Interpretable and Annotation

Download as pdf or txt
Download as pdf or txt
You are on page 1of 51

Full download test bank at ebook textbookfull.

com

Interpretable and Annotation


Efficient Learning for Medical
Image Computing Third International
Workshop iMIMIC 2020 Second
International Workshop MIL3ID 2020
CLICK LINK TO DOWLOAD
and 5th International Workshop
https://textbookfull.com/product/interpretabl
LABELS 2020 Held in Conjunction
e-and-annotation-efficient-learning-for-
with MICCAI 2020 Lima Pe Jaime
medical-image-computing-third-international-
workshop-imimic-2020-second-international-
workshop-mil3id-2020-and-5th-international-
workshop-labels-202/

textbookfull
Jaime Cardoso · Hien Van Nguyen ·
Nicholas Heller et al. (Eds.)
LNCS 12446

Interpretable and Annotation-


Efficient Learning for Medical
Image Computing
Third International Workshop, iMIMIC 2020
Second International Workshop, MIL3iD 2020
and 5th International Workshop, LABELS 2020
Held in Conjunction with MICCAI 2020
Lima, Peru, October 4–8, 2020, Proceedings
Lecture Notes in Computer Science 12446

Founding Editors
Gerhard Goos
Karlsruhe Institute of Technology, Karlsruhe, Germany
Juris Hartmanis
Cornell University, Ithaca, NY, USA

Editorial Board Members


Elisa Bertino
Purdue University, West Lafayette, IN, USA
Wen Gao
Peking University, Beijing, China
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Gerhard Woeginger
RWTH Aachen, Aachen, Germany
Moti Yung
Columbia University, New York, NY, USA
More information about this series at http://www.springer.com/series/7412
Jaime Cardoso Hien Van Nguyen
• •

Nicholas Heller et al. (Eds.)

Interpretable and Annotation-


Efficient Learning for Medical
Image Computing
Third International Workshop, iMIMIC 2020
Second International Workshop, MIL3iD 2020
and 5th International Workshop, LABELS 2020
Held in Conjunction with MICCAI 2020
Lima, Peru, October 4–8, 2020
Proceedings

123
Editors
Jaime Cardoso Hien Van Nguyen
University of Porto University of Houston
Porto, Portugal Houston, TX, USA
Nicholas Heller
University of Minnesota
Minneapolis, MN, USA

Additional Workshop Editors see next page

ISSN 0302-9743 ISSN 1611-3349 (electronic)


Lecture Notes in Computer Science
ISBN 978-3-030-61165-1 ISBN 978-3-030-61166-8 (eBook)
https://doi.org/10.1007/978-3-030-61166-8
LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics

© Springer Nature Switzerland AG 2020


This work is subject to copyright. All rights are reserved 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
Additional Workshop Editors

iMIMIC 2020 Editors


Pedro Henriques Abreu Ivana Isgum
University of Coimbra Amsterdam University Medical Center,
Coimbra, Portugal Amsterdam, The Netherlands

Wilson Silva Ricardo Cruz


University of Porto University of Porto
Porto, Portugal Porto, Portugal

Jose Pereira Amorim


University of Coimbra
Coimbra, Portugal
MIL3iD 2020 Editors
Vishal Patel Badri Roysam
Johns Hopkins University University of Houston
Baltimore, MD, USA Houston, TX, USA
Kevin Zhou Steve Jiang
Chinese Academy of Sciences UT Southwestern Medical Center
Beijing, China Dallas, TX, USA
Ngan Le Khoa Luu
University of Arkansas University of Arkansas
Fayetteville, AR, USA Fayetteville, AR, USA
LABELS 2020
Raphael Sznitman Veronika Cheplygina
University of Bern Eindhoven University of Technology
Bern, Switzerland Eindhoven, The Netherlands
Diana Mateus Emanuele Trucco
Technical University of Munich University of Dundee
Nantes, Germany Dundee, UK
Samaneh Abbasi
Eindhoven University of Technology
Eindhoven, The Netherlands
iMIMIC 2020 Preface

It is our genuine honor and great pleasure to welcome you to the Third Workshop on
Interpretability of Machine Intelligence in Medical Image Computing (iMIMIC 2020),
a satellite event at the 23rd International Conference on Medical Image Computing and
Computer Assisted Intervention (MICCAI 2020). Following in the footsteps of the two
previous successful meetings in Granada, Spain (2018) and Shenzhen, China (2019),
we gathered for this new edition.
iMIMIC is a single-track, half-day workshop consisting of high-quality, previously
unpublished papers, presented either orally or as a poster, intended to act as a forum for
research groups, engineers, and practitioners to present recent algorithmic develop-
ments, new results, and promising future directions in interpretability of machine
intelligence in medical image computing. Machine learning systems are achieving
remarkable performances at the cost of increased complexity. Hence, they become less
interpretable, which may cause distrust, potentially limiting clinical acceptance. As
these systems are pervasively being introduced to critical domains, such as medical
image computing and computer assisted intervention, it becomes imperative to develop
methodologies allowing insight into their decision making. Such methodologies would
help physicians to decide whether they should follow and trust automatic decisions.
Additionally, interpretable machine learning methods could facilitate defining the legal
framework of their clinical deployment. Ultimately, interpretability is closely related to
AI safety in healthcare.
This year’s iMIMIC was held on October 4, 2020, virtually in Lima, Peru, and was
hosted by INESC TEC and the University of Coimbra, with the support of University
of Porto and CISUC, all located in Portgual. There was a very positive response to the
call for papers for iMIMIC 2020. We received 18 full papers from 10 countries and 8
were accepted for presentation at the workshop, where each paper was reviewed by at
least three reviewers. The accepted papers present fresh ideas of Interpretability in
settings such as regression, multiple instance learning, weakly supervised learning,
local annotations, classifier re-training, and model pruning.
The high quality of the scientific program of iMIMIC 2020 was due first to the
authors who submitted excellent contributions and second to the dedicated collabora-
tion of the International Program Committee and the other researchers who reviewed
the papers. We would like to thank all the authors for submitting their contributions and
for sharing their research activities.
We are particularly indebted to the Program Committee members and to all the
reviewers for their precious evaluations, which permitted us to set up this publication.
viii iMIMIC 2020 Preface

We were also very pleased to benefit from the participation of the invited speakers
Himabindu Lakkaraju, Harvard University, USA, and Wojciech Samek, Fraunhofer
HHI, Germany. We would like to express our sincere gratitude to these world-
renowned experts.

October 2020 Jaime Cardoso


Pedro Henriques Abreu
Ivana Isgum
Wilson Silva
Ricardo Cruz
Jose Pereira Amorim
iMIMIC 2020 Organization

General Chairs iMIMIC 2020


Jaime Cardoso INESC TEC and University of Porto, Portugal
Pedro Henriques Abreu CISUC and University of Coimbra, Portugal
Ivana Isgum Amsterdam University Medical Center,
The Netherlands

Publicity Chair iMIMIC 2020


Jose Pereira Amorim CISUC and University of Coimbra, Portugal

Program Chair iMIMIC 2020


Wilson Silva INESC TEC and University of Porto, Portugal

Sponsor Chair iMIMIC 2020


Ricardo Cruz INESC TEC and University of Porto, Portugal

Program Committee iMIMIC 2020


Ben Glocker Imperial College, UK
Bettina Finzel University of Bamberg, Germany
Carlos A. Silva University of Minho, Portugal
Christoph Molnar Ludwig Maximilian University of Munich, Germany
Claes Nøhr Ladefoged Rigshospitalet, Denmark
Dwarikanath Mahapatra Inception Institute of AI, UAE
George Panoutsos The University of Sheffield, UK
Hrvoje Bogunovic Medical University of Vienna, Austria
Isabel Rio-Torto University of Porto, Portugal
Joana Cristo Santos University of Coimbra, Portugal
Kelwin Fernandes NILG.AI, Portugal
Luis Teixeira University of Porto, Portugal
Miriam Santos University of Coimbra, Portugal
Nick Pawlowski Imperial College London, UK
Ricardo Cruz University of Porto, Portugal
Sérgio Pereira Lunit, South Korea
MIL3iD 2020 Preface

Welcome to the Second International Workshop on Medical Image Learning with Less
Labels and Imperfect Data (MIL3iD 2020). The MIL3iD 2020 proceedings contain 11
high-quality papers of 8 pages that were selected through a rigorous peer-review
process.
We hope this workshop will create a forum for discussing best practices in medical
image learning with label scarcity and data imperfection. This forum is urgently needed
because the issues of label noises and data scarcity are highly practical, but largely
under-investigated in the medical image analysis community. Traditional approaches
for dealing with these challenges include transfer learning, active learning, denoising,
and sparse representation. The majority of these algorithms were developed prior to the
recent advances of deep learning and might not benefit from the power of deep
networks. The revision and improvement of these techniques in the new light of deep
learning are long overdue.
This workshop potentially helps answer many important questions. For example,
several recent studies found that deep networks are robust to massive random label
noises but more sensitive to structured label noises. What implication do these findings
have on dealing with noisy medical data? Recent work on Bayesian neural networks
demonstrates the feasibility of estimating uncertainty due to the lack of training data. In
other words, it enables our classifiers to be aware of what they do not know. Such a
framework is important for medical applications where safety is critical. How can
researchers of MICCAI community leverage this approach to improve their systems’
robustness in the case of data scarcity? Our prior work shows that a variant of capsule
networks generalizes better than convolutional neural networks with an order of
magnitude fewer training data. This gives rise to an interesting question: Are there
better classes of networks that intrinsically require less labeled data for learning?
Humans always have an edge over deep networks when it comes to learning with small
amounts of data. However, recent work on one-shot deep learning has surpassed human
in an image recognition task using only a few training samples for each task. Do these
results still hold for medical image analysis tasks?
The proceedings of the workshop are published as a joint LNCS volume alongside
other satellite events organized in conjunction with MICCAI. In addition to the LNCS
volume, to promote transparency, the papers’ reviews and preprints are publicly
available on the workshop website: https://www.hvnguyen.com/lesslabelsimperfect
dataml2020. In addition to the papers, abstracts, slides, and posters presented during
the workshop will be made publicly available on the MIL3iD website.
xii MIL3iD 2020 Preface

We would like to thank all the speakers and authors for joining our workshop, the
Program Committee for their excellent work with the peer reviews, and the workshop
chairs and editors for their help with the organization of the second MIL3iD workshop.

