Information Processing in Medical

Imaging 24th International Conference

IPMI 2015 Sabhal Mor Ostaig Isle of
Skye UK June 28 July 3 2015
Proceedings 1st Edition Sebastien
Ourselin
Visit to download the full and correct content document:
https://textbookfull.com/product/information-processing-in-medical-imaging-24th-inter
national-conference-ipmi-2015-sabhal-mor-ostaig-isle-of-skye-uk-june-28-july-3-2015-
proceedings-1st-edition-sebastien-ourselin/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Health Information Science 4th International Conference

HIS 2015 Melbourne Australia May 28 30 2015 Proceedings
1st Edition Xiaoxia Yin

https://textbookfull.com/product/health-information-science-4th-
international-conference-his-2015-melbourne-australia-
may-28-30-2015-proceedings-1st-edition-xiaoxia-yin/

Mathematics of Program Construction 12th International

Conference MPC 2015 Königswinter Germany June 29 July 1
2015 Proceedings 1st Edition Ralf Hinze

https://textbookfull.com/product/mathematics-of-program-
construction-12th-international-conference-mpc-2015-konigswinter-
germany-june-29-july-1-2015-proceedings-1st-edition-ralf-hinze/

Natural Language Processing and Information Systems

20th International Conference on Applications of
Natural Language to Information Systems NLDB 2015
Passau Germany June 17 19 2015 Proceedings 1st Edition
Chris Biemann
https://textbookfull.com/product/natural-language-processing-and-
information-systems-20th-international-conference-on-
applications-of-natural-language-to-information-systems-
nldb-2015-passau-germany-june-17-19-2015-proceedings-1st-ed/

Advanced Information Systems Engineering 27th

International Conference CAiSE 2015 Stockholm Sweden
June 8 12 2015 Proceedings 1st Edition Jelena
Zdravkovic
https://textbookfull.com/product/advanced-information-systems-
engineering-27th-international-conference-caise-2015-stockholm-
sweden-june-8-12-2015-proceedings-1st-edition-jelena-zdravkovic/
Codes Cryptology and Information Security First
International Conference C2SI 2015 Rabat Morocco May 26
28 2015 Proceedings In Honor of Thierry Berger 1st
Edition Said El Hajji
https://textbookfull.com/product/codes-cryptology-and-
information-security-first-international-
conference-c2si-2015-rabat-morocco-may-26-28-2015-proceedings-in-
honor-of-thierry-berger-1st-edition-said-el-hajji/

Ad hoc Mobile and Wireless Networks 14th International

Conference ADHOC NOW 2015 Athens Greece June 29 July 1
2015 Proceedings 1st Edition Symeon Papavassiliou

https://textbookfull.com/product/ad-hoc-mobile-and-wireless-
networks-14th-international-conference-adhoc-now-2015-athens-
greece-june-29-july-1-2015-proceedings-1st-edition-symeon-
papavassiliou/

Proceedings of Second International Conference on

Electrical Systems Technology and Information 2015
ICESTI 2015 1st Edition Felix Pasila

https://textbookfull.com/product/proceedings-of-second-
international-conference-on-electrical-systems-technology-and-
information-2015-icesti-2015-1st-edition-felix-pasila/

Experimental Algorithms 14th International Symposium

SEA 2015 Paris France June 29 July 1 2015 Proceedings
1st Edition Evripidis Bampis (Eds.)

https://textbookfull.com/product/experimental-algorithms-14th-
international-symposium-sea-2015-paris-france-
june-29-july-1-2015-proceedings-1st-edition-evripidis-bampis-eds/

Artificial Intelligence in Education 17th International

Conference AIED 2015 Madrid Spain June 22 26 2015
Proceedings 1st Edition Cristina Conati

https://textbookfull.com/product/artificial-intelligence-in-
education-17th-international-conference-aied-2015-madrid-spain-
june-22-26-2015-proceedings-1st-edition-cristina-conati/
 Sebastien Ourselin
Daniel C. Alexander
Carl-Fredrik Westin
M. Jorge Cardoso (Eds.)
LNCS 9123

Information Processing
in Medical Imaging
24th International Conference, IPMI 2015
Sabhal Mor Ostaig, Isle of Skye, UK, June 28 – July 3, 2015
Proceedings

123
Lecture Notes in Computer Science 9123
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zürich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Sebastien Ourselin Daniel C. Alexander
•

Carl-Fredrik Westin M. Jorge Cardoso (Eds.)

