method to Data Hiding in RGB Images Using Pixels
Redistribution Combining Fractal-Based Coding and DWT . . . . . . . . . 833
Hector Caballero-Hernandez, Vianney Muñoz-Jimenez,
and Marco A. Ramos
Improving the Imperceptibility of Pixel Value Difference and LSB
Substitution Based Steganography Using Modulo Encoding . . . . . . . . . . 851
Ahmad Whafa Azka Al Azkiyai, Ari Moesriami Barmawi,
and Bambang Ari Wahyudi
xii Contents

Securing Mobile Systems GPS and Camera Functions Using

TrustZone Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 868
Ammar S. Salman and Wenliang (Kevin) Du
Application of Machine Learning in Cryptanalysis Concerning
Algorithms from Symmetric Cryptography . . . . . . . . . . . . . . . . . . . . . . 885
Milena Gjorgjievska Perusheska, Vesna Dimitrova,
Aleksandra Popovska-Mitrovikj, and Stefan Andonov
A Response-Based Cryptography Engine in Distributed-Memory . . . . . 904
Christopher Philabaum, Christopher Coffey, Bertrand Cambou,
and Michael Gowanlock
Elliptic Curves of Nearly Prime Order . . . . . . . . . . . . . . . . . . . . . . . . . . 923
Daniele Di Tullio and Manoj Gyawali
Assessing Small Institutions’ Cyber Security Awareness Using
Human Aspects of Information Security Questionnaire (HAIS-Q) . . . . . 933
Glenn Papp and Petter Lovaas
Construction of Differentially Private Empirical Distributions from
a Low-Order Marginals Set Through Solving Linear Equations
with l2 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 949
Evercita C. Eugenio and Fang Liu
Secondary Use Prevention in Large-Scale Data Lakes . . . . . . . . . . . . . . 967
Shizra Sultan and Christian D. Jensen
Cybercrime in the Context of COVID-19 . . . . . . . . . . . . . . . . . . . . . . . . 986
Mohamed Chawki
Beware of Unknown Areas to Notify Adversaries: Detecting
Dynamic Binary Instrumentation Runtimes with Low-Level
Memory Scanning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1003
Federico Palmaro and Luisa Franchina
Tamper Sensitive Ternary ReRAM-Based PUFs . . . . . . . . . . . . . . . . . . 1020
Bertrand Cambou and Ying-Chen Chen
Locating the Perpetrator: Industry Perspectives of Cellebrite
Education and Roles of GIS Data in Cybersecurity and
Digital Forensics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1041
Denise Dragos and Suzanna Schmeelk
Authentication Mechanisms and Classification: A Literature Survey . . . 1051
Ivaylo Chenchev, Adelina Aleksieva-Petrova, and Milen Petrov
ARCSECURE: Centralized Hub for Securing a Network
of IoT Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1071
Kavinga Yapa Abeywardena, A. M. I. S. Abeykoon,
A. M. S. P. B. Atapattu, H. N. Jayawardhane, and C. N. Samarasekara
Contents xiii

Entropy Based Feature Pooling in Speech Command Classification . . . . 1083

Christoforos Nalmpantis, Lazaros Vrysis, Danai Vlachava,
Lefteris Papageorgiou, and Dimitris Vrakas

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1093

 Feature Analysis for Aphasic
or Abnormal Language Caused by Injury

Marisol Roldán-Palacios and Aurelio López-López(B)

Instituto Nacional de Astrofı́sica, Óptica y Electrónica, Luis Enrique Erro No. 1,

Sta. Marı́a Tonantzintla, 72840 Puebla, México
{marppalacios,allopez}@inaoep.mx
https://ccc.inaoep.mx/

Abstract. The interest of researching alterations in aphasic or abnor-

mal language has maintained and keeps growing, following a diversity of
approaches. Particularly, the period corresponding to the post-traumatic
recovery stage in impaired language caused by a traumatic brain injury
has been scarcely explored. The findings reported here specify the steps
followed during the inspection of a lexical feature set with the objective
of determining which contribute most to describe TBI language singu-
larities throughout the first two stages of recovery (after three and six
months). For this purpose, we employ language technologies, statistical
analysis, and machine learning. Starting with a 25-feature set, after con-
ducting an analytical process, we attained three selected sets with fewer
attributes, having comparable efficacy measures to those obtained from
the complete feature set taken as baseline.

Keywords: TBI-abnormal language · Feature selection · Language

technologies · Statistical learning · Machine learning

1 Introduction

Language impairment occurs from different illnesses, having a variety of causes.

In this work, we focus on the analysis of abnormal language caused by a trau-
matic brain injury (TBI). According to [15], this language disorder can not be
explained with the characteristics, or at least part of them, describing apha-
sic language cases from other causes [1,7]. Furthermore, we concentrate on the
post-traumatic recovery period which has been barely studied.
With a computational linguistic approach, a more complex analysis can be
achieved, which allows supporting or discarding a hypothesis in less time, as
well as to extend the studies regarding the characteristics describing the TBI-
abnormal language. Efforts have addressed the identification of proper metrics
solidly defining the deficiencies in the language. These try to open new possibil-
ities to be applied in therapy, to have an improvement in recovery.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

K. Arai (Ed.): Intelligent Computing, LNNS 285, pp. 1–16, 2021.
https://doi.org/10.1007/978-3-030-80129-8_1
2 M. Roldán-Palacios and A. López-López

Among the contributions of the study here reported are the following. Some
core TBI-language feature sets were identified, leading to similar efficacy values
of learning algorithms to those achieved when evaluating with the complete ini-
tial TBI-attribute set with a variety of schemes. An analytical exploration based
in correlation in the starting feature set under analysis is presented. Moreover,
a possible relation between impurity quality as a convenient evaluator directly
grading, or together with learning algorithms based on it, was found.
The paper is organized as follows. Section 2 discusses briefly some of the
relevant results reached in previous investigations, the methodologies they tried
and introduced, as well as, potential language features to describe the subject
of study in turn. Section 3 is summing up the steps carried out to complete the
present analysis. Section 4 is subdivided to include specific information about
the Corpus (4.1), implementation (4.2), preliminary results (4.3), ending the
section, and the work, with a discussion of the results presented (4.4), and some
additional observations (4.5).

2 Related Work
After Linguistics researchers started extensive investigations in this problem de-
cades ago, mainly doing their analysis and computation by hand, computational
techniques have been gradually incorporated. For instance, the application of
natural language processing (NLP) of [12] examines disabilities in the language
of a heterogeneous group of participants with different diagnostics, having a
variety of causes. A language analysis was completed based on transcripts created
by automatic speech recognition (ASR) in which three sets of characteristics were
extracted, from a Linguistic Content Analysis (LCA), Part-of-Speech (POS)
tags, and Linguistic Inquiry and Word Count (LIWC) processing.
Each of the above mentioned techniques provides a variety of characteris-
tics. The first set contributes a quantitative view of the word pattern usage.
Then from the next group, for instance, nouns by verb frequencies are extracted.
Among the categories that LIWC provides are psychology processes, linguistic
dimensions (articles, negations, and so on), relative to time and space from which
also frequencies are obtained. After ranking characteristics through an analysis
of variance (ANOVA), they were explored with varied learning algorithms in
WEKA, such as SimpleLogistic, J48 (decision tree), and Multilayered Percep-
tron (neural networks) [12].
On the same line, further investigations were conducted [4–6, 13]. Such works
incorporated statistical criteria to qualify features combined with ablation exper-
iments [5, 6]. Also, some efforts [13] were done to avoid overfitting occurring
in earlier investigations (given the limited size of datasets), and tried some
other learning algorithms, such as Naı̈ve Bayes, Support Vector Machine, besides
Logistic Regression.
Observing that the data entry quality is a cornerstone in the complexity of
these types of studies, a knowledge model [11] was introduced to work on the
feature filtering based on an ACT-R [10] architecture. Furthermore, language
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 3

models (LMs) were employed in [3, 14] which were trained on Part-Of-Speech
tag sequences. The selected resulting feature sets were tested with SVM [3, 11],
Naı̈ve Bayes [3, 11], Decision Trees [11], Neural Networks [3, 11], and Boosting
with Decision Stumps [3]. These last two works are focused on discriminating
specific language impairment (SLI) among infants or teenagers.

3 Methodology

The methodology behind our findings is illustrated in Fig. 1, and consists of the
following phases. After extracting the initial set of characteristics proper to the
language samples, a set of valuators are incorporated to analyze the features,
leading to a ranking. Parallel to this analysis, root nodes are generated and
selected employing a tree-based algorithm, to compare both criteria and then
identify the promising candidates feature sets. For evaluating these promising
features sets, a phase of wrapping which involves setting up of learning algo-
rithms, training and testing steps. According to the assessment of the results,
we can go back to valuator selection or to consider a new feature set.

Fig. 1. Methodology scheme.

4 M. Roldán-Palacios and A. López-López

4 Experiments
Next, we report details of the analysis conducted in our reported work. The
Corpus source, considered tasks, employed tools, feature set, steps, and achieved
results.

4.1 Data Set

Our samples were extracted from information collected as part of a long-term
research [15] called “Communication recovery after TBI” led by Prof. Leanne
Togher. For the current analysis, the first two phases of recovery are considered
and our data set consists of more than 40 instances per phase. To complement
our study, an additional set of non-aphasic samples was taken from a different
research [2] led by Coelho, also addressed to analyze language in TBI cases.
Table 1 summarizes details of the data set.

Table 1. Demographics

Male Female AgeRange Task

TBI-03 35 10 17–66 Retelling Cinderella story
TBI-06 36 8 17–66 Retelling Cinderella story
NTBI 31 16 16–63 A story based on The Runaway picture

First two sets described in Table 1 refer to subjects in recovery after suffering
a brain injury that caused some kind of language impairment (e.g. aphasia,
dysarthria, or apraxia), registered after three and six months, respectively. Also,
notice that the transcriptions analyzed in our TBI-study group is the retelling
of Cinderella story, while for our contrastive (negative) set is the narration of a
story based on The Runaway picture, a piece authored by Norman Rockwell.
Transcripts are files created with a special format that synthesizes the whole
information that CLAN [8] requires to generate the different indices available.
They are organized into levels, for example gems (@G), among other, supports
the identification of the task being referred by the lines following the tier. The
tag (*PAR) indicates the participant layer, and that corresponding to the inves-
tigator is marked by (*INV). Also we have the morphem (%mor) and grammar
(%gra) tiers. Next, lines of a short example of the transcripts referred to, are
given.

@G: Cinderella
*PAR: &-um the two wicked stepsisters &-um up to these old tricks.
\%mor: det:art|the det:num|two adj|wicked step#n|sister-PL adv|up
prep|to det:dem|these adj|old n|trick-PL .
\%gra: 1|2|DET 2|4|QUANT 3|4|MOD 4|0|INCROOT 5|4|JCT
6|4|JCT 7|9|DET 8|9|MOD 9|6|POBJ 10|4|PUNCT
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 5

4.2 Further Experimental Details

From CHILDES to Talkbank. The Child Language Data Exchange System

(CHILDES) Project was first conceived in 1981 and materialized in 1984. Under
a continuous growing and development, the CHILDES project started a transfor-
mation in 2001 until becoming Talkbank Project, with a long list of contributors
and representing an investment of thousands of hours in collecting, transcribing
and reviewing data. Among the different corpora included in Talkbank [8], we
focus in Togher Corpus [15].

CLAN. This acronym stands for Computerized Language ANalysis. This is a

free resource available in the Talkbank platform. Among several other functions,
CLAN tool allows to automatically compute a variety of indices oriented to study
language, i.e. in this context, language disorders [8].
The set of features we examine in the present work were extracted employing
CLAN, with the command c-nnla which automatically computes the Northwest-
ern Narrative Language Analysis profile, according to the rules established in
the NNLA manual [16]. This has proved to achieve an accuracy comparable to
those calculate by hand by experts [8]. Nowadays, c-nnla provides more than
forty features, from which we are revising a subset of only twenty five. With this
set of attributes, we are basically covering a wide lexical study. Table 2 shows
the detailed list of attributes on which we focus our analysis.1
For a few of the features included in Table 2, we provide some clarification.
In particular, for Utterance, there is no universal linguistic definition, however,
“phonetically an utterance is a unit of speech bounded by silence”.2 Open-class
notion includes nouns, verbs, adjectives, and adverbs [16], and Close-class cat-
egory encloses auxiliaries, complementizers, conjunctions, determiners, modals,
negation particle, prepositions, pronouns, quantifiers, infinitival marker, and wh-
words [16].

Selecting the More Discriminative Feature Subset. There are two main
approaches [17] regarding an attribute filtering process. When this is done prior
to the learning phase, it is called filter method. On the contrary, when the learning
algorithm in fact encompasses the selection stage, this is called wrapper method.
Usually, they are combined in a cycle of feature nomination, closely related to
the subject under analysis.

Ranking Features. As stated by [17], “there is no universally accepted measure

of relevance, although several different ones have been proposed”. So, we start by
inspecting which attribute or subset evaluator could provide more information
for our purpose.
1
Originally, the ratio Utts/Min was not part of c-nnla profile but we added it to this
set of features.
2
Source: url(https://glossary.sil.org/term/utterance).
6 M. Roldán-Palacios and A. López-López

Table 2. C-NNLA feature subset

FN Feature Description

1 DurationSec Duration in seconds

2 WordsPerMin Words / Min
3 TotalUtts # of Utterances
4 TotalWords # of Words
5 MLUWords Mean Length Utterance
6 UttsPerMin Utterances/Min
7 openClass Open-Class
8 propOpenClassPerWords %open-class/Words
9 closedClass close-class
10 propClosedClassPerWords %close-class/Words
11 PropOpenPerClosed %open-class/close-class
12 Nouns # of Nouns
13 propNounsPerWords %Nouns/Words
14 Verbs # of Verbs
15 propVerbsPerWords %Verbs/Words
16 nounPerverb Nouns/Verbs
17 adj # of Adjectives
18 adv # of Adverbs
19 det # of Determinants
20 pro # of Pronouns
& possessive determiners
21 aux # of auxiliaries
22 conj # of conjunctions
23 complementizers # of complementizers
24 modals # of Modals
25 prep # of Prepositions
Excluding “wh” interrogative pronouns and “wh” relative pronouns.