August 2020 Hien Van Nguyen


Vishal Patel
Badri Roysam
Kevin Zhou
Steve Jiang
Ngan Le
Khoa Luu
MIL3iD 2020 Organization

Chairs MIL3iD 2020


Hien Van Nguyen University of Houston, USA
Vishal Patel Johns Hopkins University, USA
Badri Roysam University of Houston, USA
Kevin Zhou Chinese Academy of Sciences, China
Steve Jiang UT Southwestern Medical School, USA
Ngan Le University of Arkansas, USA
Khoa Luu University of Arkansas, USA

Program Committee MIL3iD 2020


Zhoubing Xu Siemens Healthineers, USA
Pengyu Yuan University of Houston, USA
Pengbo Liu Chinese Academy of Sciences, China
Anjali Balagopal UT Southwestern Medical School, USA
Jue Jiang Memorial Sloan Kettering Cancer Center, USA
Jahandar Jahanipour National Institutes of Health, USA
Pietro Antonio Cicalese University of Houston, USA
Aryan Mobiny University of Houston, USA
Pasawee Wirojwatanakul New York University, USA
Samira Zare University of Houston, USA
Pengfei Guo Johns Hopkins University, USA
Xiaoyang Li University of Houston, USA
Jeya Maria Jose Valanarasu Johns Hopkins University, USA
Xiao Liang UT Southwestern Medical School, USA
Li Xiao Chinese Academy of Sciences, China
Siqi Liu Siemens Healthineers, USA
Aditi Singh University of Houston, USA
LABELS 2020 Preface

This volume contains the proceedings of the 5th International Workshop on Large-
scale Annotation of Biomedical data and Expert Label Synthesis (LABELS 2020),
which was held on October 8, 2020, in conjunction with the 23rd International
Conference on Medical Image Computing and Computer Assisted Intervention
(MICCAI 2020), originally planned for Lima, Peru, but ultimately held virtually, due to
the COVID-19 pandemic. The first workshop in the LABELS series was held in 2016
in Athens, Greece. This was followed by workshops in Quebec City, Canada, in 2017,
Granada, Spain, in 2018, and Shenzhen, China, in 2019.
As data-hungry methods continue to drive advancements in medical imaging, the
need for high-quality annotated data to train and validate these methods continues to
grow. Further, with the pressing need to address health disparities and to prevent
learned systems from internalizing biases, there has never been a greater need for
thorough study and discussion of best practices in data collection and annotation. For
the past four years, LABELS has aimed to facilitate exactly this.
Following the success of the previous four LABELS workshops, the fifth workshop
was planned for 2020. This year’s edition of the workshop included invited talks by
Anand Malpani (Johns Hopkins University, USA) and Amber Simpson (Queen’s
University, Canada), as well as several papers and abstracts. After peer review, a total
of 10 papers and 3 abstracts were selected. The papers appear in this volume, and the
abstracts are available on the workshop website: https://miccailabels.org. The research
presented this year ranged from how to quantify and mitigate demographic biases, to
probing the reproducibility of expert labels, to new tools for more efficient annotation
of emerging image modalities. LABELS takes pride in the fact that theoretical novelty
is not a prerequisite for work presented at the workshop, instead the event embraces the
messy, tedious reality of medical image collection and annotation in an effort to expose
and formalize its underlying principles.
We would like to thank all the speakers and authors for joining our workshop, the
Program Committee for their excellent work with the peer reviews, our sponsors –
Retinai and Auris Health – for their support, and the workshop chairs for their help with
the organization of the fifth LABELS workshop.
August 2020 Nicholas Heller
Raphael Sznitman
Veronika Cheplygina
Diana Mateus
Emanuele Trucco
Samaneh Abbasi
LABELS 2020 Organization

Chairs LABELS 2020


Raphael Sznitman University of Bern, Switzerland
Veronika Cheplygina Eindhoven University of Technology (TU/e),
The Netherlands
Diana Mateus Technische Universität München (TUM), Germany
Emanuele Trucco University of Dundee, UK
Samaneh Abbasi Eindhoven University of Technology (TU/e),
The Netherlands
Nicholas Heller University of Minnesota, USA

Program Committee LABELS 2020


Florian Dubost Erasmus University Medical Center, The Netherlands
Amelia Jimenez-Sanchez Pompeu Fabra University, Spain
Obioma Pelka University of Duisburg-Essen, Germany
Christoph Friedrich University of Applied Sciences and Arts Dortmund,
Germany
John Onofrey Yale University, USA
Vinkle Srivastav University of Strasbourg, France
Filipe Condessa Carnegie Mellon University, USA
Roger Tam University of British Columbia, Canada
Weidong Cai The University of Sydney, Australia
Silas Ørting University of Copenhagen, Denmark
Jack Rickman University of Minnesota, USA
Alison O’Neil Cannon Medical Research, UK
Bin Xie Central South University, China
Fausto Milletari NVIDIA, Germany
Holger Roth NVIDIA, USA
Michael Goetz DKFZ, Germany
Nishikant Deshmukh Johns Hopkins University, USA
Tom Eelbode KU Leuven, Belgium
Tuo Leng Shanghai University, China
Wen Hui Lei University of Electronic Science and Technology,
China
Xiaoyu He Central South University, China
Contents

iMIMIC 2020

Assessing Attribution Maps for Explaining CNN-Based Vertebral Fracture


Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Eren Bora Yilmaz, Alexander Oliver Mader, Tobias Fricke, Jaime Peña,
Claus-Christian Glüer, and Carsten Meyer

Projective Latent Interventions for Understanding


and Fine-Tuning Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Andreas Hinterreiter, Marc Streit, and Bernhard Kainz

Interpretable CNN Pruning for Preserving Scale-Covariant Features


in Medical Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Mara Graziani, Thomas Lompech, Henning Müller,
Adrien Depeursinge, and Vincent Andrearczyk

Improving the Performance and Explainability of Mammogram Classifiers


with Local Annotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Lior Ness, Ella Barkan, and Michal Ozery-Flato

Improving Interpretability for Computer-Aided Diagnosis Tools on Whole


Slide Imaging with Multiple Instance Learning and Gradient-Based
Explanations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Antoine Pirovano, Hippolyte Heuberger, Sylvain Berlemont,
Saïd Ladjal, and Isabelle Bloch

Explainable Disease Classification via Weakly-Supervised Segmentation . . . . 54


Aniket Joshi, Gaurav Mishra, and Jayanthi Sivaswamy

Reliable Saliency Maps for Weakly-Supervised Localization


of Disease Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Maximilian Möller, Matthias Kohl, Stefan Braunewell, Florian Kofler,
Benedikt Wiestler, Jan S. Kirschke, Björn H. Menze, and Marie Piraud

Explainability for Regression CNN in Fetal Head Circumference


Estimation from Ultrasound Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Jing Zhang, Caroline Petitjean, Florian Yger, and Samia Ainouz
xx Contents

MIL3iD 2020

Recovering the Imperfect: Cell Segmentation in the Presence


of Dynamically Localized Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Özgün Çiçek, Yassine Marrakchi, Enoch Boasiako Antwi,
Barbara Di Ventura, and Thomas Brox

Semi-supervised Instance Segmentation with a Learned Shape Prior . . . . . . . 94


Long Chen, Weiwen Zhang, Yuli Wu, Martin Strauch, and Dorit Merhof

COMe-SEE: Cross-modality Semantic Embedding Ensemble


for Generalized Zero-Shot Diagnosis of Chest Radiographs . . . . . . . . . . . . . 103
Angshuman Paul, Thomas C. Shen, Niranjan Balachandar,
Yuxing Tang, Yifan Peng, Zhiyong Lu, and Ronald M. Summers

Semi-supervised Machine Learning with MixMatch and Equivalence


Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Colin B. Hansen, Vishwesh Nath, Riqiang Gao, Camilo Bermudez,
Yuankai Huo, Kim L. Sandler, Pierre P. Massion, Jeffrey D. Blume,
Thomas A. Lasko, and Bennett A. Landman

Non-contrast CT Liver Segmentation Using CycleGAN Data Augmentation


from Contrast Enhanced CT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Chongchong Song, Baochun He, Hongyu Chen, Shuangfu Jia,
Xiaoxia Chen, and Fucang Jia

Uncertainty Estimation in Medical Image Localization: Towards Robust


Anterior Thalamus Targeting for Deep Brain Stimulation . . . . . . . . . . . . . . . 130
Han Liu, Can Cui, Dario J. Englot, and Benoit M. Dawant

A Case Study of Transfer of Lesion-Knowledge . . . . . . . . . . . . . . . . . . . . . 138


Soundarya Krishnan, Rishab Khincha, Lovekesh Vig, Tirtharaj Dash,
and Ashwin Srinivasan

Transfer Learning with Joint Optimization for Label-Efficient Medical


Image Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Xintong Li, Huijuan Yang, Zhiping Lin, and Pavitra Krishnaswamy

Unsupervised Wasserstein Distance Guided Domain Adaptation


for 3D Multi-domain Liver Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Chenyu You, Junlin Yang, Julius Chapiro, and James S. Duncan

HydraMix-Net: A Deep Multi-task Semi-supervised Learning Approach


for Cell Detection and Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Raja Muhammad Saad Bashir, Talha Qaiser, Shan E Ahmed Raza,
and Nasir M Rajpoot
Contents xxi

Semi-supervised Classification of Chest Radiographs . . . . . . . . . . . . . . . . . . 172


Eduardo H. P. Pooch, Pedro Ballester, and Rodrigo C. Barros

LABELS 2020

Risk of Training Diagnostic Algorithms on Data with Demographic Bias. . . . 183


Samaneh Abbasi-Sureshjani, Ralf Raumanns, Britt E. J. Michels,
Gerard Schouten, and Veronika Cheplygina

Semi-weakly Supervised Learning for Prostate Cancer Image Classification


with Teacher-Student Deep Convolutional Networks . . . . . . . . . . . . . . . . . . 193
Sebastian Otálora, Niccolò Marini, Henning Müller,
and Manfredo Atzori

Are Pathologist-Defined Labels Reproducible? Comparison of the


TUPAC16 Mitotic Figure Dataset with an Alternative Set of Labels . . . . . . . 204
Christof A. Bertram, Mitko Veta, Christian Marzahl,
Nikolas Stathonikos, Andreas Maier, Robert Klopfleisch,
and Marc Aubreville

EasierPath: An Open-Source Tool for Human-in-the-Loop Deep Learning


of Renal Pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Zheyu Zhu, Yuzhe Lu, Ruining Deng, Haichun Yang, Agnes B. Fogo,
and Yuankai Huo