•

Information Processing
in Medical Imaging
24th International Conference, IPMI 2015
Sabhal Mor Ostaig, Isle of Skye, UK, June 28 – July 3, 2015
Proceedings

123
Editors
Sebastien Ourselin Carl-Fredrik Westin
Centre for Medical Image Computing Dept. of Radiology
University College London Harvard Medical School Brigham
London and Women’s Hospital
UK Boston, MA
USA
Daniel C. Alexander
Centre for Medical Image Computing M. Jorge Cardoso
University College London Centre for Medical Image Computing
London University College London
UK London
UK

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-319-19991-7 ISBN 978-3-319-19992-4 (eBook)
DOI 10.1007/978-3-319-19992-4

Library of Congress Control Number: 2015940972

LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
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, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
 Preface

The 24th biennial international conference on Information Processing in Medical

Imaging (IPMI 2015) was held at the Sabhal Mor Ostaig College on the Isle of Skye,
UK, June 28 – July 3, 2015. The conference was the latest in a series of biennial
scientific meetings, the last being held in July of 2013 at the Asilomar Conference
Grounds near Pacific Grove, California, USA, during which new developments in
acquisition, formation, analysis, and display of medical images are presented, dis-
cussed, dissected, and extended.
IPMI is one of the longest running conferences devoted to these topics in medical
imaging. Over the last five decades, IPMI has evolved with the medical imaging
community it serves. The first IPMI conference was held in 1969, when a group of
young scientists working in nuclear medicine gathered to discuss the current problems
in their field. Since that time the conference has expanded into other medical imaging
acquisition modalities, including ultrasound, optics, magnetic resonance, and x-ray
imaging techniques. IPMI is now widely recognized as one of the most exciting and
influential meetings in medical imaging, with a unique emphasis on active participation
from all attendees and a strong commitment to vigorous discussion and open debate.
A wide variety of topics are covered at IPMI meetings, all within a single-track
format. This year 195 full-length manuscripts were submitted to the conference. Of
these, 22 papers were selected for oral presentation and 41 were accepted as posters.
Submissions were carefully reviewed by four members of the Scientific Review
Committee, who evaluated the novelty, methodological development, and scientific
rigor of each manuscript. The Paper Selection Committee, along with the co-chairs,
took on the difficult task of creating a meeting program. Using the rankings and
detailed comments of the reviewers of each manuscript, and adding to that their own
judgment of the merits of each manuscript, they designed the meeting program rep-
resented in this volume. The many high-quality manuscripts submitted for consider-
ation made the selection process extremely difficult, and many excellent papers did not
make it into the final program. This is an unfortunate inevitability of the high selectivity
for which IPMI is known.
One of the key goals of IPMI is to encourage participation of the most promising
and talented young researchers in the field, allowing them to explore new ideas with
some of the leading researchers in this area. Active involvement is stimulated by
preparation before the sessions in small study groups, in which everyone participates.
After reading the papers of their session, the study group members meet to discuss the
papers before they are presented, and to formulate questions and comments to kick off
the discussions. The study groups thus lead off the discussion of each paper, ensuring a
lively and vigorous dialog. The prestigious Francois Erbsmann Prize for first-time IPMI
presenters adds an extra stimulus for young researchers for active involvement in the
meeting. The Francois Erbsmann Prize is awarded for the best contribution by a young
VI Preface