4.3 Preliminary Results

Starting with one of the most extensively used criteria, i.e. information gain,
the statistics reveals that features propVerbsPerWords and nounPerVerbs, cor-
responding to feature numbers (FN) 15 and 16, are the only important compo-
nents in the group, with average merit values of 0.013 ± 0.039 and 0.013 ± 0.038
respectively. However, we observed that they appear in the last places in the
ranking column.
Additionally, several preliminary data explorations revealed openClass and
closedClass as noisy or non discriminative attribute in this 25-feature set, and
placed on top of the corresponding ranking. In consequence, we opted to try
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 7

a series of different ranking criteria. Most of these rankings had in common

openClass and closedClass features on top, with the exception of ImpurityEuclid,
which guides a part of our next experiments. We will explain a pair of rounds
of the experiments conducted throughout the analysis of the 25-feature CNNLA
set.

Outlier Case. Regarding the TBI language samples, the group of interest, it
is not possible to deal with an outlier conception because we are interested in
recognizing and modeling each alteration that language is presenting. However,
regarding the negative sample information, the notion is plausible, and the con-
tinuation of the data exploration led us to identify an outlier instance in it, given
that the sample was presenting some indexes doubling the next minimum in the
set.
We tried a set of learning algorithms to test the discriminative capability of
the selected feature group. The viability of Naı̈ve Bayes (NB) in a variety of
language studies led us to consider it as one of the first options for the analysis.
Given that Impurity indicator seems to be applicable in our context, it was
natural to consider SPAARC and Random Forest (RF) algorithms. Although
RF works on boosting, we included AdaBoostM1 too, and we considered also a
learning algorithm based on regression, i.e. ClassificationViaRegression.
Based on the observation that impurity metric was applicable to reveal some
discriminating features, and considering that the tree learning algorithm basis
to branch is to reduce the impurity grade in the new nodes created, we pro-
ceed as follows. We trained a tree learning algorithm over a random sample
split of 60% from the 25-CNNLA feature set to then analyze the root node clus-
ters generated. Recognizing potential prospects based on impurity grading along
with their Pearson correlation, we determined a part of the selected feature set.
Table 3 enumerates the selected feature group.

Table 3. Selected feature set.

FN Feature Description

3 TotalUtts # of Utterances
4 TotalWords # of Words
5 MLUWords Mean Length Utterance
13 propNounsPerWords %Nouns/Words
16 nounPerVerb Nouns/Verbs
21 aux # of auxiliaries

Tables 4 and 5 summarize the results for the first two recovery stages samples
once the outlier was removed. Every learning algorithm was applied with a leave-
one-out cross validation, given the reduced number of instances. In order to
8 M. Roldán-Palacios and A. López-López

compare our findings, we evaluate the full 25-CNNLA feature set with the whole
proposed group of learning algorithms and we set as baseline the highest value
of the overall evaluation for each recovery data set.

Table 4. Testing selected feature group once outlier was excluded - TBI03.

Learning Precision F1 Accuracy

Algorithm Recall Measure

Baseline 0.921 0.778 0.843 85.227

Sequential Min. Optimization
Naive Bayes 0.867 0.578 0.693 73.864
Sequential Min. Optimization 0.963 0.578 0.722 77.273
SPAARC 0.829 0.756 0.791 79.545
Random Forest 0.804 0.822 0.813 80.682
AdaBoostM1 0.872 0.756 0.810 81.818
Classification Via Regression 0.889 0.889 0.889 88.636

Table 5. Testing selected feature group - TBI06.

Learning Precision F1 Accuracy

Algorithm Recall Measure

Baseline 0.909 0.952 0.930 92.941

Classification Via Regression
Naive Bayes 0.909 0.952 0.930 92.941
Sequential Min. Optimization 0.833 0.714 0.769 78.823
SPAARC 0.833 0.833 0.833 83.529
Random Forest 0.878 0.857 0.867 87.059
AdaBoostM1 0.902 0.881 0.892 89.412
Classification Via Regression 0.929 0.929 0.929 92.941

Applying Normalization. Among the measure units in the 25-CNNLA fea-

ture set, we found for instance, words, utterances, utterances per minute or
percentages, i.e. they lie in different ranges. In order to analyze this factor, we
carried out a normalization to the CNNLA feature set, to have them in a unified
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 9

interval. The applied transformation, similar to that implemented in [9], is given

by formula 1
vf ∗ 100
v f = n (1)
i=1 vfi
where vf is the vector corresponding to each feature, n is the number of instances,
and f ∈ [1, 25]. This transformation aims to compute the proportion of each
instance value into their own range, using a factor of one hundred to avoid
introducing zero to small values.
Under these conditions, a pair of candidate attribute subsets was obtained,
where Tables 6, and 7 specify the promising selected feature set for the first and
second recovery stages, respectively. Tables 8, and 9 include the results achieved
when tested on the collection, corresponding to the first and second recovery
stages, respectively. Once more, for each learning algorithm tested, a leave-one-
out cross validation was applied.

Table 6. Selected feature set from normalized data - TBI03.

FN Feature Description

6 UttsPerMin Utterances/Min
7 openClass Open-Class .
8 propOpenClassPerWords %open-class/Words
10 propClosedClassPerWords %close-class/Words

Table 7. Selected feature set from normalized data - TBI06.

FN Feature Description

8 propOpenClassPerWords %open-class/Words
10 propClosedClassPerWords %close-class/Words
14 Verbs # of Verbs
16 nounPerVerb Nouns/Verbs
17 adj # of Adjectives
24 modals # of Modals
10 M. Roldán-Palacios and A. López-López

Table 8. Testing selected group from normalized feature set - TBI03.

Learning Precision F1 Accuracy

Algorithm Recall Measure

Baseline 0.813 0.867 0.839 82.954

Random Forest
Naive Bayes 0.714 0.778 0.745 72.727
Sequential Min. Optimization 0.652 0.956 0.775 71.591
SPAARC 0.702 0.733 0.717 70.454
Random Forest 0.875 0.778 0.824 82.954
AdaBoostM1 0.860 0.822 0.841 84.091
Classification Via Regression 0.894 0.933 0.913 90.909

Table 9. Testing selected group from normalized feature set - TBI06.

Learning Precision F1 Accuracy

Algorithm Recall Measure

Baseline 0.837 0.857 0.847 84.706

Random Forest
Naive Bayes 0.600 0.643 0.621 61.176
Sequential Min. Optimization 0.357 0.238 0.286 41.176
SPAARC 0.614 0.643 0.628 62.353
Random Forest 0.854 0.833 0.843 84.705
AdaBoostM1 0.659 0.690 0.674 67.059
Classification Via Regression 0.947 0.857 0.900 90.588

Table 10. Confusion matrix - classification via regression - TBI03

40 5
(True-Positive) (False-Positive)

5 38
(False-Negative) (True-Negative)
Outlier case
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 11

4.4 Discussion of Results

Regarding the first recovery stage after removing the uniquely recognized outlier,
CVR works well with an accuracy of 84.09, similar to that achieved by RF, both
showing a balance in instances correctly discriminated with the full 25-feature
set. However, they were surpassed by SMO with a 85.23 in accuracy, determining
in this way the baseline for this particular experiment.
We identify a selected group of seven features listed in Table 3, based on the
25-CNNLA feature set, achieving with CVR scheme a slightly better efficacy of
88.64 in comparison with that calculated over the full attribute group. Besides,
a balance among instances correctly discriminated is maintained, as Table 10
shows. We have to remark that three of the attributes contained in this cluster,
aux, nounPerVerb, and proNounsPerWords, were positioned in last places when
they were graded according to their InfoGain property.
For the data corresponding to the second recovery stage (i.e. six months),
a high level of accuracy of 92.94 was reached for the basic 25-attribute group
with the CVR approach. With the same 7-feature set (Table 3), it was possible
to have an efficacy as good as that obtained for the established baseline, once
again with the CVR learning algorithm. Additionally, in both settings, with a
steady balance in cases correctly discriminated.
When moving to review the results when a transformation is applied to the
data with the aim of bringing the attribute values to the same scale, the first
point to notice is the drastic reorganization in the bias that each element has into
the set under inspection. That is, proportion indexes that were null or noisy in
previous evaluations, are now positively weighted. Again, repeating the process
on the reshaped 25-CNNLA feature set for each recovery stage, we select not
one but two feature subsets, as listed in Tables 6 and 7.
For the initial stage of recovery (i.e. three months), we have a reduced
4-feature set in which three of the four elements are noisy or non discriminative
for the unscaled set, with UttsPerMin as the only one not being in that role. The
overall accuracy was computed and the baseline was established to 82.95, this
time with Random Forest which steadily achieves the same assessment to the
4-feature set, however, with a decrement in the balance of instances accurately
discriminated. The highest value of 90.91 is obtained again with CVR and with
a slightly wider margin than in earlier cases, managing a good composition of
samples rightly discriminated as the confusion matrix (2) exhibits.
 
42 3 
  (2)
 5 38

Given that when testing with this last 4-feature set with the second recov-
ery phase data sample produced poor results, a new representative feature set
(Table 7) was explored. This time, it was necessary to incorporate a new grading
test weighting instances in a neighbor based on its inverse square distance, in
reference to the evaluated instance, this along with impurity evaluation guided
the selection of the feature set (Table 7). Notice that this is a new composition
but shares a couple of elements with the 4-feature set (Table 6), and one with
the first group presented (Table 3).
12 M. Roldán-Palacios and A. López-López

Efficacy also decreases in these conditions, the accuracy baseline was deter-
mined to be 84.71, obtained with RF, which was surpassed only by CVR achiev-
ing values of 90.59, but showing a drop in the correctly discriminated samples
balance (3), charged to the negative sample.
 
36 6 
  (3)
 2 41

Regarding the whole feature set F as the union of component subsets (4),
where 0 < n ≤ m and m ≤ size(F), the idea behind finding the more descriptive
feature set, FW (5), consists of uncovering
 the more positively weighted feature,
Fw , of each subset fj . Notice that fk fl = ∅ for k = l.
m

F= fj (4)
j=n

m

FW = fw (5)
w=1

Based on that, we are extending the discussion for the selected 6-feature
set (Table 7) corresponding to the second recovery stage. Figure 2 exhibits the
strong inverse relation existing between attributes propOpenClassPerWord and
propClosedClassPerWords with a diagonal instance distribution having a nega-
tive slope, that indicates in our context that they are quite descriptive features.
Charts in Fig. 3 show the null or partial relation existing among propOpen-
ClassPerWord and the remaining attributes in the selected 6-feature set.

Fig. 2. Correlation propOpenClassPerWord vs. propClosedClassPerWords - TBI06

normalized
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 13

Fig. 3. Correlation propOpenClassPerWord vs. remaining components - TBI06

normalized

Fig. 4. PCA graph - selected normalized 6-feature set TBI06

14 M. Roldán-Palacios and A. López-López

The relations previously described can be quite evident in a principal

component analysis (PCA) graph, such as that of Fig. 4, where correspond-
ing vectors of the 6-feature subset reveal the relationship among them, for
instance, propOpenClassPerWord and propClosedClassPerWords features are
clearly inversely related. Also, certain independence among the remaining
attributes is noticeable, adding evidence to the efficacy of the six feature subset.
It should be noted that some of the features are redistributed in comparison
to the position they have in the starting 25-CNNLA feature set, i.e. it is not
possible to identify them by a direct gaze in the PCA graph corresponding to
the original attribute group.

5 Conclusions
The inherent relationships among features in the sets are complex and are
strongly related to the subject of study, and the larger the feature sets the more
complex the way they behave. In addition, frequently we have to handle het-
erogeneous group of features that convey still more elaborated connections. The
presented work is summarizing the results of an addressed micro-level analysis,
which means, an inspection bounded to a lexical level, a piece of a larger work in
the direction of understanding how TBI-abnormal language evolves during the
post-traumatic recovery stages.
Attaining reasonable efficacy measures with a quarter of the 25-feature set,
we have to keep in mind the stability of the CVR algorithm as a wrapping
method in feature selection and its steadiness, along with that of RF throughout
the whole analysis. In addition to the support that the tree-based algorithm
and ImpurityEuclid analysis, as filter methods, represent to this methodology to
guide some others examinations. Having acceptable efficacy higher or equal to
the baseline does not allow to assert that the reported results are conclusive. It
is evident that there are intrinsic relationship at the interior of the feature set
still unrevealed.
Further research basically aims for extension on three directions: (i) we will
cover a wider range of features than that considered here; (ii) we will expand
the analysis to a macro-level, i.e. incorporating syntactic components, and (iii)
additionally we plan to work on further stages of recovery to understand its
evolution. Besides it, each step unavoidable implies the revision of techniques,
methods, and methodology to better understand the results and the studied
impaired language.

Acknowledgments. The first author was supported by CONACyT, through schol-

arship 1008734. The second author was partially supported by SNI.