Imbalance-Effective Active Learning in Nucleus, Lymphocyte and Plasma


Cell Detection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Chao-Ting Li, Hung-Wen Tsai, Tseng-Lung Yang, Jung-Chi Lin,
Nan-Haw Chow, Yu Hen Hu, Kuo-Sheng Cheng, and Pau-Choo Chung

Labeling of Multilingual Breast MRI Reports . . . . . . . . . . . . . . . . . . . . . . . 233


Chen-Han Tsai, Nahum Kiryati, Eli Konen, Miri Sklair-Levy,
and Arnaldo Mayer

Predicting Scores of Medical Imaging Segmentation Methods


with Meta-learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
Tom van Sonsbeek and Veronika Cheplygina

Labelling Imaging Datasets on the Basis of Neuroradiology Reports:


A Validation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
David A. Wood, Sina Kafiabadi, Aisha Al Busaidi, Emily Guilhem,
Jeremy Lynch, Matthew Townend, Antanas Montvila, Juveria Siddiqui,
Naveen Gadapa, Matthew Benger, Gareth Barker, Sebastian Ourselin,
James H. Cole, and Thomas C. Booth
xxii Contents

Semi-supervised Learning for Instrument Detection with a Class


Imbalanced Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Jihun Yoon, Jiwon Lee, SungHyun Park, Woo Jin Hyung,
and Min-Kook Choi

Paying Per-Label Attention for Multi-label Extraction from


Radiology Reports. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Patrick Schrempf, Hannah Watson, Shadia Mikhael, Maciej Pajak,
Matúš Falis, Aneta Lisowska, Keith W. Muir, David Harris-Birtill,
and Alison Q. O’Neil

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291


iMIMIC 2020
Assessing Attribution Maps for
Explaining CNN-Based Vertebral
Fracture Classifiers

Eren Bora Yilmaz1,4(B) , Alexander Oliver Mader3 , Tobias Fricke2 ,


Jaime Peña1 , Claus-Christian Glüer1 , and Carsten Meyer3,4,5
1
Section Biomedical Imaging, Department of Radiology and Neuroradiology,
University Hospital Schleswig-Holstein (UKSH), Campus Kiel, Kiel, Germany
[email protected]
2
Department of Radiology and Neuroradiology, UKSH, Campus Kiel, Kiel, Germany
3
Institute of Computer Science, Kiel University of Applied Sciences, Kiel, Germany
4
Department of Computer Science, Faculty of Engineering, Kiel University,
Kiel, Germany
5
Department of Digital Imaging, Philips Research, Hamburg, Germany

Abstract. Automated evaluation of vertebral fracture status on com-


puted tomography (CT) scans acquired for various purposes (opportunis-
tic CT) may substantially enhance vertebral fracture detection rate. Con-
volutional neural networks (CNNs) have shown promising performance
in numerous tasks but their black box nature may hinder acceptance by
physicians. We aim (a) to evaluate CNN architectures for osteoporotic
fracture discrimination as part of a pipeline localizing and classifying
vertebrae in CT images and (b) to evaluate the benefit of using attri-
bution maps to explain a network’s decision. Training different model
architectures on 3D patches containing vertebrae, we show that CNNs
permit highly accurate discrimination of the fracture status of individual
vertebrae. Explanations were computed using selected attribution meth-
ods: Gradient, Gradient * Input, Guided BackProp, and SmoothGrad
algorithms. Quantitative and visual tests were conducted to evaluate the
meaningfulness of the explanations (sanity checks). The explanations
were found to depend on the model architecture, the realization of the
parameters, and the precise position of the target object of interest.

Keywords: Explainable AI · Sanity checks · Osteoporosis

1 Introduction

Computed tomography (CT) images are taken for a variety of medical reasons,
including diagnosis of bone fractures, internal bleedings, and tumors. Often CT
scans show the spine or sections of the spine. Even for an experienced radi-
ologist, manually searching for fractured vertebrae in such a CT image is a
time-consuming task, and is often not conducted, unless it was the primary
c Springer Nature Switzerland AG 2020
J. Cardoso et al. (Eds.): iMIMIC 2020/MIL3iD 2020/LABELS 2020, LNCS 12446, pp. 3–12, 2020.
https://doi.org/10.1007/978-3-030-61166-8_1
4 E. B. Yilmaz et al.

purpose of the CT scan. Hence, an automated tool identifying fractures in CT


images could indicate vertebral fractures to be checked by a radiologist and con-
sequently uncover otherwise missed vertebral fractures, substantially enhancing
fracture detection rate.
Tomita et al. [17] proposed a CNN/LSTM detection system automated on
patient level, but made strong assumptions regarding the position and visible
section of the spine in the CT image. Husseini et al. [8] combined an unsuper-
vised auto-encoder with a supervised multi-layer perceptron to leverage larger
quantities of unlabeled vertebrae, but require segmentation masks at training
time for all vertebrae. Nicolaes et al. [12] trained a CNN to segment spinal CT
images based on masks that are automatically derived from ground-truth coor-
dinates and annotations of the vertebrae. However, these approaches [8,12,17]
classified fractures based on Genant grading, and ignore whether a deformity is
degenerative or constitutes an osteoporotic fracture. Another possible approach
is an automated pipeline localizing and classifying vertebrae in CT images. For
the localization task, successful methods were recently proposed, e.g. in [11].
In this work, we focus on the vertebral fracture classification task, for which we
evaluate a pre-trained U-Net prefix and a custom CNN architecture1 . In combi-
nation with the vertebra localization approach proposed in [11], this constitutes
a fully automatic pipeline.
To improve the trust in the model’s decision, particularly in a clinical context,
it is desirable to shed light on the process how it came to its decision. Attribu-
tion methods—algorithms computing visual explanations (attribution maps) for
interpreting models [3]—pose a step into this direction. A number of attribution
methods has been suggested [3,14–16], but the interpretation of attribution maps
is still under debate [2,3,18]. In order to evaluate the adequacy of attribution
maps, [2] developed sanity checks that compare maps of trained models to (1)
those of models with random weights and (2) those of models trained on random
labels. Young et al. [18] compared attribution maps computed for melanoma clas-
sifiers of similar accuracy. We apply the analyses to the 3D vertebral osteoporotic
fracture classification task, computing attribution maps using a set of selected
attribution methods. Apart from assessing the sanity checks proposed in [2] we
propose two additional checks. First, we compare explanations from independent
training runs with the same architecture to test the dependence of the computed
maps on the specific realization of the learned parameters. Additionally, we test
the equivariance of attribution maps w.r.t. slight translations.

2 Methods
Given a 3D patch of a CT image containing a centered vertebra, the task of
our proposed models is to decide if the vertebra has an osteoporotic fracture
or not. Note that this means not only distinguishing between Genant score 0
and larger than 0, but also distinguishing whether a deformity is degenerative or
1
We also conducted experiments with a 3D ResNet18 variant [7] that are out of scope
for this publication but are in line with the presented results.
Assessing Attribution Maps for Vertebral Fracture Classifiers 5

constitutes an osteoporotic fracture, since this is important for an early diagnosis


of osteoporosis and decisions on therapeutic measures that may result from it.

2.1 CNN Architectures


Two architectures are evaluated: (1) A prefix of a 3D U-Net architecture that was
pre-trained for brain tumor segmentation [9] with a classification head trained
on task-related CT data, (2) a custom CNN architecture for 3D images based
on the results of a hyperparameter search.
Each model has three parallel outputs (see Fig. 1 for the custom CNN as
example) to calculate (i) a score between 0 and 1 to indicate the osteoporotic
fracture status, (ii) the Genant score [5] ranging from 0 to 3 and indicating the
severity of the deformity, and (iii) three “deformity percentages”, continuous
measures specifying the relative reduction in vertebral heights.
Output (i) uses the sigmoid activation function and is trained using the
binary cross-entropy loss, outputs (ii) and (iii) use a linear activation function
trained using the mean squared error (MSE) loss. At test time only output
(i), in combination with a threshold value optimized on the validation data, is
used to classify the vertebra. The additional outputs are used for regularization
during training. As different models may require different regularization schemes,
weighting the losses allows to control the influence of the individual outputs.
Although the architectures—as explained in the next sections—are quite dif-
ferent, they are trained in a similar setup. The Adam optimizer [10] is used in
conjunction with Stochastic Weight Averaging (SWA) [13]. SWA averages the
weights that occur after each 160 mini-batches, which we found to improve gener-
alization with little computational overhead. The averaged models are evaluated
on validation data for early stopping. To improve diversity in training data, ran-
dom multiplicative and additive Gaussian noises, shifts by up to two voxels in
each dimension, mirroring on the sagittal plane, rotation on the sagittal plane
by up to 18 degrees and cropping (replacing borders of the input with zeros) are
applied to the vertebral patches as data augmentation, motivated by prior work.
The hyperparameters listed in the following sub-sections were determined in
manual and automatic hyperparameter searches.

2.2 Prefix of Pre-trained U-Net Encoder


As a well-known architecture for semantic segmentation we adopt the U-Net
architecture to our task. Specifically, we use the encoder part of a 3D U-Net
architecture applied to multi-modal MRI scans for the task of brain tumor seg-
mentation [9]: We freeze the pre-trained weights of the first two residual blocks,
omit the remaining layers of the 3D U-Net and instead append a new fully con-
nected layer with 16 units, ReLU activation function and 25% dropout, followed
by the three classification outputs described above. The model is trained on
patches of size 32 × 64 × 96 mm3 (longitudinal × anteroposterior × lateral). The
large size in the lateral direction is required because this U-Net expects input
shapes that are multiples of 32 and a resolution of 3 mm vx is used in the lateral
6 E. B. Yilmaz et al.

Fig. 1. Custom CNN; (i)–(iii): The outputs as named in Sect. 2.1

direction. A learning rate of 10−4 and a loss weight of 1 is used for all outputs,
i.e., the total loss is the unweighted sum of the losses at all three output nodes.

2.3 Custom CNN

We propose a custom CNN consisting of 4 convolutional layers followed by one


fully connected layer with 128 units and the output layers. The convolutional
layers have a filter size of 3 × 3 × 3 and a stride of 2 × 2 × 2 (more details
in Fig. 1). All convolutional layers use zero-padding, except the first which was
added to reduce the input dimensionality for regularization and memory reasons.
Building a task-specific architecture, the patch size can be changed to 40 × 50 ×
60 mm3 = 40 × 50 × 20 vx (longitudinal × anteroposterior × lateral) to better
match the shape of the vertebrae. A learning rate of 5 · 10−4 and loss weights 1,
1, and 100 for outputs (i)–(iii), respectively, were found by the hyperparameter
search.