scientist who is the first author of a paper and a first-time IPMI oral presenter. This year
16 (of 22) oral presenters were eligible for the Erbsmann Prize.
IPMI 2015 featured a keynote talk describing the Allen Brain atlas and its impact in
neuroimaging research by Boudewijn Lelieveldt, Professor of Biomedical Imaging at
Leiden University Medical Center, and Delft University of Technology, The
Netherlands.
The conference features the traditions that past IPMI attendees have come to expect
as part of the unique character of the meeting. Most importantly, each oral presentation
has unlimited time for discussion to give the audience the opportunity to resolve all
questions regarding the methods and results. IPMI discussions can go on for hours,
involving virtually every attendee in the room! Session chairs play a key role at IPMI
as they strive to ensure that these extended discussions are productive and continuing to
add to the group’s understanding of the nature and significance of each presentation’s
original contribution.
This year, also for the first time, the conference featured an elective oral session.
Meeting attendees voted for the poster presentations they would most like to see
presented orally. The IPMI board selected several poster presenters from those most
popular among the attendees. The selected presenters gave their oral presentation
during the closing session of the conference providing extra exposure to work of key
interest to the community.
IPMI is traditionally held in a small and sometimes remote location. This year’s
venue, the Sabhal Mor Ostaig on the Isle of Skye, Scotland, is a unique venue, situated
above the shoreline of the picturesque Sound of Sleat in Scotland, known worldwide as
the University for Gaelic Language. Sabhal Mor Ostaig is spread over two campuses,
Arainn Ostaig and Arainn Chaluim Chille. The two are interlinked by a well-lit
pathway and are only a short walk apart. The venue itself is conveniently located just
two miles north-east from the main ferry terminal at Armadale village. The venue has
great facilities: The Talla Dhonaidh Caimbeul lecture theater can seat up to 190 people,
there are two additional board rooms that can accommodate up to 25 people each, and
the restaurant consists of a glass-fronted dining room that offers magnificent views over
the Sound of Skye and the nearby Knoydart peninsula.
Conference delegates relaxed and enjoyed various group excursions on the after-
noon of day 3 of the conference. The excursions made the most of the unique location
and culture ranging from whale-spotting in the Sound of Skye, through axe-throwing
competitions, to visiting the Talisker distillery (the most famous of Skye’s malt
whiskies). The afternoon of day 4 of the conference saw the traditional IPMI soccer
match, pitting the American team against “the rest of the world.” After the disap-
pointing 0-0 draw in 2013, “the rest of the world” are still looking to avenge their
first-ever defeat in the history of IPMI in 2011 – an embarrassment that still scars deep.
These proceedings contain the IPMI 2015 papers in the order presented at the
meeting. We hope that this volume serves as a valuable source of information for the
participants, as well as a reminder of the great conference experience. For those who
were not able to attend the conference, we hope that these proceedings provide you
with an excellent summary of the latest research contributions to the medical imaging
field. We look forward to the next IPMI.
 Preface VII

The organization of the 24th IPMI conference was only possible through the efforts
and contributions of several organizations and many individuals. First of all, the IPMI
2015 co-chairs would like to thank the members of the Scientific Review Committee
for providing so many high-quality reviews within a very limited time frame; because
of these reviews, we were able to make a fair selection of the best papers for the final
program. We also express our gratitude to the Paper Selection Committee members,
who each read many papers and their reviews and travelled to London for a marathon
organizational meeting that resulted in an outstanding final program.
Thanks also to Microsoft Corporation for assistance with their excellent CMT
conference management system, which we used for the automation of submissions and
review of manuscripts. We also thank Jerry Prince for providing the template for this
front matter. Finally, we thank Ron Gaston and Dominique Drai for their invaluable
administrative support, without which the conference would most certainly not have
happened, and the many local organizers at Sabhal Mor Ostaig for their dedicated
support.