References
1. Linnik, A., Bastiaanse, R., Höhle, B.: Discourse production in aphasia: a current
review of theoretical and methodological challenges. Aphasiology 30(7), 765–800
(2015). https://doi.org/10.1080/02687038.2015.1113489
 Feature Analysis for Aphasic or Abnormal Language Caused by Injury 15

2. Coelho, C.A., Grela, B., Corso, M., Gamble, A., Feinn, R.: Microlinguistic deficits
in the narrative discourse of adults with traumatic brain injury. Brain Inj. 19(13),
1139–1145 (2005)
3. Gabani, K., Sherman, M., Solorio, T., Liu, Y., Bedore L.M., Peña, E.D.: A corpus-
based approach for the prediction of language impairment in monolingual English
and Spanish-English bilingual children. Human language technologies. In: The 2009
Annual Conference of the North American Chapter of the ACL, Boulder, Colorado,
June 2009, pp. 46–55 (2009).  c 2009 Association for Computational Linguistics
4. Fraser, K.C., et al.: (2012). Automated classification of primary progressive aphasia
subtypes from narrative speech transcripts. Cortex. 55, 43–60 (2013). https://doi.
org/10.1016/j.cortex.2012.12.006
5. Kathleen, C.F., Hirst, G., Graham, N.L., Meltzer, J.A., Black, S.E., Rochon, E.:
Comparison of different feature sets for identification of variants in progressive
aphasia. Workshop on Computational Linguistics and Clinical Psychology: From
Linguistic Signal to Clinical Reality, Baltimore, Maryland USA, 27 June 2014, pp.
17–26 (2014).  c 2014 Association for Computational Linguistics
6. Fraser, K.C., Hirst, G., Meltzer, J.A., Mack, J.E., Thompson, C.K.: Using statis-
tical parsing to detect agrammatic aphasia. In: Proceedings of the 2014 Workshop
on Biomedical Natural Language Processing (BioNLP 2014), Baltimore, Maryland,
USA, 26–27 June 2014, pp. 134–142 (2014).  c 2014 Association for Computational
Linguistics
7. Bryant, L., Ferguson, A., Spencer, E.: Linguistic analysis of discourse in aphasia: a
review of the literature. Clin. Linguist. Phonetics (2016). https://doi.org/10.3109/
02699206.2016.1145740
8. MacWhinney, B.: The CHILDES Project: Tools for Analyzing Talk, 3rd edn.
Lawrence Erlbaum Associates, Mahwah (2000)
9. Marini, A., Andreetta, S., del Tin, S., Carlomagno, S.: A multi-level approach
to the analysis of narrative language in aphasia. Aphasiology 25(11), 1372–1392
(2011). https://doi.org/10.1080/02687038.2011.584690
10. Oliva, J., Serrano, J.I., Del Castillo, M.D., Iglesias, A.: Cognitive modeling of the
acquisition of a highly inflected verbal system. In: Salvucci, D., Gunzelmann, G.
(eds.) Proceedings of the 10th International Conference on Cognitive Modeling,
pp. 181–186. Drexel University, Philadelphia (2010)
11. Oliva, J., Serrano, J.I., del-Castillo, M.D., Iglesias, A.: A methodology for the
characterization and diagnosis of cognitive impairment - application to specific
language impairment. Artif. Intell. Med. 61(2014), 89–96 (2014)
12. Peintner, B., Jarrold, W., Vergyri, D., Richey, C., Gorno-Tempini, M.A., Ogar, J.:
Learning diagnostic models using speech and language measures. In: 30th Annual
International IEEE EMBS Conference Vancouver, British Columbia, Canada, 20–
24 August (2008)
13. Rentoumi, V., et al.: Automatic detection of linguistic indicators as a means of
early detection of Alzheimer’s disease and of related dementias: A Computational
Linguistics analysis. 8th IEEE International Conference on Cognitive Infocommu-
nications (CogInfoCom 2017), 11–14 September 2017 (2017)
14. Solorio, T., Yamg, L.: Using language models to identify language impairment in
Spanish-English bilingual children. In: BioNLP 2008: Current Trends in Biomedi-
cal Natural Language Processing, pp. 116–117, Columbus, Ohio, USA, June 2008
(2008). c 2008 Association for Computational Linguistics
15. Stubbs, E., et al.: Procedural discourse performance in adults with severe traumatic
brain injury at 3 and 6 months post injury. Brain Inj. 32(2), 167–181 (2018).
https://doi.org/10.1080/02699052.2017.1291989
16 M. Roldán-Palacios and A. López-López

16. Thompson C.K.: Northwestern Narrative Language Analysis (NNLA) Theory And
Methodology. Aphasia and Neurolinguistics Research Laboratory Northwestern
University, Evanston, IL (2013)
17. Witten, I.H., Frank, E.: Data Mining - Practical Machine Learning Tools and
Techniques. Elsevier, Morgan Kaufmann Publishers, Burlington (2005)
 Automatic Detection and Segmentation of Liver
Tumors in Computed Tomography Images:
Methods and Limitations

Odai S. Salman1(B) and Ran Klein1,2

1 Department of Systems and Computer Engineering, Carleton University, Ottawa, ON, Canada
odaisalman@cmail.carleton.ca
2 Department of Nuclear Medicine, The Ottawa Hospital (TOH), Ottawa, ON, Canada

rklein@toh.ca

Abstract. Liver tumor segmentation in computed tomography images is consid-

ered a difficult task, especially for highly diverse datasets. This is demonstrated by
top-ranking results of the Liver Tumor Segmentation Challenge (LiTS) achieving
~70% dice. To improve upon these results, it is important to identify sources of
limitations. In this work, we developed a tumor segmentation method following
automatic liver segmentation and conducted a detailed limitation analysis study.
Using LiTS dataset, tumor segmentation results performed comparable to state-
of-the-art literature, achieving overall dice of 71%. Tumor detection accuracy
reached ~83%. We have found that segmentation’s upper limit dice can reach
~77% if all false-positives were removed. Grouping by tumor sizes, larger tumors
tend to have better segmentation, reaching a maximum approximated dice limit of
82.29% for tumors greater than 20,000 voxels. Medium and small tumor groups
had an upper dice limit of 78.75% and 63.52% respectively. The tumor dice for
true-positives was comparable for ideal (manual) vs. automatically segmented
liver, reflecting a well-trained organ segmentation. We conclude that the segmen-
tation of very small tumors with size values < 100 voxels is especially challenging
where the system can be hyper-sensitive to consider local noise artifacts as pos-
sible tumors. The results of this work provide better insight about segmentation
system limitations to enable for better false-positive removal development strate-
gies. Removing suspected tumor regions less than 100 voxels eliminates ~80%
of the total false-positives and therefore, may be an important step for clinical
application of automated liver tumor detection and segmentation.

Keywords: Tumor detection · Tumor segmentation · Organ segmentation · Deep

learning · Image standardization

1 Introduction
The main challenge for tumor detection and segmentation is the wide variation of tumors
with respect to size, shape, intensity, contrast, and location. This creates a complex diver-
sity for an observer, human or machine, to handle. In addition, there are variations in
the global image presentation such as image acquisition and reconstruction methods,

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

K. Arai (Ed.): Intelligent Computing, LNNS 285, pp. 17–35, 2021.
https://doi.org/10.1007/978-3-030-80129-8_2
18 O. S. Salman and R. Klein

voxel size, anatomical region coverage and many other parameters. Therefore, minimiz-
ing these image variations helps focus on tumor specific variations. In a previous work
[1], we showed that image standardization significantly improved organ (liver) segmen-
tation in computed tomography (CT) scans by reducing the image specific variations.
Our method consisted of a series of steps to localize the abdominal region, then set a
fixed size and coverage for that region, and finally conduct the organ segmentation. The
segmented region is then inverse-standardized, so it corresponds to the original image
coordinates. Similarly, a system that standardizes/inverse-standardizes both imaging and
tumor variations may boost tumor detection and segmentation performance. In this work
we build on the liver organ segmentation framework to detect and segment liver tumors.
An intuitive approach is to design two modules: the first eliminates imaging-related
variations, and the second handles tumor variations. However, since tumor and imaging-
related variations are interdependent, their separation is not trivial. For instance, stan-
dardizing slice thickness may cause some tumors to disappear due to loss of axial reso-
lution when down-sampling. On the other hand, up-sampling may stretch some tumors
in a way that the image interpolation algorithms will make some tumors’ masks lose
shape information. Moreover, interpolation algorithms introduce blur, compromising
potentially useful texture information. Using interpolation based on nearest value also
does not help if specifically intended for 3D processing, where tumors can take pixelated
shapes along the axial dimension. In other words, applying any morphing standardization
may cause loss of information regarding tumors specifications, critically needed for their
detection or segmentation, especially for 3D processing. Considering the above factors,
it is important to avoid applying these morphing operations and better limit focus on stan-
dardizing only common information the images naturally share. Fortunately, crop/pad
operations can be used to standardize image representation while preserving the image
texture.
Following such option necessitates using 2D processing instead of 3D as an initial
step. Amongst CT images, the transversal view is the most consistent. Even though the
field-of-view (FOV) may vary, its impact is not harmful compared to slice thickness
variations and can be handled by copping/padding as shown by [1]. With slice thickness
being the most harmful factor requiring morphing operations to be standardized, it is
avoided by eliminating its axial dimension through processing one transversal slice
at a time. Of course, this means losing contextual information along that dimension.
Therefore, analytical post-processing context-based correction can be useful [2].
In the context of liver tumor detection and segmentation, a good starting point would
be to focus on the segmented organ rather than including other regions in the image
field-of-view. This way, the model may learn to distinguish tumors from normal tissue
specific to the organ of interest with a simplified task of producing only two major
classes, normal tissue vs. tumor tissue. This work implements such strategy, where the
preprocessing stage replaces all region outside the liver with normal tissue imitating
voxel values on a standardized image intensity scale, hence reducing intensity variation
impact.
In this work we present a novel approach for standardizing CT images including
anatomical and liver-organ segmentation, removal of region outside the liver, and image
intensity normalization. Following standardization, liver tumors are then segmented. We
 Automatic Detection and Segmentation of Liver Tumors 19

compare our performance metrics to previously published results, investigate points of

failure, explore possible solutions and discuss fundamental limitations to further enhanc-
ing segmentation performance. The rest of this article is structured as follows: Sect. 2
presents related work for liver tumor segmentation. Section 3 presents the methods.
Section 4 presents the results along with their analysis. Section 5 provides detailed
discussion. Finally, Sect. 6 presents conclusions.

2 Related Work
L. Chen et al. [3] proposed a cascaded approach to segment the liver first by fusing
segmentation results from three networks (coronal, sagittal and transversal views) and
then segment tumors within the fused liver region. For tumor segmentation, they used
adversarial neural networks to discriminate tumors from normal tissue. They used the
Liver Tumor Segmentation dataset (LiTS) [4], where they achieved a dice of 68% for
liver tumors.
G. Chlebus et al. [5] proposed using U-Net [6] to segment tumors in 2D slices.
Following tumor segmentation, they developed an object-based false-positives reduction
algorithm to achieve an overall dice score of 69% on the LiTS dataset.
L. Bi et al. [7] proposed an approach that utilizes 2D slice-based segmentation using
two ResNets [8]. The first network is fed CT image slices to produce two probability
heatmaps (normal and tumor tissues) for each slice. Then, a CT slice is stacked with its
segmentation results and fed to the second network. This means that the second network
takes segmentation estimates for further enhancement. In their work, they present two
dice scores. The first, which is relatively low (~50%) was obtained for a development
set that was taken from the training set. On the other hand, the score that they were
able to obtain on the test set reached up to 64%, raking the 4th position in the LiTS
challenge. This might indicate that training set contains more difficult images than those
in the testing set. Unfortunately, challenge coordinators do not offer ground-truth labels
for the testing set. Hence, developers who conduct research using this dataset tend to
dedicate a portion of the training set for testing.
W. Li et al. [9] used 2D patch-based voxel segmentation by classifying voxels based
on their neighborhood contents. From tumor ground-truth, they selected a patch centered
at each voxel. They trained AdaBoost [10], random forest (RF) [11] and support vector
machines (SVMs) [12] using patch’s handcrafted features, and convolutional neural
network (CNN) [13] using the patch itself to classify tumor vs. normal. They showed
that CNN achieved the highest dice of 80% on a locally acquired dataset. One major
draw-back to such approach is that it requires classifying many patches to segment a
slice (patch for each voxel), which can be computationally too expensive when having
high resolution images (e.g. small voxel size and/or thin slices). In their paper, they
state that their method has limitation in capability of segmenting heterogeneous-texture
tumors and tumors with fuzzy boundaries. Given their results, we speculate that that their
dataset may have contained these limitations at lower frequency than that encountered
in the LiTS dataset.
L. Meng et al. [14] proposed a 3D dual path convolutional neural network, where
segmentation results at the end of the architecture are fused together. They also used
20 O. S. Salman and R. Klein

conditional random fields (CRF) [15] to reduce false-positive segmentations, resulting

in an overall dice of 68% using LiTS dataset. S. Zheng et al. [16] proposed an approach
that takes 2D CT slices, denoises them using block-matching and 3D filtering (BM3D)
[17], and feeds them to a fully convolutional network (FCN) [18] to segment. Then, 2D
stacked segmentation results are modified using 3D deformable models [19]. In their
work, they trained their model using LiTS dataset and tested on 3D-IRCAD [20] dataset,
under the implicit assumption that both datasets are similar in nature, which may not
be the case, as the best tumor dice achieved in the LiTS challenge was ~70% and for
3D-IRCAD, the worst reported performance was 73% [14]. Regardless, Meng et al.
reported 84% performance on 3D-IRCAD dataset.
P. Christ et al. [21] proposed a cascaded approach for liver tumor segmentation in
abdominal images. In their work, they preprocessed CT images by applying thresholding
in Hounsfield window [–100, 400], followed by histogram equalization. Limitations to
using histogram equalization with no standardized input were discussed by [1]. They
used cascaded fully convolutional neural networks (CFCN) followed by 3D CRF. For
testing, they used 3D-IRCAD and they achieved tumor segmentation dice of 82%.
As the above literature demonstrates, performance measurements must be com-
pared using the same datasets. Furthermore, for the LiTS challenge, despite different
approaches, the upper limit of performance approaches 70% dice.