2.4 Attribution Maps

We select the following attribution methods, based on the realization of imple-


mentation invariance (see below) and on simplicity:
A simple method of highlighting important parts in the image is to compute
the Gradient of the models activation (before sigmoid) w.r.t. the CNN input [3,
4]. Large values indicate regions in the input image where a small change in the
input would have a comparably large effect on the model’s output.
Element-wise multiplication of the gradient with the input (Gradient *
Input) aims to show the contribution of the individual pixels and reduce noise
in the explanations [14].
To highlight only areas that contribute positively to a model’s decision,
Guided BackProp was suggested, setting negative gradients at activation func-
tions to zero during backpropagation [16].
SmoothGrad [15] addresses the issue of noisy gradients by averaging gra-
dients on images that are augmented with random noise. In our experiment we
use a noise level (σ/(xmax − xmin )) of 15% and average across 50 noisy images.
Assessing Attribution Maps for Vertebral Fracture Classifiers 7

When using fixed noise values for SmoothGrad and with the exception of
Guided BackProp, the above methods satisfy implementation invariance: Two
models that compute the same function should, given any fixed input image,
have the same attribution maps.

2.5 Sanity Checks


To assess the quality of the explanations in form of attribution maps, we per-
form a number of sanity checks. Each sanity check compares attribution maps
computed in two different settings (see below). As in [2,18], the similarity is
evaluated using the structural similarity index (SSIM) computed between the
whole 3D attribution maps. The reader is informed that the SSIM is a value
between −1 and 1 where 0 corresponds to no correlation.
As suggested by [2], we compare attribution maps computed for a trained
model to those of a model with re-initialized “Random Weights”. The frozen,
pre-trained weights of the U-Net are not changed. The output layers of the model
are also not re-initialized, since they were originally initialized with zeros leading
to all-zero attribution maps. Strong similarity of the maps would indicate that
learned parameters are independent from the explanations, while—as a result of
learning—they clearly have strong influence on the model’s output.
In the second setting proposed by [2], the attribution maps are compared for
a model trained on “Random Labels” and a correctly trained model. More
precisely, all labels in the training dataset are re-assigned to random vertebrae
before training, preserving class ratio. A strong similarity in this setting indicates
that attribution maps do not reflect task-specific supervision.
As a further check to assess the meaning of the explanations, “Re-Training”,
we compare the explanations computed for two trained models with the same
architecture on the same data, where only the random components during train-
ing vary. This includes the weight initialization, the choice of images in each
mini-batch, dropout, and data augmentation. The total number of mini-batches
applied during training may vary due to early stopping. If weak (or no) similar-
ity is observed, either completely different features were learned—which seems
unlikely—or the attribution maps are not consistent in the visualization of the
learned features.
Finally, we address equivariance of attribution maps w.r.t. slight translations,
“2-Voxel-Shift”: We compute attribution maps on patches of vertebrae that
were translated by 2 voxels and compare them to translated attribution maps
computed on original patches. As translation introduces zeros at the border
of the patches, the respective borders are cropped from the patches and the
explanations, in order to not influence the SSIM values. Since the object of
interest is fully contained in the image patch in both scenarios, we expect both
the model’s output and attribution maps to be similar in both settings.
8 E. B. Yilmaz et al.

3 Experiments
The dataset used in this study contains 159 low-dose CT images of distinct
patients (136 female), typically showing vertebrae T5–L4. The images were
acquired in seven centers participating in the in the Diagnostik Bilanz study
of the BioAsset project [6]. SpineAnalyzerTM [1] was used by a radiologist to
annotate—for each visible vertebra—the Genant score [5], “deformity percent-
ages” indicating height reduction, vertebra centers on a 2D sagittal slice, and a
differential diagnosis indicating if the vertebra shows either a “deformity” (1019
cases), an “osteoporotic fracture” (128 cases), is “unevaluable” (due to noise,
5 cases) or “normal” (802 cases). For the binary classification task, vertebrae
with an “osteoporotic fracture” are labeled 1, “deformity” and “normal” corre-
spond to 0, and “unevaluable” vertebrae were excluded. The lateral coordinate
was computed using a state-of-the-art vertebra localization tool [11] and man-
ually checked for correctness. Centered on these coordinates, for each vertebra,
a 3D patch is extracted of fixed size from the CT image, serving as input for
the respective CNN. As the only preprocessing steps, images were scaled to
3
a resolution of 1 × 1 × 3 mmvx (longitudinal × anteroposterior × lateral) and
Hounsfield-values were divided by 2048. The images were split on patient level
into four subsets defining a 4-fold cross-validation setup. In each run, two data
subsets were used as training data, one as validation data (for early stopping and
choosing the classification threshold) and one subset as test data. To address the
class imbalance, at training time each vertebra in each mini-batch was chosen
randomly, such that with 50% probability a vertebra labeled 1 was chosen. The
data augmentation methods listed in Sect. 2.1 were applied to each vertebra in
each mini-batch.

Table 1. Vertebral fracture discrimination results for the U-Net prefix and the Cus-
tom CNN. Mean and standard deviation are computed across 4 folds. Bold: Best per
column. ROC-AUC: area under the ROC curve. AP: average precision.

Method ROC-AUC (±σ) AP (±σ) Specificity (±σ) Sensitivity (±σ)


U-Net prefix 0.939 (±0.011) 0.703 (±0.042) 0.912 (±0.030) 0.824 (±0.049)
Custom CNN 0.989 (±0.0088) 0.907 (±0.015) 0.958 (±0.024) 0.906 (±0.081)

3.1 Vertebral Fracture Discrimination

Both investigated architectures accurately discriminate osteoporotic fracture sta-


tus of individual vertebrae, with ROC-AUCs of above 0.9 (results in Table 1).
The custom CNN achieved on average a higher ROC-AUC (0.989) and aver-
age precision (0.907) than the U-Net prefix. The Pearson correlation coefficient
of the architectures’ outputs—the scores between 0 and 1—on vertebrae in the
respective test sets was 0.72.
Assessing Attribution Maps for Vertebral Fracture Classifiers 9

3.2 Attribution Maps


Some examples for attribution maps created in the different settings are dis-
played in Fig. 2 (top). The lower part of Fig. 2 plots SSIM scores for the sanity
checks, together with the Pearson correlation of the respective models’ output.
The following observations are based on the results displayed in the figure and
manual inspection of explanations for other vertebrae.

Similarly Accurate Models Can Have Different Explanations. Inspect-


ing the Pearson correlation, we found the compared models’ outputs to show
stronger correlation in the settings “Re-Training”, and “2-Voxel-Shift”, than for

Fig. 2. Attribution maps and sanity checks for the U-Net prefix (left) and the Custom
CNN (right). Top: Comparison of the central sagittal slices of attribution maps using
a true positive L3 vertebra as example, where blue, white and red pixels correspond to
negative, near-zero and positive values. The “Baseline” row shows the attribution maps
for a trained instance of the corresponding CNN; the following rows show resulting maps
when performing the described sanity checks. Images were normalized individually by
division with the maximum absolute value. Bottom: A quantitative comparison, where
colored bars display the SSIM values averaged across the vertebrae in the respective
test sets and black bars show the Pearson correlation of the models’ outputs. Error
bars indicate the standard deviation across vertebrae.
10 E. B. Yilmaz et al.

“Random Weights” and “Random Labels”. In contrast, the SSIM of the attri-
bution maps was only slightly larger, if at all (see Fig. 2 and Sect. 3.1).

Gradient * Input Seems to Highlight Input Structures. In our experi-


ments, Gradient * Input visually highlights the vertebra’s outline in most set-
tings, leading to comparably high SSIM values for the performed sanity checks.

Guided BackProp Shows Fracture-Relevant Features When Trained


on Random Labels. While the Guided BackProp explanations for the Custom
CNN seems to highlight fracture-relevant features (vertebral end plates), this is
also the case in the “Random Labels” setting.

The Quantitative Results Do Not Always Match the Qualitative


Impressions. One example is the comparison of “Random Weights” and “Ran-
dom Labels” for the Custom CNN: While visually (inspecting the central sagittal
slice) the similarity seems to be much higher in “Random Labels”, the SSIM is
almost the same in the two settings.

4 Discussion and Conclusions

Automatic vertebral fracture discrimination may help to improve the fracture


detection rate by automatically assessing fracture status in CT scans taken for
various other medical reasons. The investigated CNN architectures were found
to permit highly accurate discrimination of the fracture status of vertebrae.
In contrast to previous publications [8,12], we focus on identifying osteo-
porotic fractures (as opposed to deformities in general). It must be noted that
different authors worked on different datasets with varying size and difficulty,
so results may not be comparable. Still, the quantitative results on our in-house
dataset compare well to previous work: [8] and [12] report specificities of 0.905
and 0.669 and sensitivities of 0.938 and 0.854, respectively. In additional experi-
ments, we combined the vertebra level models with an automatic vertebra local-
ization algorithm [11] to achieve fracture status assessment on a patient level,
achieving ROC-AUC values of 0.876 for the U-Net prefix and 0.940 for the Cus-
tom CNN.
Furthermore, to shed light on the complex decision process of the CNNs,
we have computed attribution maps for the fracture status classification task
on vertebra level. In addition to previously proposed sanity checks [2,18], we
introduced two further tests for explanation methods, namely retraining of a
given model configuration and testing equivariance of attribution maps with
regard to slight translations. In both cases, the computed explanation maps
were quantitatively (SSIM2 ) less similar than expected. We found that models
2
Other correlation measures (Pearson and Spearman coefficients) lead to similar con-
clusions.
Assessing Attribution Maps for Vertebral Fracture Classifiers 11

of similar accuracy can produce different attribution maps, which matches the
results of [18]. Gradient * Input, designed to reduce noise in the explanations,
highlights mostly the outline of the input structure, which was also observed
by [2]. SmoothGrad, also designed to reduce noise, visually and quantitatively
behaved similar to the original gradient.
In conclusion, the explanations exhibit a strong dependence on the model
architecture, the realization of the parameters, and the precise position of the
target object of interest. Since explanations of a model’s decision would be most
helpful to convince physicians that automated approaches perform trustworthy
evaluations, future work should address the implications of these findings for the
clinical practice.