June 2015 Sebastien Ourselin

Daniel C. Alexander
Carl-Fredrik Westin
M. Jorge Cardoso
 Organization

Conference Chairs
Sebastien Ourselin University College London, UK
Daniel C. Alexander University College London, UK
Carl-Fredrik Westin Harvard University, USA

Organizing Committee
Ron Gaston University College London, UK
Dominique Drai University College London, UK

Paper Selection Committee

Chris Taylor Manchester University, UK
Marleen de Bruijne Copenhagen University, Denmark
Marc Niethammer University of North Carolina, USA
Sarang Joshi University of Utah, USA
M. Jorge Cardoso University College London, UK

Committee
Simon Arridge University College London, UK
John Ashburner University College London, UK
Suyash Awate University of Utah, USA
Christian Barillot IRISA/CNRS, France
Pierre-Louis Bazin Max Planck Gesellschaft, Germany
Ismail Ben Ayed Western University, USA
Sylvain Bouix BWH Harvard, USA
Djamal Boukerroui Mirada Medical, UK
Michael Brady University of Oxford, UK
Owen Carmichael University of California - Davis, USA
Gary Christensen University of Iowa, USA
Albert Chung HKSUT, HK
Ela Claridge The University of Birmingham, UK
Olivier Commowick Inria Rennes, France
Tim Cootes Manchester University, UK
Sune Darkner DIKU, Denmark
Benoit Dawant Vanderbilt University, USA
Rachid Deriche Inria Sophia Antipolis, France
X Organization

Maxime Descoteaux Sherbrooke, Canada

Jim Duncan Yale University, USA
Stanley Durrleman Inria/ICM, France
Yong Fan University of Pennsylvania, USA
Aaron Fenster Robarts Research Institute, USA
Aasa Feragen DIKU, Denmark
Thomas Fletcher University of Utah, USA
Alejandro Frangi University of Sheffield, UK
James Gee University of Pennsylvania, USA
Guido Gerig University of Utah, USA
Ben Glocker Imperial College, UK
Polina Golland Massachusetts Institute of Technology, USA
Michael Goris Stanford University School of Medicine, USA
Matthias Guenther Fraunhofer MEVIS, Germany
Horst Hahn Fraunhofer MEVIS, Germany
Justin Haldar University of Southern California, USA
Dave Hawkes University College London, UK
Mattias Heinrich University of Lübeck, Germany
Joachim Hornegger University of Erlangen, Germany
Brian F. Hutton University College London, UK
Juan E. Iglesias Basque Center on Cognition, Brain and Language,
Spain
Ivana Isgum University Medical Center Utrecht, The Netherlands
Anand Joshi University of Southern California, USA
Enrico Kaden University College London, UK
Nico Karssemeijer Radboud University, The Netherlands
Boklye Kim University of Michigan, USA
Andy King King’s College London, UK
Ender Konukoglu Massachusetts General Hospital, USA
Frithjof Kruggel University of California Irvine, USA
Sebastian Kurtek The Ohio State University, USA
Jan Kybic Czech Technical University, Czech Republic
Georg Langs Medical University of Vienna, Austria
Tobias Lasser Technical University of Munich, Germany
Richard Leahy University of Southern California, USA
Boudewijn Lelieveldt Leiden University Medical Center, The Netherlands
Marco Lorenzi University College London, UK
Frederik Maes Katholieke Universiteit Leuven, Belgium
Stephen Marsland Massey University, New Zealand
Bernard Ng Stanford University, USA
Mads Nielsen University of Copenhagen, Denmark
Wiro Niessen Erasmus Medical Center Rotterdam, The Netherlands
Marc Niethammer UNC Chapel Hill, USA
Alison Noble University of Oxford, UK
Lauren O’Donnell Harvard Medical School, USA
Evren Ozarslan Harvard University, USA
 Organization XI