3 Method
Our overall strategy consists of two main components. Module 1 is a segmentation
module that processes 2D transversal slices one at a time. This module provides initial
tumor segmentation, and it is expected to include false-positive regions. Module 2 is
a patch-based approach designed to classify tumor regions identified in the first step
into true-positive and false-positive regions. This article is solely dedicated to covering
Module 1, with Module 2 to be covered in future work. We explore the limitations of
Module 1 to identify the range of improvements that may be expected from Module 2.
The workflow diagram is illustrated in Fig. 1.

3.1 Dataset
For this study, the LiTS [4] segmentation dataset was used, consisting of 131 CT scans.
Of these, 81 were randomly selected for training, and 50 for testing. Generally, this
dataset consists of images with a tight FOV around the patient. Slice thicknesses varied
as 1.51 ± 1.18 [0.70, 5.00] mm (Mean ± Std [min, max]) as did the number of slices
per image 447 ± 275 [74, 987]. With respect to anatomical region coverage in the FOV,
of the 131 images, 109 contain pelvis, 52 contain chest, 1 contains legs, and 2 contain
head-neck.

3.2 Preprocessing
Each image was automatically preprocessed to include the liver and mask out all other
regions as previously described. Regions outside the liver ROI were subsequently erased
 Automatic Detection and Segmentation of Liver Tumors 21

Fig. 1. CT preprocessing.

by replacing them with voxel intensities resembling normal liver tissues as detailed
below. Two volumes were generated. The first was contrast improved CT (CICT) by
windowing to Hounsfield values ranging [–100, 400]. In the second volume (CICT-
Hist), histogram equalization was applied exclusively to the liver voxels of the first
volume, maximizing tumor to liver contrast. However, this has the undesired effect of
also increasing contrast between normal liver voxels in the absence of tumors. The two
volumes were then combined on the slice level, forming a multi-channel 2D slice stacks.
This configuration maintains actual intensity and emphasizes contrast for the subsequent
tumor segmentation stages.
Reducing Dataset Intensity Diversity: For such purpose, intensity standardization
was applied. However, rather than subtracting the mean and dividing by standard devi-
ation for all images of the training set, the mean and standard deviation used for this
adjustment were those of liver’s normal tissue. Each image voxels were rescaled to
a standard intensity based on intensity distribution derived from normal tissue voxel
intensities in the training images. Since the training set contains diverse tumor contrast,
relying on the absolute liver intensity without standardization can create some instabil-
ity due to absence of intensity reference. Thus, standard intensity was obtained for each
voxel by subtracting the normal tissue mean (μ) and dividing by its standard-deviation
(σ) and finally scaling to range [0, 1], representing [μ − σ, μ + σ]. Hence, normal tissue
voxel intensity distribution was considered as reference to standardize upon. This helps
22 O. S. Salman and R. Klein

to relatively map tumor intensity as abnormality on the normal tissue statistical distri-
bution. Two means and two standard deviations were used (for CICT and CICT-Hist
respectively).

Neutralizing Impact of Contributions from Outside the Liver ROI: Given the pro-
cedure in which an image is convolved, and that the system is divided into two classes
(tumor vs. background), most voxels outside the liver region will significantly contribute
to the background class. The values within this region should be carefully set. In general,
the intensity of tumors is lower than for normal tissue. Setting the background black can
harm the learning because it is closer to tumor intensity. Therefore, to avoid texture bias
in the learning, it is essential that the region outside the liver’s ROI shares similar texture
properties to normal tissue (where both belong to the “background” class). Therefore,
the background for CICT was set to solid intensity of 0.5 (after image intensity scaling)
while CICT-Hist’s was set to uniform noise with intensity values 0, 0.5 and 1; Fig. 1
shows the scheme. Once both volumes were prepared, their corresponding slices were
stacked, and the 2D slices were passed to U-Net to segment tumor regions.

3.3 Tumor Segmentation

This module takes 2D stacked transversal slices one at a time, and segments them using
U-Net architecture. Weighted dice coefficient was used to overcome class imbalance.
The initial learning rate was set to 10–4 with a drop rate of 80% every two epochs for a
maximum of 10 epochs. The training used a minimum batch size of 4 stacked images, and
it was applied exclusively on slices with tumors. Since there are chances of encountering
low intensity around organ boundaries (due to inaccurate segmentation), an optional
step to apply binary erosion (scratching a thin layer from the outside surface of the liver
segmentation) was implemented. The eroded images may exclude false-positives and
possibly small true-positive tumors residing on the liver boundaries. Hence, this can
reduce the detection accuracy. In the results section, we presented both configurations,
with and without erosion.

3.4 Performance Metrics

For this study, two performance metrics were used. The first metric is the dice coefficient,
which is defined as twice the intersection volume between segmentation and ground-
truth, divided by the summation of their volumes. Since there are images with different
slice thickness and hence axial pixel height, a normalized average is needed to get the
same contribution from two tumors with the same physical volume and segmentation,
while taken from images with different resolutions. For instance, a tumor from an image
that has a fine slice thickness of 0.7 mm per height pixel can contribute ten times a
tumor of the same size from an image with slice thickness of 7 mm per height pixel.
The dice depends on the voxel counts of the volume. In all images, x-y resolution is
usually 512 × 512 pixels, with the liver approximately covering 250 × 250. However,
the height z-resolution varies between images, and that change the proportional aspect
ratio. This results in inaccurate dice values even if the tumors actual contribution is the
 Automatic Detection and Segmentation of Liver Tumors 23

same. To minimize and stabilize the volatility of this effect, post-segmentation height was
normalized into a common scale that reflects an approximate average amongst dataset
images. Height was normalized using nearest pixel method. Hence, the dice can now be
more representative to true segmentation performance with no bias caused by the image
resolution. The dice score was calculated for the whole test set following the definition
of the first metric.
To study the dependence of dice performance on tumor size, these were plotted with
respect to one-another. Segmented tumors were grouped into three, arbitrarily selected
size groups: small, medium, and large corresponding to (x ≤ 5000), (5000 < x ≤ 20,000)
and (x > 20,000), respectively, where x is the number of voxels in the segmented tumor
region. For each size range, the plot included scattered distribution with each point
presenting the dice for a single segmented region. In addition, an “approximate” dice
was calculated as a size-based weighted average of all singular regions’ dices within the
group. We considered this average as approximate in relation to the overall dice defined
above. The range dice can indicate the behavior of the group in relation to the overall
dice.
The other metric was the tumor detection accuracy, which is the ratio of the detected
ground-truth positives to the total ground-truth positives for the test set. A segment
that intersected a ground-truth region was considered a true-positive. A false-positive
tumor is a segmented region that has no intersection with ground-truth. The number
of true-positive segmented regions was less than the actual positives due to network’s
detection capacity and segmentation efficiency for this problem, where in the latter
multiple ground-truth regions may intersect with a single true-positive segmented region.

3.5 Region Analysis and Performance Conditions

To study limitations of the segmentation module, for each tumor size group we measured
the dice using four consecutive stages of tumor exclusion conditions:

1) Base value–included all detected tumors within the liver region, representing results
without any false-positive removal.
2) False-positive exclusion–excluded all false-positive detected tumors, representing
best possible dice if all false-positives were removed.
3) Speckle exclusion–excluded all detected tumors less than 500 voxels in size, rep-
resenting a simple criterion by which tumors that are most likely to be noise are
removed.
4) Low-overlap exclusion–excluded detected tumors that had a relatively low dice (poor
overlap with ground-truth), representing detected tumors that disagreed with the
ground-truth and hence may be harder to detect (e.g. low contrast). The aim of this
exclusion is to isolate regions with bad performance to study their properties. This
can be useful to designing a system that is specialized in processing similar instances.
Such properties may include contrast, intensity, and location.

We reported the number of removed and remaining regions after applying each exclusion
condition.
24 O. S. Salman and R. Klein

4 Results and Analysis

Table 1 shows the results based on different conditions. Following automatic liver
segmentation, tumor segmentation dice achieved base value (without false-positives
removal) of 71.11%. Manual (or ideal) organ segmentation starts with the ground-truth
liver, while automatic segmentation is based on liver segmentation produced by a trained
network. The tumor segmentation in both cases was made using identical procedure. The
only difference is the starting liver segmentation. The automatic results are less than the
ideal ones but comparable. Erosion scratches some amounts from the surfaces for the
sake of reducing false-positives. Erosion hurts accuracy of the automatic segmentation.
The difference between manual and automatic for both dice and detection accuracy is
due to the variations in the liver segmentation. These points are discussed later.

Table 1. Tumor dice and detection accuracy values following manual and automatic (trained
network) liver segmentation methods. For both methods, results were presented in presence and
absence of applying binary erosion.

(n = 50 test images) Dice (%) Dice limit (%) Detection accuracy (%)
All positive FPs removed
Manual liver 76.17 81.83 87.31
segmentation
Manual liver 77.50 81.34 87.31
segmentation-erosion
Automatic liver 69.20 76.53 83.40
segmentation
Automatic liver 71.11 76.44 79.25
segmentation-erosion

In some sense we have dealt with the false-positives as a separate independent prob-
lem. With all false-positives included, the result gives a lower limit to the performance and
when all are removed, the dice gives an upper limit to the segmentation performance of
the segmentation module, independent of how a false-positive removal module operates.
In other words, the test reflects the segmentation learning capacity for true tumors.
Table 2 shows the number of detected tumors and false-positives based on size and
segmented region exclusion conditions. The table shows the amounts removed, and the
remaining ones for each exclusion applied as explained in the following figures. Table
3 shows the dice under various conditions. The dice improved from 25.67% to 60.61%
for the small tumors when false-positives were removed, reflecting their large number
for that range. By contrast, when removing 157 extremely small tumors (<500 voxels),
which dominate this range, the dice improved to 63.52%. Removing all regions with
dice less than the moving average-sigma dramatically improved performance by ~10%
points.
 Automatic Detection and Segmentation of Liver Tumors 25

Table 2. Number of detected tumors and number of false-positives grouped by size and based
on consecutive stages of region exclusions (Rmv.: no. removed regions, Rmn.: no. remaining
regions).

(n = 50 test Base False-positives + Speckles + Removing low

images) segmentation excluded excluded (regions overlap regions
result (no of size ≤ 500) with dice < mov
exclusion) avg – mov σ
Size All points Rmv. Rmn. Rmv. Rmn. Rmv. Rmn.
Automatic liver segmentation
Very small 1336 1179 157 157 0 0 0
(≤500)
Small 126 75 51 0 51 10 41
(500 < x ≤ 5K)
Medium 26 3 23 0 23 4 19
(5K < x ≤ 20K)
Large 19 1 18 0 18 4 14
(>20K)
Total 1507 1258 249 157 92 18 74
Manual liver segmentation
Very small 1547 1388 159 159 0 0 0
(≤500)
Small 200 113 87 0 87 16 71
(500 < x ≤ 5K)
Medium 45 6 39 0 39 9 30
(5K < x ≤ 20K)
Large 32 0 32 0 32 5 27
(>20K)
Total 1824 1507 317 159 158 30 128

For medium and large sizes, removing false-positives improved dice ~9% and ~5%
points, respectively. Removing regions with dice less than the moving average-sigma
further improved dice ~4% points for medium, and ~4% for the large. Considering only
false-positive exclusion, the total approximate dice improved from ~68% to ~80%. With
all exclusions applied (≤500 and <moving average-sigma), the overall dice reached
~85%.
The histograms of false-positive tumors in Fig. 2 show that very small size segmented
tumors are much more common than medium and large false-positive tumors with >92%
of these having fewer than 500 voxels regardless of the liver segmentation method.
Without false-positive exclusion, removing them improved the mean dice for the small
size category by ~8% and ~6% for automatic and manual liver segmentation, respectively
(Table 3).
26 O. S. Salman and R. Klein

Table 3. Average dice for different tumors sizes under certain exclusions.