References
1. SpineAnalyzer. Optasia Medical Ltd., Cheadle Hulme, United Kingdom (2013)
2. Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity
checks for saliency maps. In: Advances in Neural Information Processing Systems,
Montréal, Canada, pp. 9525–9536. Curran Associates Inc. (2018)
3. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Gradient-based attribution meth-
ods. In: Samek, W., Montavon, G., Vedaldi, A., Hansen, L.K., Müller, K.-R. (eds.)
Explainable AI: Interpreting, Explaining and Visualizing Deep Learning. LNCS
(LNAI), vol. 11700, pp. 169–191. Springer, Cham (2019). https://doi.org/10.1007/
978-3-030-28954-6 9
4. Erhan, D., Bengio, Y., Courville, A., Vincent, P.: Visualizing higher - layer features
of a deep network. Technical report, Univeristé de Montréal (2009)
5. Genant, H.K., Wu, C.Y., van Kuijk, C., Nevitt, M.C.: Vertebral fracture assessment
using a semiquantitative technique. J. Bone Miner. Res. 8(9), 1137–1148 (1993)
6. Glüer, C.C., et al.: New horizons for the in vivo assessment of major aspects of bone
quality microstructure and material properties assessed by Quantitative Computed
Tomography and Quantitative Ultrasound methods developed by the BioAsset
consortium. Osteologie 22, 223–233 (2013)
7. Haarburger, C., et al.: Multi scale curriculum CNN for context-aware breast MRI
malignancy classification. In: Shen, D., et al. (eds.) MICCAI 2019. LNCS, vol.
11767, pp. 495–503. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-
32251-9 54
8. Husseini, M., Sekuboyina, A., Bayat, A., Menze, B.H., Loeffler, M., Kirschke, J.S.:
Conditioned variational auto-encoder for detecting osteoporotic vertebral fractures.
In: Cai, Y., Wang, L., Audette, M., Zheng, G., Li, S. (eds.) CSI 2019. LNCS,
vol. 11963, pp. 29–38. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-
39752-4 3
9. Isensee, F., Kickingereder, P., Wick, W., Bendszus, M., Maier-Hein, K.H.: Brain
tumor segmentation and radiomics survival prediction: contribution to the BRATS
2017 challenge. In: Crimi, A., Bakas, S., Kuijf, H., Menze, B., Reyes, M. (eds.)
BrainLes 2017. LNCS, vol. 10670, pp. 287–297. Springer, Cham (2018). https://
doi.org/10.1007/978-3-319-75238-9 25
10. Kingma, D.P., Ba, J.L.: Adam: a method for stochastic optimization. In: 3rd Inter-
national Conference on Learning Representations, San Diego, May 2015
12 E. B. Yilmaz et al.

11. Mader, A.O., Lorenz, C., von Berg, J., Meyer, C.: Automatically localizing a large
set of spatially correlated key points: a case study in spine imaging. In: Shen, D.,
et al. (eds.) MICCAI 2019. LNCS, vol. 11769, pp. 384–392. Springer, Cham (2019).
https://doi.org/10.1007/978-3-030-32226-7 43
12. Nicolaes, J., et al.: Detection of vertebral fractures in CT using 3D convolutional
neural networks. In: Cai, Y., Wang, L., Audette, M., Zheng, G., Li, S. (eds.) CSI
2019. LNCS, vol. 11963, pp. 3–14. Springer, Cham (2020). https://doi.org/10.1007/
978-3-030-39752-4 1
13. Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., Wilson, A.G.: Averaging
weights leads to wider optima and better generalization. In: Uncertain Artificial
Intelligence, Monterey, California, pp. 876–885. AUAI Press, Corvallis, March 2018
14. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not Just a Black Box:
Learning Important Features Through Propagating Activation Differences (2016)
15. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: SmoothGrad: remov-
ing noise by adding noise (2017). arXiv:Learning
16. Springenberg, J., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity:
the all convolutional net. In: International Conference on Learning Representations
(2015)
17. Tomita, N., Cheung, Y.Y., Hassanpour, S.: Deep neural networks for automatic
detection of osteoporotic vertebral fractures on CT scans. Comput. Biol. Med. 98,
8–15 (2018)
18. Young, K., Booth, G., Simpson, B., Dutton, R., Shrapnel, S.: Deep neural network
or dermatologist? In: Suzuki, K., et al. (eds.) ML-CDS/IMIMIC -2019. LNCS,
vol. 11797, pp. 48–55. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-
33850-3 6
Another random document with
no related content on Scribd:
que beaucoup de Juifs, mes camarades, défenseur à l’excès des
traditions pharisiennes. »
Nous n’avons aucun motif d’induire que Paul, quinze ou vingt ans
après, exagérait ses violences pour mieux attester : Je me suis
converti malgré moi, sans nul mérite, sans que rien m’y préparât.
L’étrange, c’est plutôt le ton dégagé de sa confession ; pas un
mot ne laisse entendre que la mémoire de ses violences l’a bourrelé
de remords. Il expliquera très simplement, plus tard, à Timothée,
pourquoi il a pu trouver grâce devant Dieu :
« Le Christ Jésus m’a établi dans son service, moi qui étais
auparavant blasphémateur, persécuteur, tourmenteur. Il m’a pris en
pitié parce que j’avais agi sans savoir, dans le manque de foi. »
Les fureurs de Saul sortaient donc d’un zèle exaspéré pour la
religion qu’il croyait uniquement vraie. Ses cruautés trouveraient une
explication dans ce mot aigu de Pascal :
« Jamais on ne fait le mal si pleinement ni si gaiement que quand
on le fait par conscience. »
Mais il faut aussi comprendre quelle pouvait être l’âme d’un Juif
au I er siècle, ce qu’était le monde autour de lui.
On aurait grand tort de se figurer Israël comme un peuple, avant
tout, féroce. Dans son histoire, les traits de miséricorde et de
tendresse n’ont rien d’anormal. Sur l’âpreté des tempéraments, le
précepte divin posait son onction :
« Tu aimeras le Seigneur ton Dieu de tout ton cœur, de toute ta
force. »
Entre Iahvé et son peuple, un principe de suavité tempérait la
crainte :
« Le Seigneur ton Dieu t’a porté, lui disait Moïse, comme un
homme porte (sur l’épaule) son fils tout petit [45] . »
[45] Deutéronome I , 31.

A l’intérieur de la famille, une loi sainte gouvernait les rapports du


père et des enfants, des frères et des proches entre eux. C’était une
loi exigeant « la circoncision du cœur [46] » ; la bonté, le pardon y
tenaient leur place. Avant le père de la parabole évangélique, on
s’était souvenu d’Ésaü étreignant dans les larmes Jacob faible,
humilié ; de Joseph se réconciliant avec ses frères indignes ; de
David pleurant le misérable Absalom et criant : « Qui me donnera de
mourir à ta place, Absalom, mon fils, ô mon fils Absalom ! »
[46] Id. X , 16.

Selon le code mosaïque les juges devaient rendre justice au


pérégrin comme à l’Hébreu, au petit comme au grand, sans faire
acception de personnes, parce que leur jugement « était le jugement
de Dieu [47] ». Quand on entrait en guerre, devant une ville ennemie,
avant de donner l’assaut, il fallait « lui offrir la paix [48] ». C’était une
obligation de respecter durant un mois la femme captive [49] .
[47] Id. I , 17.
[48] Deutéronome XX , 10-11.
[49] Id. XXI , 11-14.

Moïse interdisait de livrer l’esclave fugitif à son maître [50] , de


garder plus d’un jour le gage du débiteur pauvre [51] . Il commandait
au riche d’ouvrir sa main à l’indigent, de laisser, pour l’orphelin et la
veuve, sur l’olivier quelques olives, dans la vigne quelques
grappes [52] . Il enseignait même la pitié pour les animaux : « Si, en
marchant sur une route, tu trouves dans les branches d’un arbre ou
à terre, un nid d’oiseau, et la mère couvant ses œufs ou ses petits,
tu ne la retiendras pas captive [53] . » Josèphe, célébrant l’humanité
de la loi juive, observe qu’elle défendait de tuer les animaux, « s’ils
entraient en suppliants dans une maison [54] ».
[50] Id. XXIII , 15.
[51] Id. XXIV , 12-13.
[52] Id. XXIV , 21.
[53] Id. XXII , 6.
[54] Contre Apion, l. II, ch. VI .

Un peuple où l’on avait conçu et compris, du moins littéralement,


le Cantique des Cantiques, les Psaumes, les Livres des Prophètes,
ne pouvait ignorer les délicatesses ni les violences de l’amour
humain ou divin. Nul n’a senti d’une façon plus véhémente que
l’amour est fait de pitié.
Mais les Juifs pouvaient-ils échapper à la dureté foncière de tout
l’Orient sémitique ? Quand on pense aux tyrans assyriens, aux
atrocités rituelles qu’attestent les bas-reliefs et les inscriptions de
ces pays, on s’étonne moins de voir Israël, en guerre contre des
voisins terribles, exterminer dans les villes hommes, femmes, petits
enfants, incendier les maisons, ne laisser que des cendres et de
l’horreur derrière lui. Les Hébreux savaient ce qui les attendait s’ils
épargnaient les idolâtres ; ils exécutaient sur eux le juste châtiment
d’Iahvé, et, plus encore, en les exterminant, ils se préservaient de
leurs dieux redoutables.
Israël eut besoin d’être fanatique ; autrement il aurait succombé,
et, avec lui, le pacte d’alliance, le témoignage du seul Dieu vrai. Il se
savait élu entre tous les peuples ; sa fierté d’un tel privilège était
farouche. Jamais orgueil nobiliaire n’a pu être comparé à celui des
Juifs. Un grand orgueil offensé devient cruel en se croyant juste.
D’où, chez eux, des vindictes inflexibles dont celles des hidalgos
espagnols seraient une faible réplique.
Le pays où ils se fixèrent, malgré ses parties fertiles, est dur
comme son climat.
Pays de hautes vallées et de faîtes abrupts, peu accessible par
la mer, et qui repousse l’étranger. Six mois d’été sans pluie ; un hiver
assez rude. Les villages, sur les pentes, ressemblent à des tas de
pierres. Nulle part au monde la pierre ne règne aussi implacable ; on
s’explique la lapidation, supplice éminemment juif ; sous les
monceaux de silex on chercherait les os d’un lapidé. Je ne connais
rien de plus désolant, surtout en automne, que la descente de
Jérusalem à Jéricho : des bosses de terres nues, après des bosses
de terres nues, çà et là broussailleuses, ou d’un gris de lèpre,
chargées de boursouflures livides, au-dessus d’une gorge rougeâtre
qui fait saillir et béer ses roches comme des gueules de bêtes
altérées.
De telles régions ne pouvaient convenir qu’à des brigands ou à
des clans rigides, intraitables pour tout ce qui violait les mœurs et les
principes de la communauté.
La Loi mosaïque les enserrait dans des haies de préceptes et de
rites, dans les craintes minutieuses des cas d’impureté. Elle exigeait
de ces paysans rapaces le sacrifice de leurs bestiaux, des victimes,
certains jours, sans nombre [55] . Aux grandes fêtes, le parvis du
Temple devenait un énorme abattoir ; le gémissement des animaux
égorgés couvrait les voix des prêtres ; ceux-ci n’étaient plus que
d’infatigables bouchers. Les lévites parfois devaient monter sur des
escabeaux pour ne pas tremper leurs jambes dans les nappes de
sang qui débordaient [56] . Le matin de Kippour, lors du grand jeûne
d’octobre, quand on avait imposé les mains au bouc qu’on chargeait
des péchés du peuple, les assistants crachaient tous sur lui, le
piquaient avec des épines [57] . Il était coiffé d’une bande de laine
écarlate ; puis, à coups de fouet, les prêtres le chassaient hors de la
ville, en un lieu désert. Là, on lui arrachait du dos sa toison qu’on
éparpillait sur les broussailles, et on le jetait dans un précipice. S’il
se relevait, personne ne lui donnait à manger ; il s’en allait mourir
comme un maudit, dans un trou.
[55] Pour la dédicace du Temple de Salomon, les
Paralipomènes (l. II, VII , 5) dénombrent l’immolation de
vingt-deux mille bœufs et de cent vingt mille béliers. Pour
la même cérémonie, Josèphe (A. J., VIII , 2) parle de
douze mille veaux et de cent vingt mille agneaux.
[56] Voir M. Marnas, Miriam, p. 220.
[57] Voir l’épître dite de Barnabé. Le passage sur le
bouc est une citation tirée on ne sait d’où.