Xavier Pennec Inria, France

Jens Petersen University of Copenhagen, Denmark
Dzung Pham Center for Neuroscience and Regenerative Medicine,
USA
Stephen Pizer The University of North Carolina at Chapel Hill, USA
Kilian Pohl SRI/Stanford, USA
Marcel Prastawa University of Utah, USA
Jerry Prince Johns Hopkins University, USA
Jinyi Qi University of California Davis, USA
Anqi Qiu National University of Singapore, Singapore
Yogesh Rathi BWH Harvard, USA
Joseph Reinhardt University of Iowa, USA
Emma Robinson Oxford University, UK
Karl Rohr University of Heidelberg, Germany
Daniel Rueckert Imperial College London, UK
Mert Sabuncu Massachusetts General Hospital, USA
Benoit Scherrer Boston Children’s Hospital, USA
Julia Schnabel University of Oxford, UK
Christof Seiler Stanford University, USA
Pengcheng Shi Rochester Institute of Technology, USA
Kaleem Siddiqi McGill University, Canada
Nikhil P. Singh University of North Carolina Chapel Hill, USA
Lawrence Staib Yale University, USA
Colin Studholme Washington University, USA
Martin Styner University of North Carolina at Chapel Hill, USA
Gabor Szekely ETH Zurich, Switzerland
Maxime Taquet Boston Children’s Hospital, USA
Bertrand Thirion Inria, France
Matthew Toews Harvard Medical School, USA
Carole Twining The University of Manchester, UK
Koen Van Leemput Harvard Medical School/Massachusetts General
Hospital, USA
Gael Varoquaux Inria, France
Baba Vemuri University of Florida, USA
Tom Vercauteren University College London, UK
Hongzhi Wang IBM Almaden Research Center, USA
Simon Warfield Harvard Medical School and Children’s Hospital, USA
Demian Wassermann Inria, France
William M. Wells BWH Harvard, USA
Ross Whittaker University of Utah, USA
Pew-Thian Yap The University of North Carolina at Chapel Hill, USA
Paul Yushkevich University of Pennsylvania, USA
Gary Zhang University College London, UK
Kevin Zhou Siemens Corporate Research, USA
Lilla Zollei Harvard Medical School, USA
XII Organization

IPMI Board

Harrison H.Barrett Richard M. Leahy

Christian Barillot Stephen M. Pizer
Aaron B. Brill Jerry L. Prince
Gary E. Christensen Gåbor Székely
Alan C.F. Colchester Chris Taylor
James S. Duncan Andrew Todd-Pokropek
Michael L. Goris William M. Wells
Nico Karssemeijer
 Contents

Probabilistic Graphical Models

Colocalization Estimation Using Graphical Modeling and Variational
Bayesian Expectation Maximization: Towards a Parameter-Free Approach . . . 3
Suyash P. Awate and Thyagarajan Radhakrishnan

Template-Based Multimodal Joint Generative Model of Brain Data . . . . . . . . 17

M. Jorge Cardoso, Carole H. Sudre, Marc Modat,
and Sebastien Ourselin

Generative Method to Discover Genetically Driven Image Biomarkers. . . . . . 30

Nematollah K. Batmanghelich, Ardavan Saeedi, Michael Cho,
Raul San Jose Estepar, and Polina Golland

MRI Reconstruction

A Joint Acquisition-Estimation Framework for MR Phase Imaging . . . . . . . . 45

Joseph Dagher

A Compressed-Sensing Approach for Super-Resolution Reconstruction

of Diffusion MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Lipeng Ning, Kawin Setsompop, Oleg Michailovich, Nikos Makris,
Carl-Fredrik Westin, and Yogesh Rathi

Accelerated High Spatial Resolution Diffusion-Weighted Imaging . . . . . . . . . 69

Benoit Scherrer, Onur Afacan, Maxime Taquet, Sanjay P. Prabhu,
Ali Gholipour, and Simon K. Warfield

Clustering

Joint Spectral Decomposition for the Parcellation of the Human Cerebral

Cortex Using Resting-State fMRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Salim Arslan, Sarah Parisot, and Daniel Rueckert

Joint Clustering and Component Analysis of Correspondenceless Point

Sets: Application to Cardiac Statistical Modeling . . . . . . . . . . . . . . . . . . . . 98
Ali Gooya, Karim Lekadir, Xenia Alba, Andrew J. Swift,
Jim M. Wild, and Alejandro F. Frangi
XIV Contents

Statistical Methods

Bootstrapped Permutation Test for Multiresponse Inference on Brain

Behavior Associations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Bernard Ng, Jean Baptiste Poline, Bertrand Thirion,
Michael Greicius, and IMAGEN Consortium