(n = 50 test Base False-positives + Speckles + Removing low

images) segmentation excluded excluded overlap regions
result (no (regions of size with dice < mov
exclusion) ≤ 500) avg – mov σ
Automatic liver segmentation
Small 25.67 60.61 63.52 73.85
(≤5K)
\\Medium 69.54 78.75 78.75 83.08
(5K < x ≤ 20K)
Large 76.86 82.29 82.29 86.26
(>20K)
Overall 68.15 80.14 80.58 84.96
Excluding regions ≤ 500vx following base segmentation results improved dice from 25.67%
to 33.28%
Manual liver segmentation
Small 21.74 54.40 57.17 68.37
(≤5K)
Medium 61.92 71.91 71.91 84.95
(5K < x ≤ 20K)
Large 85.43 85.43 85.43 88.64
(>20K)
Overall 75.85 82.49 82.78 87.50
Excluding regions ≤ 500vx following base segmentation results improved dice from 21.74%
to 27.49%

The behavior of the tumor dice following manual and automatic liver segmentation
is summarized in Tables 1, 2 and 3 and compared with respect to the moving average
exclusion condition in detail in the Figs. 3, 4, 5, 6, 7, 8, 9 and 10. The numbers of
detected tumors differ between the manual and automatic liver segmentation, because
tumor segmentation may combine two regions from the ground-truth or vice versa. With
different liver ROI, the tumor segmentation algorithm is expected to act differently.
Another more likely reason is that the automatic segmentation could have missed portions
of the liver, which was reflected in the lower tumor detection accuracies (Table 1).
Exploring the Variety of Random
Documents with Different Content
climbed up to the ramparts to look at the sunset from the heights of
the Burg, while the lines of a ballad by Uhland rang in his memory,
he did not think that evil masters and false prophets would twice in
a quarter of a century unchain the lightning over Europe and the
rest of the world, and that through them so many treasures of art
and beauty would be destroyed, so many human lives sacrificed, so
much suffering piled up. Indeed, when one studies the genesis of
this unheard-of drama there can be no question of romanticism;
what we have to deal with rather is a perverted romanticism, a
morbid perversion of the sense of greatness, and the mind is baffled
by the true significance of the ideas of National Socialism—ideas
which I shall touch upon only in passing to show how they led the
Defendant Rosenberg, since it is he of whom I am speaking, and his
codefendants to commit the crimes which are held against them:
The concept of race, to begin with, which we see arising in a
country which in other respects resembles any other but where the
intermingling of ethnic types of every variety took place through the
centuries on a gigantic scale; this anti-scientific confusion which
mixes the physiological features of man with the concept of nations;
this neo-paganism which aims at abolishing the moral code, the
justice, and security which 20 centuries of Christianity have brought
to the world; this myth of blood which attempts to justify racial
discrimination and its consequences: slavery, massacre, looting, and
the mutilation of living beings!
I shall not dwell, Mr. President, on what we consider a jumble of
nonsense which claims to be philosophy and in which may be found
to be the most heterogeneous fragments of all kinds taken from
every source, from the megalomaniac concepts of Mussolini, Hindu
legends, and the Japan of the samurai, the cradle of fascism, which
swept over the world like a tidal wave. The previous presentations
have already adequately dealt with these conceptions. I shall simply
stress today that these pseudo-philosophic conceptions tended solely
to set back humanity thousands of years by reviving the clan
conception, which assumes the law of might as the supreme law—
the Faustrecht already formulated by the Iron Chancellor, the right to
cheat others, the right to take the property of others, the right to
reduce man to slavery, the right to kill, the right to torture.
But homo sapiens refuses to return to the state of homo lupus.
International law is not morality without obligation or sanction. The
Charter of 8 August has recalled and specified the obligation; it is for
you, Gentlemen, to apply the sanction.
One of the consequences of these theories of the superiority of
the race or of the so-called “Germanic Race” was to lead certain of
the conspirators, particularly the Defendant Rosenberg, of whom we
are speaking, to become plunderers; and it is this aspect of the
activities of the Defendant Rosenberg which I should like very briefly
to stress, for it concerns France and the occupied countries of the
West and had deeply harmful consequences for their artistic,
intellectual, or merely utilitarian heritage.
I wish to speak of all the measures decreed or applied by
Rosenberg with the aim of removing from France and the western
countries cultural treasures, works of art, and property belonging to
groups or individuals, and transferring to Germany all these riches.
Gentlemen, owing to the limited time which we have at our
disposal, I shall limit myself today to recalling how certain organisms
were made to collaborate in this pillaging through orders from higher
quarters. I shall indicate, first of all, the part played by the Gestapo,
which was ordained by a decree issued by the Defendant Keitel,
dated 5 July 1940, which bears the Document Number 137-PS and
which was submitted by the American Delegation, under Exhibit
Number USA-379, on 18 December 1945 (Exhibit Number RF-1400).
I refer likewise to a second order dated 30 October 1940, which
reinforced and detailed the orders given in regard to pillaging by
what was known as the Einsatzstab Rosenberg. This is Exhibit
Number RF-1304 (Document Number 140-PS), which was quoted by
the Economic Section of the French Prosecution.
Thus, Keitel and Rosenberg went back to the conception of a
booty exacted by the triumphant German people from the Jewish
people with regard to whom it was not bound by the conditions of
the Compiègne Armistice. This intervention by the chief of the army,
as indicated by the orders to which I have just referred, suffices in
my opinion to prove the important part played by the German Army
in this looting; and the Tribunal will not fail to remember that when
it makes its decisions as to the guilt of the Defendant Keitel and the
Defendant Göring.
If I mention the Defendant Göring, it is because a third
document proves that this defendant gave the operation his full
support, inviting all the organizations of the Party, the State, and the
Army to afford the fullest possible support and assistance to
Reichsleiter Rosenberg and his collaborator Utikal, whom Rosenberg
himself had appointed Chief of the Einsatzstab on 1 April 1941. This
is the order of 1 May 1941, which we produced under our Exhibit
Number RF-1406 (Document Number 1614-PS). If we examine the
text of this decree carefully we cannot fail to be struck by the first
paragraph. The Tribunal will surely allow me to reread it rapidly:
“The struggle against the Jews, the Freemasons, and other
ideologically opposed forces allied to them, is a most urgent task of
National Socialism during the war.”
Thus, it was enough for one to have a philosophy of life different
from that described as the Nazi Weltanschauung, to be exposed to
the danger of seeing one’s cultural property seized and transferred
to Germany. But the Tribunal will surely remember from the
documents already presented to it, that not only cultural property
was involved, but that anything with any kind of value was taken
away.
The Defendant Rosenberg tried, in the course of an interrogation
carried out by the superior officers in charge of the preliminary
investigations to claim, without much conviction, it seems to me,
that the cultural property in question was intended solely to adorn
the collections of the National Socialist Hohen Schulen. We shall see
presently, in presenting the text of this interrogation, how we may
judge this. But it is a fact which I wish to present now that, from the
documents which we possess, at least, it does not seem that the
Defendant Rosenberg appropriated works of art, precious stones, or
other objects of value for himself. Consequently, in the light of the
proceedings as conducted thus far, no accusation of this kind can be
brought against him. We shall not say as much for the Defendant
Hermann Göring, of whom we shall speak a little later and who,
according to the documents that we possess, may be convicted of
having appropriated to his own use part of the objects of art taken
from the countries of the East and the West.
I shall not dwell on the discussion which might arise about these
misappropriations. I shall go straight on to the interrogatory of the
Defendant Rosenberg. This is the document that was introduced
yesterday by the Economic Section of the French Prosecution, which
bears the Exhibit Number RF-1331, and which we use today as
Document Number ECH-25.
I think that the Tribunal will easily be able to refer to this
interrogatory, but meanwhile I should like very briefly to summarize
the essential points which I think should be brought up.
Colonel Hinkel, questioning the Defendant Rosenberg, asked him
on what legal grounds such looting could be justified. The Defendant
Rosenberg first answered that these seizures were justified by the
hostility which certain groups had manifested toward the National
Socialist ideology. But a little further on, on Page 4, the Defendant
Rosenberg made the following verbatim statement:
“I considered them”—he is referring to the measures which he
himself had taken—“a necessity caused by the war and by the
reasons which caused the war.”
A few moments later, pressed by Colonel Hinkel, the Defendant
Rosenberg invoked the necessity of putting into safekeeping
property thus seized, a necessity which will certainly constitute one
of the main points of his defense. But Colonel Hinkel replied to the
Defendant Rosenberg:
“And so if your idea was to safeguard art objects, it sounds rather
strange, doesn’t it, that you were going to safeguard only some art
objects and not others?
“On the other hand, with regard to the maintenance of the objects,
there were objects at least equal in value to those which had been
removed, but to which no one paid any attention.”
 Finally, the Defendant Rosenberg admitted that he had very often
given no receipt to those concerned, which in itself precluded any
idea of eventually returning the property to the legitimate owners.
The truth of the matter is that these were treasures of very
considerable value, and the Defendant Rosenberg in the end
admitted that he regarded these acquisitions as an accomplished
fact. We consider that the fact of having thus removed works of art
and objects of value is purely and simply what is known in civil law
as misappropriation. These misappropriations were made on a vast
scale with the grandiose means which the Third Reich had at its
disposal, means which were further facilitated by the intervention of
the Army and the Luftwaffe. But it is nonetheless true that the
criminal character of these misappropriations remains; and we urge
the Tribunal, when it delivers judgment, to declare that it was by
fraudulent seizure that the Defendant Rosenberg and his
codefendants robbed France and the western countries of all the
objects of value and all the art treasures and cultural treasures.
As to what the objects themselves consisted of, Mr. President and
Your Honors, I would respectfully refer the Tribunal to the report
submitted by the Economic Section yesterday, which was made by
Dr. Scholz, the associate of the Einsatzstab Rosenberg. This report
was submitted by the Economic Section under Exhibit Number RF-
1323 (Document Number 1015-PS), and in it the Tribunal will find
enumerated everything that the Einsatzstab took out of France. In
this connection I shall make an incidental remark in answer to the
question that the President asked my colleague yesterday about the
Rothschild collections. The President asked my colleague, “Have you
proof that certain collections and objects of value were taken from
the Rothschild collections?”
I should like, Mr. President, to point out that there are two proofs
of this. The first is the immediate result of the Rosenberg
interrogation of 23 September 1945. I have just spoken to the
Tribunal of the all-important questions put to the Defendant
Rosenberg as to the legitimacy and legal basis of these removals. I
beg the Tribunal to refer to Page 5 of these minutes. I read from the
text the question asked by the American officer in charge of the
interrogation, my eminent friend, Colonel Hinkel:
“Question: ‘How do you justify the confiscation of art treasures
belonging to the Rothschild family?’ ”—A very precise question. It
concerned the art treasures taken from the Rothschild family by
Rosenberg’s organization.
“Answer: ‘Still from the same general point of view.’ ”
That means that the Defendant Rosenberg claimed to justify the
confiscations made to the detriment of the Rothschilds by the
reasons which I had the honor of analyzing to the Tribunal a few
moments ago.
A second consequence: The Defendant Rosenberg thus admitted
with his own lips that the Rothschild family was among those
despoiled. That confession, Mr. President, Your Honors, can be
considered as one of the proofs, one of the main proofs. This is the
first answer, then, to the question that the President asked
yesterday.
The second proof which I wish to present to the Tribunal is the
following: I beg the Tribunal to refer to the report by Dr. Scholz
mentioned above and produced yesterday in the document book of
the Economic Section. This is Exhibit Number RF-1323 (Document
Number 1015-PS).
If the Tribunal will kindly refer to it, that is to say, the report by
Dr. Scholz, the second paragraph of Page 1, it will find the following
statement, “The special staff not only seized a very considerable part
of the collection. . . .”
THE PRESIDENT: [Interposing.] As I said the other day, we
cannot keep all the books before us; but it seems to me that, as you
have shown that the Defendant Rosenberg agreed that this
collection had been taken, that is quite sufficient.
M. MOUNIER: Mr. President, I understand perfectly your point of
view. I should like respectfully to point out to you that I was to
speak immediately after my colleague, and if I had done so you
would have had this document book before you. We had a delay of
one day, and I apologize for not having thought of asking you to
bring this document again this morning.
However, I respectfully ask the Tribunal to be good enough to
note this reference which it will easily find. It is a very short
passage, which I should like to read to the Tribunal. It will not take
very much time.
THE PRESIDENT: Certainly.
M. MOUNIER: This declaration is simply the following:
“The special staff”—that is to say, the Einsatzstab Rosenberg—“not
only seized a very considerable portion of the collection which the
Rothschilds had left behind in their Paris mansion. . . .”
I shall not read the rest.
Here then, Gentlemen, is an official report which cannot be
disputed and which demonstrates, like the previous proof, that the
Rothschild collection was among those pillaged.
I do not insist on these facts, which are known to you. It seems
to me that the two points on which I have just cast a ray of light
suffice to make it clear that illegal seizures—fraudulent seizures—
were really operated by the Defendant Rosenberg to the detriment
of France and to the detriment, likewise, of the western countries.
As for their importance, I do not want to abuse the patience of the