Atroces pour nous, ces rites expiatoires l’étaient bien moins que
ceux des idolâtres, offrant leurs fils au bûcher de Moloch, ou se
mutilant, comme faisaient les prêtres de Cybèle, en public, avec
frénésie. Ils provoquaient les Juifs à la pénitence, commémorant les
peines dont Iahvé avait frappé leurs pères impies ou fornicateurs. Ils
préfiguraient la victime substituée, elle volontaire et parfaite, le Christ
percé d’épines, flagellé, honni. Mais, chez des âmes brutales, ils
excitaient le goût du sang, une sorte d’irritation luxurieuse déviée en
ivresse de tuerie.
D’ailleurs, asservis à des maîtres iniques, les Juifs, tout en
courbant l’échine, avaient médité d’affreuses représailles. Si on
touchait à leur culte et à la Loi, ils résistaient sauvagement, et les
répressions étaient inexorables. Lorsque Antiochus Épiphane
prétendit helléniser Jérusalem, établir dans le Temple une statue de
Zeus, lorsqu’il eut interdit la circoncision, les pharisiens s’obstinèrent
à faire circoncire les nouveau-nés. Tous ceux qui étaient dénoncés
étaient battus de verges, mutilés, mis en croix ; et les bourreaux,
après avoir étranglé les enfants, pendaient leurs cadavres au cou
des crucifiés [58] . Hérode ayant fait clouer sur le portail du Temple un
aigle d’or, deux docteurs, Judas et Mathias, l’arrachèrent en plein
midi, devant la foule, et le brisèrent à coups de hache. Arrêtés, ils
justifièrent leur violence avec ce seul argument : « Nous avons
vengé l’outrage fait à Dieu et l’honneur de la Loi dont nous sommes
les disciples. » Pour déchaîner un mouvement furieux, il suffit que
Pilate voulût faire promener à travers les rues de Jérusalem des
enseignes militaires où figurait le médaillon de César [59] . Caligula,
quand il essaya d’imposer dans le Temple sa statue en « nouveau
Jupiter », faillit soulever toute la Judée contre Rome.
[58] Josèphe, Antiq., l. XII, VII .
[59] Id., l. XVIII, IV .

A mesure que la nation juive se vit plus étroitement harcelée par


l’hellénisme [60] , pressée par l’arrogance et la rapacité romaines, son
esprit de révolte se renforça ; mais il devait se perdre dans l’anarchie
des factions. Sadducéens, bourgeois et sceptiques, semblables à
nos radicaux d’aujourd’hui, pharisiens intransigeants, zélotes et
démagogues illuminés s’exécraient les uns les autres. Les bandes
armées, les brigandages se multipliaient. Les grands prêtres
soudoyaient des séditieux qui provoquaient des rixes ; ils les
envoyaient saisir dans les granges des dîmes appartenant aux
sacrificateurs « dont quelques-uns étaient si pauvres qu’ils
mouraient de faim [61] ». Des sicaires, les jours de fête, arrivaient à
Jérusalem, cachant des dagues sous leurs manteaux.
[60] Sous Caligula, les Grecs massacrèrent, dans les
progroms d’Alexandrie, au dire de Josèphe, cinquante
mille Juifs ; d’où l’ambassade conduite par Philon auprès
de l’Empereur.
[61] Josèphe, Antiq., l. XX, VI .

Ils poignardaient les gens au milieu des cérémonies, et, les


voyant tomber morts, se penchaient sur eux comme pour les
secourir, échappant ainsi aux soupçons [62] .
[62] Id. XX, VII . Et Eusèbe, H. E., II, XX .

La férocité des mœurs, l’exaspération des caractères atteignaient


déjà ce paroxysme qui aboutira aux atrocités héroïques du siège de
Jérusalem, aux épouvantes de Massada. Josèphe, homme cultivé,
raconte, comme une chose toute naturelle, de quelle manière il traita
un factieux de Tibériade. Celui-ci était venu simplement lui réclamer
une somme qu’il ne devait pas :
« Je le fis battre de verges, je lui fis couper une main qu’on lui
attacha au cou, et je le leur renvoyai en cet état [63] . »
[63] Vie, ch. XXIV .

Plus loin, il invite un autre séditieux à se trancher lui-même, d’un


coup d’épée, la main gauche. Et cet homme s’empresse
d’obtempérer.
Saul, dans son offensive contre les Nazaréens, se comporta
donc selon la rigueur d’un bon pharisien sectaire. Il y ajoutait
l’emportement de sa jeunesse, la fierté d’exceller dans une œuvre
juste. Nul doute qu’il n’écoutât en même temps des impulsions
démoniaques. Il dira dans la suite : « Nous n’avons pas seulement à
combattre contre la chair et le sang, mais contre les Puissances,
contre les Maîtres de ce monde ténébreux [64] . » Les Puissances
d’en bas l’armaient de leur furie. Elles l’avaient élu comme un parfait
agent d’extermination.
[64] Éphésiens, VI , 12.

Au reste, il croyait connaître l’histoire de Jésus, ses


enseignements ; il les interprétait, sans intelligence, « en homme
charnel » ; il en demeurait scandalisé, outré. Il était pharisien, et
Jésus avait écrasé, sous une réprobation, la superbe, l’hypocrisie
des pharisiens. Israël attendait un Messie qui établirait sa revanche
sur les oppresseurs et même lui soumettrait l’univers. Isaïe l’avait
annoncé : L’empire sera sur son épaule. A cette prophétie, mal
comprise, vulgarisée dans tout l’Orient, les Romains eux-mêmes
prêtaient attention [65] . Jésus, trompant les espoirs terrestres
d’Israël, semblait l’ennemi à détruire dans la personne de ses
disciples, faux prophètes qui blasphémaient l’éternité de la Loi,
l’avenir du peuple saint. Que devenait son privilège, si toutes les
nations étaient appelées au Royaume ? En persécutant les
Galiléens, Saul pensait « rendre hommage à Dieu [66] ».
[65] Tacite, Hist., V , 13, et Suétone, Vespasien, IV .
[66] « L’heure vient où quiconque vous tuera croira
rendre hommage à Dieu » (Jean, XVI , 2).

On a contesté qu’il ait pu les traquer hors de Palestine, jusqu’en


Syrie. Mais, dans une phase de trouble, le sanhédrin se hâtait de
ressaisir une compétence pénale dont il était jaloux. Or, la Syrie
appartenait en ce temps-là au roi Arétas, beau-père d’Hérode le
Tétrarque ; et les Juifs, nous le savons par Paul lui-même,
s’entendaient fort bien avec Arétas [67] . En fait, partout où vivait une
communauté juive, l’émissaire du sanhédrin exerçait un droit de
police.
[67] « A Damas, l’ethnarque du roi Arétas faisait
garder la ville pour se saisir de moi » (II Cor. XI , 32).
Juster (op. cit., t. II, p. 134-139) estime incertaine, mais
possible, la compétence du sanhédrin hors de la
Palestine ; car Hérode avait eu le droit de se faire
remettre par les autorités romaines des criminels enfuis à
l’étranger. Ce droit, surtout en matière de crimes
religieux, avait pu passer d’Hérode au sanhédrin.

Tout persécuteur devient un persécuté. Dans l’idée que des


victimes lui échappent, il ne dort plus. Pour allonger sa liste de
suspects, il n’a jamais assez d’espions. Fatalement, son inquisition
s’étendra aussi loin qu’il peut faire devant lui le vide par la terreur.
C’est pourquoi nous trouverons Saul, avec une escorte de policiers,
en marche vers Damas, « soufflant la menace et le meurtre »,
frémissant d’anéantir une Église qui se croyait à l’abri.
Mais, avant de le joindre sur la route brûlante où le Christ lui
donna rendez-vous, il convient de le mieux connaître et d’atteindre
les premiers linéaments de sa personnalité.

*
* *

QUI ÉTAIT SAUL ?

Il s’est chargé de nous répondre ; il a dressé un sommaire état


civil, exhibé, pour les biffer aussitôt d’un trait méprisant, ses titres de
noblesse juive :
« Si quelque autre s’imagine être puissant selon la chair, moi
encore plus : circoncis le huitième jour, Israélite par ma race, de la
tribu de Benjamin, Hébreu issu d’Hébreux ; à l’égard de la Loi,
pharisien [68] … »
[68] Philipp. III , 4-3.
Aujourd’hui, on jugerait un peu vague le signalement. Ce n’est
point négligeable, pourtant, d’apprendre que Saul, « Hébreu issu
d’Hébreux », était de la tribu de Benjamin, et pharisien.
Né hors de Palestine, il devait s’attacher d’autant plus à la pureté
de ses ascendants, certifier qu’il tenait du judaïsme tout ce qu’il était.
Mais il se prévalait d’un autre avantage : sa famille avait rang dans
la tribu de Benjamin, celle qui marchait en tête des processions,
ayant, la première, traversé la mer Rouge, de Benjamin qui, seule
avec Juda, après la grande captivité, avait relevé les murailles de
Sion [69] . Ce n’était pas tout ; pharisien, il appartenait à une caste
supérieure, un peu comme le religieux d’un Ordre vis-à-vis des
séculiers. Les pharisiens, « les gens à part », se posaient eux-
mêmes au-dessus du commun des Juifs ; ils avaient seuls la haute
science, la vertu sans reproche ; car peut-on être agréable au
Seigneur, si on ne connaît toute la Loi ? Et ils se targuaient de la
méditer nuit et jour ; plus ils en resserraient les préceptes, plus ils en
aggravaient les contraintes, plus ils s’estimaient devant Dieu.
[69] Esdras, l. II, XI .