Controlling False Discovery Rate in Signal Space

for Transformation-Invariant Thresholding of Statistical Maps. . . . . . . . . . . . 125
Junning Li, Yonggang Shi, and Arthur W. Toga

Longitudinal Analysis

Group Testing for Longitudinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Yi Hong, Nikhil Singh, Roland Kwitt, and Marc Niethammer

Spatio-Temporal Signatures to Predict Retinal Disease Recurrence . . . . . . . . 152

Wolf-Dieter Vogl, Sebastian M. Waldstein, Bianca S. Gerendas,
Christian Simader, Ana-Maria Glodan, Dominika Podkowinski,
Ursula Schmidt-Erfurth, and Georg Langs

Microstructure Imaging

A Unifying Framework for Spatial and Temporal Diffusion

in Diffusion MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Rutger Fick, Demian Wassermann, Marco Pizzolato,
and Rachid Deriche

Ground Truth for Diffusion MRI in Cancer: A Model-Based Investigation

of a Novel Tissue-Mimetic Material . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Damien J. McHugh, Fenglei Zhou, Penny L. Hubbard Cristinacce,
Josephine H. Naish, and Geoffrey J.M. Parker

Shape Analysis

Anisotropic Distributions on Manifolds: Template Estimation and Most

Probable Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Stefan Sommer

A Riemannian Framework for Intrinsic Comparison of Closed

Genus-Zero Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Boris A. Gutman, P. Thomas Fletcher, M. Jorge Cardoso,
Greg M. Fleishman, Marco Lorenzi, Paul M. Thompson,
and Sebastien Ourselin
 Contents XV

Multi-atlas Fusion

Multi-atlas Segmentation as a Graph Labelling Problem: Application

to Partially Annotated Atlas Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Lisa M. Koch, Martin Rajchl, Tong Tong, Jonathan Passerat-Palmbach,
Paul Aljabar, and Daniel Rueckert

Keypoint Transfer Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

C. Wachinger, M. Toews, G. Langs, W. Wells, and P. Golland

Fast Image Registration

Finite-Dimensional Lie Algebras for Fast Diffeomorphic

Image Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Miaomiao Zhang and P. Thomas Fletcher

Fast Optimal Transport Averaging of Neuroimaging Data . . . . . . . . . . . . . . 261

A. Gramfort, G. Peyré, and M. Cuturi

Deformation Models

Joint Morphometry of Fiber Tracts and Gray Matter Structures Using

Double Diffeomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Pietro Gori, Olivier Colliot, Linda Marrakchi-Kacem, Yulia Worbe,
Alexandre Routier, Cyril Poupon, Andreas Hartmann, Nicholas Ayache,
and Stanley Durrleman

A Robust Probabilistic Model for Motion Layer Separation in X-ray

Fluoroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
Peter Fischer, Thomas Pohl, Thomas Köhler, Andreas Maier,
and Joachim Hornegger

Poster Papers

Weighted Hashing with Multiple Cues for Cell-Level Analysis

of Histopathological Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
Xiaofan Zhang, Hai Su, Lin Yang, and Shaoting Zhang

Multiresolution Diffeomorphic Mapping for Cortical Surfaces. . . . . . . . . . . . 315

Mingzhen Tan and Anqi Qiu

A Comprehensive Computer-Aided Polyp Detection System

for Colonoscopy Videos. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Nima Tajbakhsh, Suryakanth R. Gurudu, and Jianming Liang
XVI Contents

A Feature-Based Approach to Big Data Analysis of Medical Images . . . . . . . 339

Matthew Toews, Christian Wachinger, Raul San Jose Estepar,
and William M. Wells III

Joint Segmentation and Registration Through the Duality of Congealing

and Maximum Likelihood Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Boris Flach and Archibald Pontier

Self-Aligning Manifolds for Matching Disparate Medical Image Datasets . . . . 363