Tribunal by quoting statistics. I respectfully ask the Tribunal to refer
to the Scholz report which I have twice mentioned in the course of
my previous statements.
I should not, however, wish to leave the case of Rosenberg, for
the time being, without quoting to the Tribunal a passage from an
article by the French writer François Mauriac, of the French
Academy. François Mauriac was present on 7 November 1945 at the
inaugural session of the National Constituent Assembly at the Palais
Bourbon. On this occasion François Mauriac invoked a memory which
was recalled in Le Figaro of 6 November 1945 in the following terms:
“Almost five years ago to a day, from the height of this rostrum, the
most illustrious in Europe, a man spoke to other men dressed in field
grey. His name was Alfred Rosenberg. I can testify to the exact date.
It was 25 November 1940.
“Rosenberg leaned his elbows on this rostrum, where the voices of
Jaurès and of Albert De Mun were once heard and where, on 11
November 1918, Clemenceau nearly died of joy. Here are his words:
“ ‘In one gigantic revolutionary burst’—he said—‘the German nation
has reaped such a harvest as never before in its history. The French
will admit one day, if they are honest, that Germany has freed them
from the parasites of which they could not rid themselves unaided.’
“And the Nazi philosopher”—continues Mauriac—“then proclaimed
the victory of blood. He meant”—writes Mauriac—“the victory of
race; but it happens that a man may utter prophetic words
unwittingly and without realizing the full import of the words which
God places upon his lips. As Rosenberg predicted at the Palais
Bourbon on 25 November 1940, it was indeed blood that won the
victory. It was the blood of the martyrs which in the end choked the
executioners.”
M. President, with the approval of the Court, and with the same
brevity as heretofore—and I hope the Tribunal will appreciate the
care I am taking not to abuse its patience—I should like to say a few
words on the individual charge against the Defendant Fritz Sauckel.
Your Honors, the Tribunal is already acquainted with the really
remarkable work, the genuinely positive work, presented to it some
time ago by my colleague and friend, M. Jacques Bernard Herzog.
This is why, with your permission, I shall pass over the facts
themselves, which are known to you, and limit myself to the part
beginning on Page 3 of my brief; and we shall examine together, if it
please the Tribunal, the grounds for the pleas advanced up to now
by the Defendant Fritz Sauckel.
One question must be asked first of all: Was Fritz Sauckel acting
under orders when he carried out this recruiting—so-called voluntary
in part but compulsory in most cases—this recruiting of laborers
destined to supply the needs of the German Reich?
 According to Sauckel, when he was appointed Plenipotentiary for
the Allocation of Labor on 27 March 1942, his initial program did not
include the conscription of foreign workers; and it is supposed to
have been Hitler who intervened then. For it is striking, Your Honors,
when you read the minutes of the interrogations and also, I am
sure, when the defendants speak before the Tribunal, you will see
that most of them take refuge behind two great shadows; the
shadow of the former Führer and the shadow of his accursed
second, Himmler. Here we can see Hitler intervening to tell Sauckel,
according to the latter, that the use of foreign workers in the
occupied territories is not contrary to the Hague Convention for two
reasons; firstly, the countries involved surrendered unconditionally
and consequently we can impose any kind of labor conditions on
them, and secondly because Russia has not signed this convention.
If, therefore, we use Russian workers on compulsory labor and make
them work to death, we are not violating the Hague Convention.
This, Your Honors, is the reasoning of the Defendant Sauckel on
this point, without the addition of a single word. Hitler is supposed
to have ordered him to recruit workers, at first using persuasion and
then all the means of compulsion which you already know;
suppression of ration cards, for instance, which compelled men, who
saw their wives and children starving, to volunteer for work which
would be used against their own fellow citizens and against the
soldiers of the Allied armies with whom all their sympathies lay.
The Tribunal will know how to deal with such an excuse for, in
the first place, Sauckel, by virtue of the powers conferred upon him
by his office, enjoyed full authority in regard to everything to do with
the labor necessary for the execution of the Four Year Plan. On the
other hand, on taking up his appointment as Plenipotentiary for
Labor Allocation, Sauckel knew that he would be unable to carry out
his mission without resorting sooner or later to means of coercion.
In any case, Sauckel, as well as most of the defendants who are
before you, enjoyed the most extensive powers, indeed autonomous
powers. Consequently, he cannot shelter behind orders received.
THE PRESIDENT: M. Mounier, you must forgive me if I interrupt
you; but as I pointed out yesterday, I think, we have already had an
opening statement which contained argument from the United
States, from Great Britain, and from M. De Menthon on behalf of
France, and we have, in the past, confined other counsel. . . .
Do you hear me? I was saying that after having heard the
opening statement from the United States, from Great Britain, and
from France, we have in the past, confined the counsel who have
followed them to a presentation of evidence and have not permitted
them to go into an argument.
I am not sure that that rule has been strictly carried out in all
cases because it is, perhaps, somewhat difficult to confine the
matter; but we have, on several occasions, pointed out to counsel
who have followed the counsel who made the leading statement that
they ought to confine themselves to a presentation of the evidence.
I think the Tribunal would wish you, if possible, to adhere to that
rule and, therefore, not to argue the case but to present the
evidence, that is to say, to refer us to the evidence insofar as it has
already been put in evidence; to refer us to it by its number, possibly
stating what the substance of the evidence is; and, in reference to
any document which has not yet been put in evidence, to read such
parts of that document as you think necessary.
M. MOUNIER: Very well, Mr. President, to meet the wishes of the
Tribunal, I shall limit myself, as concerns the Defendant Sauckel, to
referring to figures which, it seems to me, do not admit of
argument, since they are the figures given by the Defendant Sauckel
himself under interrogation. This does not seem to me to infringe
upon the rule which the President has just drawn to my attention.
The figures stated are the following: In 1942 there were already
a million foreign workers in Germany. In one year Sauckel
incorporated into the economy of the Reich some 1,600,000 war
prisoners to meet the needs of war economy.
I beg to refer the Tribunal to Exhibit Number RF-1411 in my
document book. This is an interrogation of the Defendant Speer
under the date of 18 October 1945, which has already been
submitted by the United States Prosecution on 12 December 1945,
under Exhibit Number USA-220 (Document Number 3720-PS). In this
interrogation the Defendant Speer states that 40 percent of all
prisoners of war were employed in the production of arms and
munitions and in related industries.
I likewise offer under Exhibit Number RF-1412 (Document
Number 1292-PS) of 13 December 1945, a memorandum signed by
Lammers, Secretary of the Reich Chancellery, giving an account of
the discussion which occurred at a conference held on 4 January
1944. On that date, 4 January 1944, in the course of a conference,
at which, in addition to the Defendant Sauckel, the Führer himself,
Himmler, Speer, Keitel, Field Marshal Milch, and others were present,
the number of new workers to be furnished by Sauckel was fixed at
four million.
I must mention in this connection that in the course of this
meeting, Sauckel expressed doubts as to the possibility of furnishing
this number of workers unless he were given sufficient police forces.
Himmler replied that he would try to help Sauckel to achieve this
objective by means of increased pressure.
Consequently, when the Defendant Sauckel claims, as he
probably will do, that he had absolutely nothing to do with the
institution now spurned by everyone, known as the Gestapo, we
may answer him by official German documents showing that for the
recruitment of labor he really did employ the police with all the more
or less condemned means already pointed out to you.
As for France alone, the demand for workers at the beginning of
1944 amounted to one million; and this figure was over and above
the number of men and women workers already sent to Germany,
who in June 1944 numbered one million to one and a half million.
The Defendant Sauckel, therefore, committed the offenses
already known to the Court. We have an old adage, an old slogan
we may say, according to which “The court is the law”; and it is
proper to present only the facts. I shall, therefore, abstain from
reading the passage on Page 9 of my presentation dealing with
those articles of the law under which the activities of the accused,
Sauckel, are punishable.
Mr. President, Your Honors, I should like now to summarize the
activity of the Defendant Speer, for as regards France and the
western countries the Defendant Speer incurs responsibilities of the
same nature as those of the Defendant Sauckel. Like the defendant
of whom I have just spoken, he permitted violations of the laws of
war, violations of the laws of humanity, in working towards the
drafting and carrying out of a vast program of forced deportation
and enslavement of the occupied countries.
Speer, Mr. President, first took part in working out the program of
forced labor and collaborated in its adoption. In the course of his
interrogatory, he stated under oath: First, that he took part in the
discussion at which the decision to use forced labor was made;
second, that he collaborated in the execution of this plan; third, that
the basis of this program was the removal to Germany by force of
foreign workers on the authority of Sauckel, Plenipotentiary for
Allocation of Labor under the Four Year Plan. The Tribunal will kindly
refer to Document Number 3720-PS, submitted by the United States
Delegation on 12 December 1945, which I quote under Exhibit
Number RF-1411 of our documentation.
As regards France, in particular, Hitler and the Defendant Speer
held a conference on 4 January 1943 in the course of which it was
decided that more severe measures would be taken to expedite the
recruiting of French civilian workers without discrimination between
skilled and unskilled workers. This is made clear by a note to which I
would ask the Tribunal to refer. That is a note signed by Fritz
Sauckel, himself. It has already been presented by the American
Prosecution under Document Number 556-PS (Exhibit Number RF-
67).
The Defendant Speer knew that the levies for forced labor in the
occupied territories were obtained by violence and terror. He
approved the continuation of these methods from September 1942
onward. He knew, for instance, that workers were deported by force
from the Ukraine to work in the Reich. He knew, likewise, that the
great majority of workers in the occupied regions of the West were
sent to Germany against their will. He even declared before the
American magistrate who was questioning him that he considered
these methods regular and legal.
The Defendant Speer, knowing that the foreign workers were
recruited and deported for forced labor in Germany, made specific
demands for foreign workers and provided for their employment in
the various branches of activity placed under his direction.
The preceding paragraphs summarize all the declarations made
by the defendant in the course of the interrogation already
mentioned and to which I have just referred.
I beg to remind you that Speer, in addition, was a member of the
Central Committee of the Four Year Plan. On account of this, and in
common with Field Marshal Milch, only Hitler and Göring were
superior to him as far as demands for labor were concerned. He
likewise took part, in this capacity, in discussions which took place
with Hitler to settle the numbers of foreign workers required. He
knew that most of these forces were obtained by means of
deportation, through coercion and enslavement of the occupied
countries. Proof of this is furnished by various passages of the
minutes of the Central Committee of the Plan and from Speer’s
conferences with Hitler. I refer to Document Numbers R-123 and R-
124 which have been submitted under Exhibit Number USA-179, on
12 December 1945 (Exhibit Number RF-1414).
Speer did not hesitate to resort to methods of terrorism and
brutality as a means of achieving a peak output from the forced
workers. He found justification for the action of the SS and of the
police and for the use of concentration camps to subdue
recalcitrants.
I beg to recall to the Tribunal the document relating to the
minutes of the 21st meeting of the Central Committee of the Four
Year Plan, 30 October 1942, Page 1059, already quoted. This is the
document which I quoted previously, Exhibit Number USA-179,
Document Numbers R-123 and R-124 on 12 December 1945 (Exhibit
Number RF-1414).
The Defendant Speer likewise bears responsibility for the use of
prisoners of war in military operations directed against their
countries; for in his capacity as chief of the Todt Organization, he
forced citizens of the Allied nations to work for this organization,
particularly, in the building of fortifications and, among other things,
the famous West Wall. He likewise forced Frenchmen, Belgians,
Luxembourgers, Dutchmen, Norwegians, and Danes to manufacture
arms to be utilized against the allies of the countries to which they
themselves belonged.
Finally—and this is a very important question regarding the
responsibility of the Defendant Speer—he participated directly in the
use of internees from the concentration camps. He proposed the use
of internees from the concentration camps in the armament
factories. Now, in view of the wretched physical condition of the
prisoners, no profit but only the extermination of the prisoners could
be expected from this measure. The use of internees from the
concentration camps in the factories had the effect of increasing the
demand for this type of labor; and this demand was satisfied in part,
at least, by sending to the concentration camps persons who, in
ordinary times, would never have been sent there.
Speer went so far as to establish, near the factories,
concentration camps which served solely to feed them with labor.
He knew the Mauthausen Camp. The Spanish witness, Boix,
whom the Tribunal heard a few days ago, attested under oath that
he had seen, with his own eyes, the Defendant Speer visit the camp
at Mauthausen and congratulate the directors of this camp. He even
declared that he had worked on the preparation of photographs of
this scene. Consequently this visit to the camp must be considered,
absolutely beyond doubt. He therefore saw for himself the barbarous
conditions in which the prisoners lived. Nevertheless, he persisted in
utilizing labor from the Camp of Mauthausen in the factories under
his authority.
I have concluded the case against Speer.
THE PRESIDENT: We will adjourn now for 10 minutes.
[A recess was taken.]