Chez Saul, l’orgueil théocratique fut sans doute immense.


Converti, il reconnaîtra que la fierté du sang est une vanité
misérable, une de ces choses « qu’on jette aux chiens [70] ». Pour
l’instant, excusons-le ; jamais peuple n’a pu justifier, comme Israël, la
gloire de ses origines ; il était l’unique nation choisie par le Tout-
Puissant, conduite par lui, en ses grandeurs comme en ses
désastres, afin qu’elle gardât les vérités essentielles et la semence
d’où l’Homme-Dieu prendrait sa chair.
[70] Philipp. III , 8. Tel est le sens exact du mot violent
qu’il emploie : skybala.

Le Messie étant venu, le peuple juif aurait pu mourir, comme


l’arbuste des solitudes quand, au sommet de sa tige, la fleur pourpre
a surgi. Il a survécu en qualité de témoin ; la conscience de sa
mission divine l’avait doué d’une telle force qu’il est resté, dans sa
déchéance, un peuple-roi. Que lui importe d’avoir bu, durant des
siècles, les affronts comme l’eau ? Il garde dans la bouche le goût du
vin des Maîtres ; il n’a jamais douté de lui-même ; cette foi tenace le
prédestinait à dominer les nations ; et, maintenant, il a fait de toutes
« l’escabeau de ses pieds ».
A Jérusalem, un vendredi, vers 4 heures — c’est le moment où
les Israélites dévots vont allumer des cierges contre le mur des
pleurs et psalmodier — j’ai remarqué un petit bossu qui se
rengorgeait en marmottant des prières et se balançait, un livre à la
main, avec une mine de satisfaction presque arrogante. Je me suis
dit : « Voilà Saul ! »
La fierté juive, en Saul, était doublée de la hauteur pharisienne.
Pour celle-ci Jésus n’avait eu que des mots terribles. Le réquisitoire
qu’abrège saint Mathieu (ch. XXIII ) est un magnifique portrait de la
caste, et qui vise, au delà, toute l’enflure des orgueils sociaux. Les
pharisiens n’agissent que pour être vus ; il élargissent leurs
phylactères, ils allongent les franges de leurs robes. Ils aiment les
lits d’honneur dans les banquets, les bancs d’honneur dans les
synagogues. Ils veulent qu’on les salue sur les places, qu’on les
appelle : Rabbi, Rabbi !…
Saul pouvait donc se vanter d’appartenir, comme on dirait, à une
honorable famille juive. Son père n’était pourtant pas un Hébreu de
Judée, mais un helléniste établi à l’étranger depuis assez
longtemps ; il portait le titre de citoyen romain, et son fils en hérita.
Pas une seule fois dans les Épîtres, Tarse n’est nommée ; ce
sont les Actes qui font dire à Paul [71] — et il le dit en araméen : —
« Je suis né à Tarse, en Cilicie. »
[71] XXII , 3. Il eut, vraisemblablement, dès son
enfance, deux noms : un nom juif, Saul, et un romain à
désinence grecque, Paulos.

Tarse, proche de la mer, au débouché de la seule route par où


les caravanes venant de l’Asie Mineure franchissaient le défilé des
portes ciliciennes, était alors une des plus grandes villes de l’Orient.
La plaine de Cilicie, ample et magnifique, avec l’opulence de ses
cotons et de ses blés, ferait songer à l’Égypte, si elle ne s’appuyait
aux rampes du Taurus, dont les crêtes coiffées de nuages la barrent
à l’Occident.
Point de jonction entre la haute Asie et la côte — le long du
Cydnus les bateaux de toute la Méditerranée remontaient jusqu’à
ses quais — elle s’offrait comme un confluent de civilisations.
L’Hellade y superposait son empreinte à celle de l’Assyrie, de la
Perse, de la Phénicie. Ses monnaies portent souvent un Baal, figuré
en Zeus, ayant un aigle à son côté [72] . Tarse amalgamait l’élégance
grecque avec les rites et les voluptés du vieil Orient. C’était là que,
sur la proue d’or de sa galère, sous des voiles de soie parfumées,
Cléopâtre avait attendu Antoine. Quand Paul citera aux Corinthiens
le proverbe grec [73] :

Mangeons et buvons, car demain nous mourrons,

il se souviendra peut-être aussi de l’inscription assyrienne que


portait, non loin de Tarse, la statue de Sardanapale, la statue aux
doigts disposés comme si elle voulait les faire craquer :
[72] Voir Ramsay, The Cities of Paul, p. 129.
[73] 1 Cor. XV , 32. Ce proverbe se rencontre dans la
Thaïs de Ménandre, mais il était déjà dans Isaïe (XXII ,
13), où Paul a dû le prendre.

« Passant, mange, bois, divertis-toi ; car tout le reste ne vaut pas


cela [74] . »
[74] Strabon, XIV, V . On voit encore à Tarse une
construction massive avec des murailles d’une
prodigieuse épaisseur qu’on dénomme le tombeau de
Sardanapale.

Les Tarsiens possédaient, à un étrange degré, la facilité qu’ont


les Orientaux communément d’apprendre les langues et
d’improviser. Les écoles de Tarse envoyaient à Rome des
grammairiens et des philosophes. Le stoïcisme y florissait.
Sur les bords du Cydnus, un gymnase célèbre groupait les
meilleurs maîtres. Saul, enfant, le fréquenta-t-il ? Apprit-il le grec
dans une école juive, près de la synagogue, ou avec un pédagogue,
dans la maison de son père ? L’essentiel pour nous, c’est qu’il eut
une pleine connaissance de la langue des idées, de l’idiome qui
pouvait le mieux rendre universelle sa doctrine. S’il était né à
Jérusalem, il l’aurait ignorée ou mal sue. Les rabbins défendaient de
l’apprendre aux garçons, sauf, disaient-ils en raillant, « lorsqu’il ne
faisait ni jour ni nuit [75] ». Elle était, pour leur méfiance, le véhicule
du mensonge païen ; quiconque avait sucé le miel des fables
helléniques trouvait dans la vérité des Écritures une amertume.
[75] Talmud. Péa, I ; Josèphe (Antiq., XX, 18) constate
le mépris des Juifs pour l’étude des langues profanes.

Mieux encore que le grec, Saul retint ce qu’on apprenait à tous


les petits Hébreux, les dix-huit bénédictions de l’Amida, la psalmodie
du Hallel. Il vit, dans la synagogue, le lecteur tirer de l’armoire l’étui
qui enfermait les rouleaux de la Loi, et, chez son père, le soir de
chaque vendredi, s’allumer les lampes du sabbat.
Il reçut aussi les rudiments d’un métier manuel. Nous savons par
les Actes [76] qu’il était « un faiseur de tentes ». Rien ne prouve que
son père lui-même fût artisan. Mais, selon les docteurs, tout bon Juif
devait savoir œuvrer de ses mains ; et le fameux Chammaï exhibait
à son oreille le copeau du charpentier.
[76] XVIII , 3.

Saul sut fabriquer ces tentes noires en poil de chèvre où


s’abritent encore les bergers ciliciens. J’ai vu, dans un faubourg de
Tarse, des ouvriers en faire le tissu d’après des méthodes simplistes
qui, depuis les temps de Paul, n’ont guère dû changer.
Ils étaient trois dans un hangar ouvert sur les côtés, trois
hommes maigres, un peu chauves, grisonnants, avec des visages
ascétiques. Le premier, debout, mettait en action un rouet d’où
pendaient deux bouts de corde ; d’une sacoche suspendue contre
son tablier il tirait un à un les poils qu’il tordait autour de la corde en
mouvement. Il filait ainsi, marchant à reculons depuis le fond du
hangar jusqu’à l’entrée, et là, il abaissait la longueur du fil, le
déposait auprès des autres.
Ses deux compagnons travaillaient, assis à terre sur une peau de
mouton, les pieds dans un trou. Chacun d’eux avait devant soi un
vaste métier incliné quelque peu en arrière ; il y disposait la chaîne,
l’écartait avec un couteau de bois, passait agilement la navette entre
les fils tendus, les arrêtait ; puis il étirait la chaîne et la trame d’un
coup sec de son cardoir, outil massif en bois poli, qui ressemblait,
sauf ses dents, à un joug pour les bœufs.
On s’explique comment Paul, après avoir manié des heures ce
cardoir pesant, écrivait d’une main gourde. Quand il dit aux Galates :
« Voyez les gros caractères de mon écriture [77] », il ne fait pas
allusion à ses mauvais yeux qui l’eussent contraint de grossir les
lettres. Il écrivait comme un ouvrier dont les doigts sont raidis par la
manœuvre d’une chose très lourde [78] .
[77] VI , 11.
[78] Cet outil pèse près de deux kilos.

Auprès des artisans de Tarse, je recueillis quelques détails


propres à préciser certains faits dans sa vie. Leur métier est
lucratif [79] ; il a dû l’être toujours. De la sorte, Paul subvenait à ses
besoins et à ceux des indigents, sans être une charge pour
personne ; il pouvait pratiquer largement la maxime du Seigneur
Jésus :
[79] Ils gagnent, par journée, cinquante à soixante
francs.