Christian F. Baumgartner, Alberto Gomez, Lisa M. Koch,
James R. Housden, Christoph Kolbitsch, Jamie R. McClelland,
Daniel Rueckert, and Andy P. King

Leveraging EAP-Sparsity for Compressed Sensing of MS-HARDI

in ðk; qÞ-Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Jiaqi Sun, Elham Sakhaee, Alireza Entezari, and Baba C. Vemuri

Multi-stage Biomarker Models for Progression Estimation

in Alzheimer’s Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Alexander Schmidt-Richberg, Ricardo Guerrero, Christian Ledig,
Helena Molina-Abril, Alejandro F. Frangi, Daniel Rueckert,
and on behalf of the Alzheimers Disease Neuroimaging Initiative

Measuring Asymmetric Interactions in Resting State Brain Networks. . . . . . . 399

Anand A. Joshi, Ronald Salloum, Chitresh Bhushan,
and Richard M. Leahy

Shape Classification Using Wasserstein Distance for Brain

Morphometry Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Zhengyu Su, Wei Zeng, Yalin Wang, Zhong-Lin Lu, and Xianfeng Gu

Temporal Trajectory and Progression Score Estimation from Voxelwise

Longitudinal Imaging Measures: Application to Amyloid Imaging. . . . . . . . . 424
Murat Bilgel, Bruno Jedynak, Dean F. Wong, Susan M. Resnick,
and Jerry L. Prince

Predicting Semantic Descriptions from Medical Images with Convolutional

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Thomas Schlegl, Sebastian M. Waldstein, Wolf-Dieter Vogl,
Ursula Schmidt-Erfurth, and Georg Langs

Bodypart Recognition Using Multi-stage Deep Learning . . . . . . . . . . . . . . . 449

Zhennan Yan, Yiqiang Zhan, Zhigang Peng, Shu Liao,
Yoshihisa Shinagawa, Dimitris N. Metaxas, and Xiang Sean Zhou
 Contents XVII

Multi-subject Manifold Alignment of Functional Network Structures via

Joint Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
Karl-Heinz Nenning, Kathrin Kollndorfer, Veronika Schöpf,
Daniela Prayer, and Georg Langs

Brain Transfer: Spectral Analysis of Cortical Surfaces

and Functional Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
Herve Lombaert, Michael Arcaro, and Nicholas Ayache

Finding a Path for Segmentation Through Sequential Learning . . . . . . . . . . . 488

Hongzhi Wang, Yu Cao, and Tanveer F. Syed-Mahmood

Pancreatic Tumor Growth Prediction with Multiplicative Growth

and Image-Derived Motion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
Ken C.L. Wong, Ronald M. Summers, Electron Kebebew,
and Jianhua Yao

IMaGe: Iterative Multilevel Probabilistic Graphical Model for Detection

and Segmentation of Multiple Sclerosis Lesions in Brain MRI . . . . . . . . . . . 514
Nagesh Subbanna, Doina Precup, Douglas Arnold, and Tal Arbel

Moving Frames for Heart Fiber Reconstruction. . . . . . . . . . . . . . . . . . . . . . 527

Emmanuel Piuze, Jon Sporring, and Kaleem Siddiqi

Detail-Preserving PET Reconstruction with Sparse Image Representation

and Anatomical Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
Jieqing Jiao, Pawel Markiewicz, Ninon Burgos, David Atkinson,
Brian Hutton, Simon Arridge, and Sebastien Ourselin

Automatic Detection of the Uterus and Fallopian Tube Junctions

in Laparoscopic Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
Kristina Prokopetc, Toby Collins, and Adrien Bartoli

A Mixed-Effects Model with Time Reparametrization for Longitudinal

Univariate Manifold-Valued Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
J.-B. Schiratti, S. Allassonnière, A. Routier, the Alzheimers Disease
Neuroimaging Initiative, O. Colliot, and S. Durrleman

Prediction of Longitudinal Development of Infant Cortical Surface Shape

Using a 4D Current-Based Learning Framework . . . . . . . . . . . . . . . . . . . . . 576
Islem Rekik, Gang Li, Weili Lin, and Dinggang Shen