M. MOUNIER: Mr. President, Your Honors, considering the strictly

limited time at my disposal, I shall be compelled, in dealing with the
Defendant Göring, of whom I shall have the honor to speak to you,
to skip Pages 1, 2, and 3 of this presentation. I ask the Court now to
turn to Page 3 of my statement.
 I should like to present to the Tribunal the question of the
responsibility of the Defendant Göring for the measures taken
against the commandos and against Allied airmen who fell into the
hands of the Germans during their missions.
During the Trial we have on several occasions mentioned an
order given by Hitler on 18 October 1942, which was first submitted
by the American Delegation on 2 January 1946 under the Document
Number 498-PS (Exhibit Number RF-1417). It is an order detailing
the measures to be taken against commandos in operations in
Europe and Africa. They were to be exterminated to the last man,
even if they were in military uniform, and no matter what their mode
of transport might be: boat, plane, or parachute. An order was given
to take no prisoners. In the occupied territories isolated members of
commandos who might fall into the hands of the German forces
were to be handed over immediately to the Sicherheitsdienst, RSHA
branch. This order did not apply to enemy soldiers who were
captured or who surrendered in open battle and within the scope of
combat operations.
Among those notified was the Oberkommando of the Luftwaffe.
Consequently, the Defendant Göring knew of this order; and in his
capacity as Commander-in-Chief of the Air Force, as well as in his
capacity as Commander-in-Chief of one of the three military
services, he has joint responsibility with the leaders of the other
services.
We know, also, that on the same date, 18 October 1942, Hitler
had a memorandum distributed annotating the previous instructions
and announcing that if one or two prisoners were spared for the
time being, so that information might be obtained from them, they
were to be put to death as soon as they had been interrogated.
I refer to Exhibit Number RF-1418 (Document Number 503-PS) of
9 January 1946. The American Prosecution which produced this
document has also submitted to the Tribunal—and I shall not come
back to this fact—a certain number of cases proving that this order
was frequently carried out.
On the other hand, the Tribunal already knows that numerous
Allied airmen, who found themselves in German territory after losing
their planes, were maltreated and lynched by the Germans with the
connivance of the authorities. As evidence we present only the order
of 10 August 1943 by which Himmler forbade the police to take part
in these lynchings and forbade them equally to oppose them. I refer
to Document Number R-110, presented 19 December 1945 as
Exhibit RF-1419.
Goebbels, in an article in the Völkischer Beobachter, intervened
in the same way. Bormann, in a memorandum of 30 May 1944,
confirmed these instructions and stipulated that they should be
passed on to the administrative authorities, not in writing but by
word of mouth only. I refer to Document Number 057-PS (Exhibit
Number RF-1420), cited on 17 December 1945 by the American
Delegation.
These instructions were carried out to the letter, to such an
extent that the American forces have brought to trial, since the
capitulation, a considerable number of German civilians who had
murdered unarmed Allied airmen.
But the Defendant Göring was not satisfied simply to let these
things happen. At a conference which took place on 15 and 16 May
1944 he stated that he would suggest to the Führer that not only
parachutists but also American or English crews who attacked,
indiscriminately, cities and civilian trains in motion should be put to
death on the spot forthwith. This is Exhibit Number RF-1421
(Document Number L-166), cited by the French Prosecution, 31
January 1946, under Exhibit Number RF-377.
In fact, Göring saw Hitler between 20 and 22 May 1944. The Air
Force General, Korten, sent the Defendant Keitel a memorandum
pointing out that Hitler had decided that enemy airmen who were
shot down should be put to death without trial if they had
participated in acts described as terroristic. This is Document
Number 731-PS (Exhibit Number RF-1407), which we submit to the
Court in the form of a photostatic copy. I ask the Tribunal’s
permission not to read this document. I think the Tribunal will prefer
to read it for themselves. However, I am at their disposal if they wish
me to read it.
THE PRESIDENT: No; it has already been put in, has it not?
 M. MOUNIER: Yes, Mr. President.
In consequence, an agreement was made with the OKW that
Himmler, Göring, and Ribbentrop should be consulted on the
measures to be taken in this matter. Ribbentrop proposed that any
attack upon German cities should be considered as an act of
terrorism. General Warlimont also, in the name of the OKW,
proposed two means: Lynching and what he called
Sonderbehandlung or special treatment, which consisted in
delivering the parties concerned to the Sicherheitsdienst where they
were subjected to diverse treatments, one of the most notorious
being the well known Kugel action, of which the Tribunal has already
heard and which was simply a way of doing away with those in
question. Document Number 735-PS (Exhibit Number RF-1452) was
submitted to this effect on 9 January 1946.
On 17 June 1944 Keitel wrote to Göring to ask him to approve
the definition of acts of terrorism drawn up by Warlimont. On 19
June 1944 Göring replied through his aide-de-camp that the
population should be forbidden to act as it had done against enemy
airmen and that these enemy airmen should be brought to trial,
since the Allied Governments had forbidden their airmen to commit
acts of terrorism. I refer here to Document Number 732-PS, which I
submit to the Tribunal under Exhibit Number RF-1405.
Consequently, I draw the Tribunal’s attention to this document,
dated 26 June 1944, where Reich Marshal Göring declared that he
would support the taking of judicial action against these airmen.
Remember this date, 19 June 1944, because it is important.
But on 26 June 1944 the Defendant Göring’s aide-de-camp
telephoned to the OKW headquarters staff, who had insisted upon a
definite reply, and notified them that his chief, Reich Marshal Göring,
was in agreement with their definition of acts of terrorism and the
procedure proposed which, as I recall it, included two alternatives:
The handing over of those in question for Sonderbehandlung or their
immediate execution. I refer to Document Numbers 733-PS and 740-
PS, cited on 30 January 1946 by the French Prosecution, under the
Exhibit Numbers RF-374 and RF-375 (Exhibit Numbers RF-1423 and
RF-1424).
 In a memorandum dated 4 July 1944 Hitler made it known that
since the British and the Americans had bombed small towns of no
military importance as a reprisal for V-1, he was asking the German
radio and press to announce that all enemy airmen shot down in an
attack of that kind would be put to death as soon as they were
caught. Such are the facts found in these absolutely irrefutable
documents, and if I cited in detail the reply made on 19 June 1944
by the Defendant Göring, or to be more exact, by his aide-de-camp,
it is because I am anxious to introduce into the proceedings the
documents concerning this question in their entirety.
But I see that in spite of the existence of the order of 19 June
1944 I am obliged to infer the full responsibility of the Defendant
Hermann Göring.
In fact, the Defendant Hermann Göring states that he never
agreed to these measures, and that Captain Breuer, who telephoned
to the General Staff of the OKW, acted—according to the Defendant
Göring—without having previously consulted him. Göring added, in
the statements which he made, that he could not be held
responsible for all the absurd or insignificant actions carried out by
his subordinates.
But, Gentlemen, without even reference to the famous
Leadership Principle—for I see no reason to apply German law to the
accused in any way—the Defendant Göring is in any case responsible
in his capacity as leader. Responsibility begins with authority.
Moreover, what did he do to stop the massacre of airmen by people
whom he had ordered to do the opposite, according to orders which
it was forbidden to formulate in writing?
Even if we consider the position which he takes up in the order
dated 19 June 1944, to which I have referred as establishing
accurately his views at that date on the massacre of airmen and
parachutists, we are compelled to see that at that date, 19 June
1944, even in Germany, the most shortsighted knew that the
German forces would soon succumb to the weight of the Allied
Armies.
Allied aviators were put to death in Germany throughout the war.
Moreover, if the Defendant Hermann Göring maintains that the letter
of 19 June 1944 was written by his aide-de-camp, he is obliged to
admit that the letter of 26 June 1944, also written by the aide-de-
camp, can be imputed to him, although signed by one of his
subordinates. We consider, then, that this document signed by an
aide-de-camp involves Göring as much as if he had signed it himself.
Mr. President and Gentlemen, I shall not enlarge upon the
responsibility of the Defendant Göring for compulsory labor, but I
respectfully beg the Court to refer in due course to certain rays of
illumination that I have tried to indicate in this brief in order to
clarify the position of the defendant in this matter.
I shall make no further mention of the employment of prisoners
of war and internees from concentration camps, which I detailed on
Page 10 of my brief. I should like simply to say a word concerning
economic pillaging and the pillaging of art treasures. These
questions are dealt with at the bottom of Page 11 of my brief.
Concerning economic pillage, Gentlemen, I shall not stress the
considerable part played by the Defendant Göring as leader of the
Four Year Plan in all the measures which contributed to strip literally
all the western countries of their substance. I shall simply point out
one fact which, I believe, has not yet been brought to your
knowledge but which is found in the next to the last subheading on
Page 12. This fact is the following: After the Armistice in 1940, the
Defendant Göring had brought about through Roechling, the official
sequestrator, the cession to the Hermann Göring Werke of all the
factories of Lorraine belonging to the family of Wendel.
This is connected with all the operations of economic pillaging
about which the Economic Section of the French Prosecution have
already informed the Court. With regard to this, the Court will not
fail to realize that the Defendant Göring shares jointly with the
Defendants Rosenberg, Ribbentrop, and Seyss-Inquart—for the
Netherlands—the responsibility for this spoliation.
With regard to the pillaging of works of art, Gentlemen, we have
documents which permit us to draw our conclusions with regard to
this matter which is obviously an unpleasant one for a man who has
occupied the position of the Defendant Göring, namely, that a part of
the works of art and objects of value which were pillaged from the
western countries were reserved for him without any kind of
compensation. I shall not discuss the exact meaning of this act in
municipal law; I leave it to the Tribunal to apply the proper legal
terms for this matter, when it delivers its judgment. But what I
should like to say today is that the appropriation of works of art by
the Defendant Hermann Göring for his private purposes is proved in
documents which cannot be contested and which have already been
submitted to the Tribunal. I refer particularly to Exhibit Number USA-
368 (Document Number 141-PS) submitted on 18 December 1945.
This document was submitted by the Economic Section of the French
Prosecution under the Exhibit Number RF-1309.
I may rapidly recall that this document prescribes that works of
art brought to the Louvre are to be classified in a certain way:
“Firstly, those works of art of which the Führer reserved the right to
dispose of himself. Secondly, those works of art destined to complete
the collection of the Reich Marshal”—et cetera.
I won’t read the rest of the document.
What followed these levies or these privative appropriations? Did
the Defendant Göring pay anything for these? The opposite seems to
be the case; for in the interrogation of the Defendant Rosenberg,
which was given under the Exhibit Number RF-1332 and to which I
referred in the course of the hearing, it is pointed out that the
Defendant Göring made his selection from the works of art
assembled by Rosenberg’s staff and made no corresponding
payment to the Reich treasury.
Not to abuse the patience of the Tribunal, I respectfully beg it to
go back to Page 10 of the transcript previously cited, where it will
see the part played by the Defendant Göring in the appropriation of
works of art, and the fact that no money was paid in compensation.
I simply emphasize, in passing, that at the top of Page 11 you
will find this statement, in reply to a question asked by Colonel
Hinkel. Colonel Hinkel said this to him.
THE PRESIDENT: You are referring to Page 10 and Page 11 of
which document?
 M. MOUNIER: Page 11, Mr. President, of Document Number ECH-
25, which was submitted yesterday under the Exhibit Number RF-
1331, by my colleague M. Gerthoffer. It is not there, for reasons
which I have already pointed out to the Tribunal.
Colonel Hinkel, at the bottom of Page 10, asked the following
question:
“Well, doesn’t that letter state in the last paragraph that you don’t
think that Göring should pay for these articles that he had selected
because he was going to put these articles in an art gallery?”
The reply of the Defendant Rosenberg:
“Not exactly. I would like to add the following:”—which I consider
important—“I was rather uneasy when at the outset I heard art
treasures which the Einsatzstab had sent to Germany. . . .”
That is all, Gentlemen, I won’t say anything more. I merely want to
point out to you the annoyance which the chief of the Einsatzstab
himself felt on learning this fact.
Mr. President, Gentlemen, in regard to the participation of the
Defendant Göring in Crimes against Humanity, particularly the
concentration camps, I shall not insist; but I shall ask the Tribunal,
when they have time, to refer to a few paragraphs in which I briefly
recall the question. But there is a document which, as far as I know,
has not been submitted to the Tribunal and which I should like to
submit today. It concerns pseudo-medical experiments which I
believe have not yet been discussed.
You have frequently been told of Dr. Rascher’s experiments in the
exposure of certain persons to alternate heat and cold, but there is a
question which I treat on Page 17 of my brief and which concerns
the document which I submit today as Exhibit Number RF-1427. This
is a document which originally had the Number L-170. It is a report
made by Major Leo Alexander of the United States Army, on an
institution known as the Kaiser Wilhelm Institut. Major Leo
Alexander, at the time of the defeat of Germany by the Allied Forces,
had to conduct certain investigations. He conducted one in
connection with experiments made by Dr. Rascher and another in
connection with these carried out in the Kaiser Wilhelm Institut. This
report which I submit to the Tribunal is entitled, “Neuropathology in
Wartime Germany.” This Kaiser Wilhelm Institut was an institute
designed for cerebral research. This institution had formerly been in
Berlin-Buch (Page 18 in my brief) and was split up into three
establishments, the first in Munich—I pass over the one in Munich—
the third in Göttingen. The second, the one which interests me, was
established at Dillenburg, in Hessen-Nassau, where there was a
department for special pathology directed by Dr. Hallervorden. What
is interesting, Mr. President. . . .
THE PRESIDENT: Could we see the original?
M. MOUNIER: The original? Here it is, Mr. President.
THE PRESIDENT: Is the series “L” referred to in Major Coogan’s
affidavit?
M. MOUNIER: Mr. President, I should like to point out that this
Number L-170 is the same as that referring to that same Major Leo
Alexander’s document book concerning the experiments of Dr.
Rascher. It is the same number. . . .
THE PRESIDENT: As this document has already been produced in
evidence in the series “L”—it is L-170 I think—the Tribunal will treat
it for the moment as being in evidence and will further consider its
admissibility.
M. MOUNIER: Yes, sir. At all events, I should like to remind the
President, who has certainly noticed it, that I reproduce in this brief,
which has already been communicated to the Defense, the passage
which I regard as relevant to my brief. The passage is quoted in full
in my brief.
THE PRESIDENT: [Turning to Dr. Stahmer.] Yes, we will listen to
you in a few minutes.
[Turning to M. Mounier.] Which passage do you wish to refer to?
M. MOUNIER: Pages 20 and 21 in my brief.
THE PRESIDENT: Yes, do you wish to read them?
M. MOUNIER: I accept the decision of the Tribunal. If the Court
considers this reading superfluous, I shall limit myself to pointing out
that what I find striking in this document is the manner in which Dr.
Hallervorden ordered the delivery of brains for examination when he
says:
“ ‘I had heard that they were going to do that.’ ”—That is, to say, to
kill some sick people in different establishments by means of carbon-
monoxide.—Dr. Hallervorden explained to his American interrogator,
Major Alexander.
“ ‘. . . I went up to them and told them “Look here now, boys, if you
are going to kill all these people, at least take the brains out so that
the material could be utilized.
“ ‘They asked me, “How many can you examine?” and so I told them
an unlimited number—the more the better. I gave them the fixatives,
jars and boxes, and instructions for removing and fixing the
brains. . . .’ ”