« Donner, c’est plus de béatitude que recevoir [80] . » J’appris, en


outre, des camarades lointains de l’Apôtre — savaient-ils son nom ?
— qu’ils besognaient, comme lui, « nuit et jour [81] ». Ils dormaient
sur des sacs ; peut-être en faisait-il autant. Mais leur travail, tout
uniforme et machinal, laisserait libre leur esprit, s’ils avaient le goût
de penser ; et Paul devait, sans interrompre le jeu de la navette et du
cardoir, songer aux églises ou prêcher autour de lui.
[80] Actes XX , 35.
[81] I Thessalon. II , 9 : « Nuit et jour au travail pour
n’être à charge à personne d’entre vous » ; et I Cor. IV ,
12 : « Nous nous épuisons à travailler de nos mains. »

A l’âge où il n’était qu’un apprenti amateur, il ne soupçonnait


guère qu’en tissant des poils de chèvre il préparait son apostolat.
L’étude des Écritures et les sentences des rabbins le captivaient
davantage.
Son père voulut qu’il achevât sa formation de pharisien et de
lettré. Vers douze ans, il fit, comme tous les adolescents juifs, un
pèlerinage rituel à Jérusalem. Si l’on entend au sens littéral ce qu’il
dit aux Juifs [82] de son éducation, il aurait même grandi dans la ville
sainte. Mais il ajoute qu’il fut « nourri aux pieds de Gamaliel » ; le
fameux rabbi l’aurait-il admis comme auditeur, s’il n’avait eu déjà la
maturité d’un étudiant ?
[82] Actes XXII , 3. Le terme qu’il emploie veut dire :
nourri en grandissant.

Nous imaginons volontiers Saul, assis aux pieds du maître, les


genoux entre ses mains croisées, semblable à ces jeunes
musulmans qui, dans les mosquées, font cercle autour d’un imam,
silencieux et ravis, les yeux pleins d’une sorte d’extase, tandis que le
docteur, derrière une petite table, pérore avec un feu prophétique.
L’éducation d’un étudiant juif se concevrait assez bien d’après
celle d’un séminariste dans un milieu sacerdotal fermé. La science
qu’il absorbait se ramassait autour de l’Écriture et de la Loi. Il devait
posséder à fond le Pentateuque, lire les prophètes, devenir exégète
et théologien. Il lisait aussi des écrits ésotériques comme le Livre
d’Hénoch, l’Assomption de Moïse [83] .
[83] Voir Lagrange, le Messianisme chez les Juifs,
passim.
Mais l’exégèse juive se plaisait à l’imprévu des conclusions ; elle
accrochait aux textes des allégories, une dialectique retorse et
paradoxale. Ainsi, à propos du premier homme :
« Le Saint (béni soit-il) a fait dans le moule d’Adam tous les
hommes de la terre, et personne n’est semblable à l’autre. Aussi,
chacun doit se dire que le monde a été créé pour lui…
« Dieu a varié dans l’homme trois choses : le visage pour éviter
des confusions ; la pensée, afin d’éviter des vols ; la voix, pour éviter
(la nuit) des unions illégitimes [84] . »
[84] Baba Bathra, trad. Schwab, p. 270.

Les Écritures elles-mêmes étaient submergées sous les


commentaires de la Loi. Retenir sans notes — car il était interdit de
noter les décisions des rabbins — tous les cas qu’elle posait, les
solutions contradictoires, les possibilités qu’elle impliquait, c’était un
travail embrouillé, comme celui d’apprendre le grand alphabet
chinois.
Aucun acte ne devait être accompli sans bénir le Seigneur, et
chaque bénédiction exigeait une formule spéciale. On ne prononçait
pas les mêmes mots, à table, pour bénir des raves coupées en petits
morceaux ou des raves coupées en long [85] . Avant de boire, les
pharisiens se lavaient les doigts. Mais, selon Chammaï, il fallait
d’abord faire cette ablution, ensuite verser l’eau dans le calice ; selon
Hillel, verser l’eau avant de se laver.
[85] Voir Marnas, op. cit., p. 26.

Et puis il y avait les ergoteries sans fin sur les impuretés, sur ce
qui est permis ou prohibé les jours de sabbat [86] . Il y avait la
casuistique des dommages qui entraînent ou non un châtiment :
[86] Faut-il rappeler l’opposition d’Hillel et de
Chammaï autour de ce point grave, si, un jour de fête, on
pouvait manger un œuf pondu dans la journée ?
« Si un coq, en voltigeant d’un endroit à un autre, cause des
dégâts par son contact, le propriétaire sera responsable du dégât
entier. Mais, si le dommage est survenu par le vent des ailes, le
propriétaire ne doit payer que la moitié de la valeur [87] . »
[87] Baba Gama, trad. Schwab, p. 12.

« Deux ânes se suivent ; l’un glisse et tombe ; puis l’autre arrive,


se heurte contre lui et tombe aussi ; enfin le troisième heurte celui-ci
et tombe à son tour ; le maître du premier devra payer à celui du
deuxième le dommage survenu, et le deuxième au troisième [88] . »
[88] Baba Gama, p. 25.

« Si un homme frappant son père ou sa mère les blesse, ou s’il


blesse quelqu’un le jour du sabbat, il ne paie rien, car il est
condamné à mort. Si un homme blesse son propre esclave païen, il
n’est pas condamné au paiement [89] . »
[89] Id. p. 65.

On ne s’étonnera guère que la logique, chez saint Paul, conserve


des traces d’arguties, une tendance aux coudes brusques dans le
cheminement des idées [90] . Un trait consigné par le Talmud sur
Rabbi Gamaliel, son maître, révèle, s’il est véridique, l’ironie subtile
d’un sophiste qui trouve réponse à tout :
[90] De même, les citations composites, ou qui
mettent des textes en enfilade, sont une réminiscence
des méthodes rabbiniques (voir Prat, Théologie de saint
Paul, I, p. 32-33).

« R. Gamaliel allait prendre ses bains à Acco dans une maison


de bains qui appartenait à la déesse Aphrodite (le temple de cette
déesse, ses prêtres et le personnel étaient entretenus des revenus
qu’on tirait de la maison de bains). Un païen nommé Proclus ben
Philosophos lui demanda comment il pouvait se permettre d’aller
prendre des bains dans une maison affectée au service d’une idole,
quand la Loi mosaïque défendait de tirer profit des objets consacrés
aux divinités païennes. Une fois sorti, R. Gamaliel répondit : « Je ne
vais pas dans le domaine de l’idole, c’est elle qui vient dans le mien ;
on n’a pas construit la maison de bains en l’honneur d’Aphrodite ;
c’est elle qui sert d’ornement à la maison de bains [91] . »
[91] Aboda Zara, trad. Schwab, p. 212.

Gamaliel, comme son aïeul Hillel, se distinguait par une relative


largeur de vues où il corrigeait la casuistique dure, pointilleuse de
Chammaï. Il représentait, parmi les pharisiens, l’école libérale. Le
langage qu’il tient dans le sanhédrin, au sujet des Apôtres, énonce
une bizarre doctrine de laisser faire et de fatalisme providentiel :
« … Et maintenant je vous dis : Écartez-vous de ces hommes et
laissez-les ; parce que, si leur volonté, leur œuvre ne sont
qu’humaines, elles tomberont d’elles-mêmes. Mais, si elles sont de
Dieu, vous ne pouvez les abattre, de peur qu’on ne vous trouve
combattant contre Dieu [92] . »
[92] Actes V , 35-39.

On a supposé qu’une inspiration divine lui suggéra cette politique


évasive. La tradition [93] le représente comme secrètement chrétien.
[93] Les Recognitiones, livre apocryphe, hérétique,
mais qui date du II e siècle, disent de lui : « Gamaliel,
prince du peuple, était secrètement notre frère. » Sur la
légende de Gamaliel, voir Lagrange, Saint Étienne et
son sanctuaire, p. 42-59.

Quoi qu’il en soit, entre lui et Saul éclate une antithèse : voilà un
Maître, pondéré, souple, théoricien de l’indulgence, et son disciple
agit à l’encontre de sa doctrine autant qu’un Jacobin de 93
démentait un Necker ou un Montesquieu.
Il serait oiseux de vouloir élucider ce problème, comme de
trancher si Paul fut ou non rabbin. Très souvent le disciple est
l’opposé du maître ; de même que le fils est la négation du père.
Gamaliel eut un fils fanatique et hostile aux chrétiens. Saul, dans sa
jeunesse, était quelqu’un de très indépendant. Porté aux extrêmes, il
suivait en ses haines la fougue de ses énergies. S’il admirait la
science et l’autorité de Gamaliel, il estimait dangereux son
libéralisme. En tant que Juif, avait-il tort ?
Une hypothèse semble absurde, effroyable, celle de concevoir la
foi chrétienne étouffée dans sa première croissance. A ne
l’envisager qu’humainement, elle aurait pu l’être, si on l’avait
exterminée avec suite et sans merci. Mais les empereurs ne la
persécuteront par système qu’au second et au troisième siècle,
quand il sera trop tard pour la tuer. Et la persécution juive a été
brève, intermittente, indécise. Une puissance supérieure la
contrecarrait, la paralysait. Hérode aura beau tenir Pierre lié de deux
chaînes entre les soldats. Un Ange touchera les chaînes ; elles se
dénoueront ; d’elle-même, la porte de fer s’ouvrira. Et Saul, au
moment où il se croit victorieux de Jésus le Nazaréen, va devenir
son esclave, « le vase d’élection ».
II
SAUL LE VOYANT

SUR LA ROUTE DE DAMAS

Il est midi ; l’heure où, assis à l’entrée de sa tente, Abraham vit


tout d’un coup les trois hommes debout devant lui, l’heure où Jésus
vint s’asseoir, n’en pouvant plus, sur la margelle de la fontaine, et dit
à la femme qui puisait : « Donne-moi à boire. »
Un nuage poudreux chemine au long de la route, entre des rocs
brillants comme des cônes de sel. Voici la caravane se hâtant vers
Damas. Des lignes confuses de verdure, au pied de collines fauves,
jaillissent là-bas, du côté d’où souffle l’aquilon : les vergers dont la
ville est ceinte ! Depuis une semaine on marche. Enfin, c’est la
douceur du terme proche, la fraîcheur de l’eau humée dans l’air
dévorant. Les âniers pressent leurs bêtes ; les ombres des
chameaux porteurs de bagages se balancent moins lentes sur la
poussière qui brûle les yeux.
Escorté par des gens de police, bâtons en mains, Saul allonge le
pas. Petit [94] , mais impétueux, décisif, il s’avance, comme César, le
capitaine, d’une allure qui entraînerait derrière lui une armée. Sent-il
le poids du soleil sur sa tête ? Cet empire du feu qu’il traverse, ce
désert dont les roches semblent vibrer sous le choc des rayons, ce
mirage des vaines splendeurs existe à peine pour son regard. En
découvrant les murailles de Damas, se souvient-il qu’au temps
d’Abraham déjà, cette ville, patrie d’Éliézer, était un des grands
caravansérails de l’Orient ? Oui, peut-être ; mais l’idée qui le
tyrannise emporte dans sa frénésie toute notion des faits lointains.

You might also like