I call the attention of the Tribunal to the truly horrible nature of
the measures taken in regard to the people who were to be killed
merely to have their brains examined, for they were, so he said,
“ ‘. . . selected from the various wards of the institutions according to
an excessively simple and quick method. Most institutions did not
have enough physicians, and what physicians there were were either
too busy or did not care, and they delegated the selection to the
nurses and attendants. Whoever looked sick or was otherwise a
problem patient from the nurses’ or attendants’ point of view, was
put on a list and was transported to the killing center. The worst
thing about this business was that it produced a certain brutalization
of the nursing personnel. They got to simply picking out those whom
they did not like, and the doctors had so many patients that they did
not even know them, and put their names on the list.’ ”
I shall stop my citation there, Mr. President, but what I should
like to do subsequently, unless the Tribunal is going to call upon Dr.
Stahmer to speak. . . .
THE PRESIDENT: Yes, we are now going to hear what Dr.
Stahmer wants to say.
 DR. OTTO STAHMER (Counsel for the Defendant Göring): I am
sorry that I must contradict what has just been said, for there is no
proof that these things took place or that the Defendant Göring is
responsible. The Defendant Göring states that he was quite unaware
of these events and that he had nothing whatever to do with
matters of that kind. As far as I know, the Prosecution itself. . . .
THE PRESIDENT: I have to interrupt you, Dr. Stahmer. You will
have a full opportunity of presenting arguments to us to show that
the evidence which is adduced, which is brought forward now
against the Defendant Göring, has really no reference to him. You
will have a full opportunity to do that at the appropriate stage when
you present the defense. The only question we are considering now,
the technical question, is whether this document is a document
which is admissible. We are considering it, of course, but it is not the
appropriate time for you to present your argument that the
document does not refer to Göring and that Göring had no
knowledge of it. That will be your defense. It isn’t an objection to
the admissibility of the document. It is an argument to show that
Göring didn’t know anything about the document and didn’t know
anything about the experiments.
Do you understand what I mean?
DR. STAHMER: Yes, sir.
M. MOUNIER: Mr. President, I only wanted, by introducing. . . .
THE PRESIDENT: Yes, M. Mounier, continue.
M. MOUNIER: Mr. President, I take leave to point out to you that
my friend, Mr. Elwyn Jones, has just pointed out to me that this is
admitted as proof in view of the conditions under which it was
submitted. This is the document entitled, “Neuropathology and
Neurophysiology, including Electroencephalography, in Wartime
Germany.” Besides this reference is found in the English copy which I
submitted in the modest document book which I submitted to the
Tribunal just now. I should like to tell you, Mr. President, in citing this
short passage. . . .
THE PRESIDENT: Maybe the Tribunal had better keep the original
document for the present.
 M. MOUNIER: My aim, Mr. President, in citing this short passage,
is to demonstrate the truly atrocious way in which they treated
people in order to procure the necessary material for these so-called
experiments. According to the Prosecution this relates to Hermann
Göring, for the Tribunal will take into account the fact that these
experiments were made for the purpose of obtaining information of
a scientific or pseudo-scientific nature concerning the effects upon
the brains of airmen of all the accidents which might happen to
them.
These experiments are connected with those of Dr. Rascher,
concerning which some correspondence took place. The Defendant
Hermann Göring cannot have been ignorant of this correspondence,
for it directly concerned the Air Force, which he commanded. I cite,
for instance, a letter dated 24 October 1942, which was addressed
by Himmler to Dr. Rascher and which I submit to the Tribunal under
the Exhibit Number RF-1409 (Document Number 1609-PS).
To save the time of the Tribunal I shall not read this letter. I shall
simply refer to another document which has already been cited as
Document Number 343-PS. It was submitted by the American
Prosecution as Exhibit Number USA-463, 20 December 1945 (Exhibit
Number RF-1428), and it is a letter which proves that as early as 20
May 1942 Field Marshal Milch was charged by the Defendant Göring
with the task of transmitting to the SS his special thanks for the aid
which they had given the Luftwaffe with these pseudo-medical
experiments. Consequently, we consider that in this respect the
responsibility of the Defendant Hermann Göring is clearly
established.
Mr. President and Gentlemen, I have concluded the points
concerning the Defendant Hermann Göring to which I wanted to
draw the attention of the Tribunal. There is a conclusion in my brief
against the Defendant Hermann Göring. With the permission of the
Tribunal I shall not read it. I shall say that this conclusion is an
extract from an old book dating from 1669, which is certainly known
to everyone in Germany at least. Its title is Simplizius Simplizissimus
by Grimmelshausen. It is a work in which persons are seen invoking
dreams. Unfortunately the realization seems to have been achieved
by the National Socialist regime.
I now go on to the Defendant Seyss-Inquart, whose case
concerns most particularly our friends in the Netherlands on behalf
of whom France is acting as counsel.
Consequently, Mr. President and Gentlemen, as regards the
Defendant Seyss-Inquart, the French Prosecution is going to outline
as briefly as possible both in the name of the Netherlands
Government and in its own name the separate charges against this
defendant. The part played by the Defendant Seyss-Inquart, his
participation in the annexation of Austria, were carefully studied
during the course of this Trial. But it is his operations in Holland
which deserve to be thrown into special relief today.
On 13 May 1940 the Netherlands Government left Holland for a
friendly Allied country. Its presence there was indicative of its firm
determination not to yield up in any way its sovereign rights.
On 29 May 1940 the Defendant Seyss-Inquart, who had the rank
of Reich Minister without Portfolio, was appointed Reich
Commissioner for the occupied Netherlands. The Defendant Seyss-
Inquart has therefore been considered responsible, by virtue of his
functions, for all the acts committed by the so-called German Civil
Government from that date up to the capitulation of the German
Army. The speeches which he made afford evidence that he was
invested not only with purely administrative functions but also with
political authority.
It is, therefore, useless for him to try, as he did when he was
interrogated by my friend Mr. Thomas Dodd, to maintain that in
Holland he was nothing more than an official empowered to put his
seal on orders, in the same way that in Austria earlier he was
practically only a telegraph operator. This interrogation is dated 18
September 1945, Pages 20 to 22. I do not insist further, as I did not
wish to produce these interrogations in order to avoid wasting the
time of the Court with the numerous interrogations which would
have had to be cited in cross-examination, and these documents will
really remain for the edification of the Court.
THE PRESIDENT: M. Mounier, has the interrogation been put in?
 M. MOUNIER: No, Mr. President.
THE PRESIDENT: Well, as a matter of technical procedure. . . .
M. MOUNIER: I know in advance that you cannot accept this as
proof already constituted in your eyes, considering the rule. . . .
THE PRESIDENT: Yes, it can be given if the rule is complied with.
M. MOUNIER: My intention, Mr. President, is the following—to
state. . . .
THE PRESIDENT: M. Mounier, I think you are misunderstanding
me. Under the article the prosecutors have got the right to
interrogate any of the defendants, and this was an interrogation of
one of the defendants.
If the Prosecution choose to do so, they can offer their
interrogation in evidence. If they do not choose to do so, they need
not do so. Under such circumstances the interrogation is not in
evidence, and need not be furnished to the defendant until it is.
M. MOUNIER: Yes, Mr. President, I have not alluded to these
statements made by the defendant. I simply wish to point out that
when the defendant of whom I am now speaking is cross-examined,
we shall be able to confront him with the statements he made, or, at
least, I hope so.
With the permission of the Court I shall first take up the subject
of the Defendant Seyss-Inquart’s terrorist activities. These are
shown by the following measures:
First, a whole system of collective fines. In March 1941 he
established a system of collective fines which were imposed upon
the Dutch cities where he thought that elements of the resistance
movement existed. Thus the city of Amsterdam had to pay a fine of
two and a half million.
The Defendant Seyss-Inquart also established a system of
hostages. On 18 May 1942 he published a proclamation announcing
the arrest of 450 persons in important official positions, who were
only suspected of being in relation with the resistance movement.
In fact, the defendant has admitted before Mr. Dodd. . . . No, I
stop, Mr. President, I did not submit these interrogations. I shall pass
over this passage and only point it out in a general way, and I beg
the Court not to consider this fact as an infringement of the Charter.
I am simply pointing out to the Court that in this case, too, the
Defendant Seyss-Inquart tried to hide behind the shadow of the
Reich Chancellor, the shadow of the Führer, Hitler.
By the decree of 7 July 1942, the defendant ordered that the
German tribunals, the judges of which he himself appointed, were to
try not only the German citizens in Holland, but also citizens
suspected of activities hostile to the Reich, to the Nazi Party, or to
the German people.
At the same time the Defendant Seyss-Inquart introduced the
death penalty for those who had not properly performed the security
jobs assigned them by the Wehrmacht or the Security Police or who
had failed to inform the German command posts of all criminal
projects directed against the occupation forces which came to their
knowledge.
THE PRESIDENT: M. Mounier, you were citing then a
proclamation dated 18 May 1942. You did not give us any number as
yet.
M. MOUNIER: Mr. President, I ought to say that I am referring in
a general way to the official report of the Netherlands Government
(Document Number RF-1429). The government submitted a
report. . . .
THE PRESIDENT: Is it stated there?
M. MOUNIER: Yes, Mr. President.
THE PRESIDENT: Did that also apply to the document of 7 July
1942 that you just spoke of?
M. MOUNIER: Yes, Mr. President. The Defendant Seyss-Inquart
also appointed the SS Obergruppenführer Rauter, General
Commissioner for Security. The latter is responsible for the murder of
thousands of Dutchmen executed with the passive consent of Seyss-
Inquart, inasmuch as Rauter’s appointment was always maintained
and was never terminated.
On the other hand, the Netherlands Government charges the
Defendant Seyss-Inquart with the creation of a whole series of
exceptional courts. In May 1943 he established summary police
jurisdiction, and in fact through an ordinance issued by Hitler, Dutch
prisoners of war who had been freed shortly after the cessation of
hostilities were once more interned. A tough resistance showed itself
in the Dutch factories and the newly established summary
jurisdiction sentenced several Dutch citizens who were executed.
Moreover, Seyss-Inquart did not fail to boast of all these terrorist
measures at a meeting of Dutch collaborators and claimed
responsibility for them.
The Defendant Seyss-Inquart was Hitler’s supreme representative
in Holland. He should be considered as responsible, along with the
Defendant Sauckel, for the mass deportation of workers from
Holland to the Reich between 1940 and 1945. Whether or not the
German military authorities played any part themselves in the
mobilization of labor, Sauckel’s officials in Holland were normally
placed under the authority of the Reich Commissioner Seyss-Inquart,
and he must be considered as responsible for their actions. It was
the Defendant Seyss-Inquart who signed the decree of the Reich
Commissioner, Number 26 of 1942, which is found in the official
Dutch report, in an official publication ordering the compulsory
transport of Dutch labor to Germany. Those who would not work for
Germany got nothing to eat; the occupation authorities even went
so far as to make huge roundups in the streets of Rotterdam and
The Hague in order to procure labor for the fortifications of the
Wehrmacht.
In regard to economic pillage during the Defendant Seyss-
Inquart’s period of office as Commissioner, the Dutch economic
system was plundered like that of the other occupied countries. In
the winter of 1941-42 woollen goods were requisitioned by order of
Seyss-Inquart for the German Army on the Eastern front. In 1943
textiles and every-day household articles were requisitioned for the
benefit of the bombed-out German population. Under what the
occupation authorities called the “Action Böhm,” people of the
Netherlands were compelled to sell wines and various objects
destined to form gifts for the German population for the celebration
of Christmas 1943.
The same thing happened with regard to the organization of the
black market, for, in order to carry out the Four Year Plan, Seyss-
Inquart gave the Defendant Göring and the Defendant Speer
competent assistance in the pillage of the Dutch economic system.
We can say in this way that a huge black market was fostered and
maintained. The Four Year Plan utilized “snatchers” for these alleged
purchases but when Dutch prosecutors tried to intervene they were
prevented from doing so by the German police.
In 1940 the Defendant Seyss-Inquart issued an ordinance
permitting the German authorities in Holland to confiscate the
property of all persons who could be accused of hostile activities
against the German Reich. The property of the royal family was, on
the Defendant Seyss-Inquart’s orders, confiscated by the General
Commission for Security. The occupation troops could help
themselves to everything that was of use to them.
This pillage was manifested in a particularly cruel manner by the
abuses which went on in connection with the requisition of food
products.
In fact, the official report of the Dutch Government and the
document already submitted by the Economic Section of the French
Prosecution under Document Number RF-139 (Exhibit Number RF-
139), and Document Number RF-140 (Exhibit Number RF-140) show
that, from the very beginning of the occupation, food stocks were
systematically removed with the consent of Seyss-Inquart—as was
also the case with agricultural produce, which was transported to
Germany. When a railway strike broke out in the north in September
1944, soon after the liberation of southern Holland, Seyss-Inquart, in
order to break the strike, gave orders that no food stocks were to be
moved from the northeast to the West. As a result of this, it was
impossible to establish food stocks in the West.
Consequently, Seyss-Inquart must also be held responsible for
the famine which ensued during the winter of 1944-45, causing the
death of some 25,000 Dutchmen.
In regard to works of art, the pillage was carried on in the same
way. The Defendant Seyss-Inquart must be considered responsible
for organizing the removal of works of art from Holland, since he
expressly called in his friend, Dr. Mühlmann, who was a specialist in
this branch.
 In this connection I refer to the document submitted by the
Economic Section of the French Prosecution under Document
Numbers RF-1343 and RF-1344. The Defendant Seyss-Inquart issued
a whole series of measures contrary to international law which did
considerable harm to the Netherlands.
In 1941 the Dutch authorities had established a currency control
system which allowed them to keep track of purchases made with
German money, either of goods or public funds, with the aim of
preventing abuses which would lead to the plundering of Holland’s
wealth in the form of materials or of currency.
On 31 March 1941 the Defendant Seyss-Inquart abolished the
“currency” frontier existing between the Reich and the occupied
Dutch territory. By so doing, he paved the way for all the abuses
committed in monetary matters by the occupying power, in addition
to the impossible sums demanded by Germany to defray the
expenses of occupation: 500 million Reichsmark on 24 March 1941.
The frontier control between Dutch occupied territory and
Germany was also abolished by order of Göring, in order to expedite
the pillage of the Netherlands’ economic system. When the war
began to go badly for the Wehrmacht, especially after 1 September
1944, the destruction became systematic. The objectives aimed at
by the Germans in the Netherlands were the following: First, to
demolish or put out of action factories, shipyards, basins and docks,
port installations, mines, bridges, railway equipment. Second, to
flood the western parts of Holland. Third, to seize raw materials,
semi-manufactured products, manufactured goods and machines,
sometimes by requisitioning, sometimes in return for payment in
money, but in many cases simply by force of arms. Fourth, to break
open safe-deposits containing securities, diamonds, et cetera, and to
take illegal possession of these. The result of these measures,
responsibility for which devolves wholly or to a great extent on the
Defendant Seyss-Inquart, was to throw Holland into a state of
unspeakable and undeserved misery.
I have now concluded, Mr. President, the case of the Defendant
Seyss-Inquart.