Multi-scale Convolutional Neural Networks for Lung Nodule

Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
Wei Shen, Mu Zhou, Feng Yang, Caiyun Yang, and Jie Tian
XVIII Contents

Tractography-Driven Groupwise Multi-scale Parcellation of the Cortex . . . . . 600

Sarah Parisot, Salim Arslan, Jonathan Passerat-Palmbach,
William M. Wells III, and Daniel Rueckert

Illumination Compensation and Normalization Using Low-Rank

Decomposition of Multispectral Images in Dermatology. . . . . . . . . . . . . . . . 613
Alexandru Duliu, Richard Brosig, Saahil Ognawala, Tobias Lasser,
Mahzad Ziai, and Nassir Navab

Efficient Gaussian Process-Based Modelling and Prediction

of Image Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626
Marco Lorenzi, Gabriel Ziegler, Daniel C. Alexander,
and Sebastien Ourselin for ADNI

A Simulation Framework for Quantitative Validation of Artefact Correction

in Diffusion MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
Mark S. Graham, Ivana Drobnjak, and Hui Zhang

Towards a Quantified Network Portrait of a Population . . . . . . . . . . . . . . . . 650

Birkan Tunç, Varsha Shankar, Drew Parker, Robert T. Schultz,
and Ragini Verma

Segmenting the Brain Surface from CT Images with Artifacts Using

Dictionary Learning for Non-rigid MR-CT Registration . . . . . . . . . . . . . . . . 662
John A. Onofrey, Lawrence H. Staib, and Xenophon Papademetris

AxTract: Microstructure-Driven Tractography Based on the Ensemble

Average Propagator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Gabriel Girard, Rutger Fick, Maxime Descoteaux, Rachid Deriche,
and Demian Wassermann

Sampling from Determinantal Point Processes for Scalable

Manifold Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
Christian Wachinger and Polina Golland

Model-Based Estimation of Microscopic Anisotropy in Macroscopically

Isotropic Substrates Using Diffusion MRI . . . . . . . . . . . . . . . . . . . . . . . . . 699
Andrada Ianuş, Ivana Drobnjak, and Daniel C. Alexander

Multiple Orderings of Events in Disease Progression . . . . . . . . . . . . . . . . . . 711

Alexandra L. Young, Neil P. Oxtoby, Jonathan Huang,
Razvan V. Marinescu, Pankaj Daga, David M. Cash, Nick C. Fox,
Sebastien Ourselin, Jonathan M. Schott, Daniel C. Alexander,
and for the Alzheimers Disease Neuroimaging Initiative
 Contents XIX

Construction of An Unbiased Spatio-Temporal Atlas of the Tongue

During Speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
Jonghye Woo, Fangxu Xing, Junghoon Lee, Maureen Stone,
and Jerry L. Prince

Tree-Encoded Conditional Random Fields for Image Synthesis . . . . . . . . . . . 733

Amod Jog, Aaron Carass, Dzung L. Pham, and Jerry L. Prince

Simultaneous Longitudinal Registration with Group-Wise Similarity Prior . . . 746

Greg M. Fleishman, Boris A. Gutman, P. Thomas Fletcher,
and Paul M. Thompson

Spatially Weighted Principal Component Regression for High-Dimensional

Prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 758
Dan Shen and Hongtu Zhu

Coupled Stable Overlapping Replicator Dynamics for Multimodal Brain

Subnetwork Identification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
Burak Yoldemir, Bernard Ng, and Rafeef Abugharbieh

Joint 6D k-q Space Compressed Sensing for Accelerated High Angular

Resolution Diffusion MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
Jian Cheng, Dinggang Shen, Peter J. Basser, and Pew-Thian Yap

Functional Nonlinear Mixed Effects Models for Longitudinal Image Data . . . 794
Xinchao Luo, Lixing Zhu, Linglong Kong, and Hongtu Zhu

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807

Probabilistic Graphical Models
 Colocalization Estimation Using Graphical
Modeling and Variational Bayesian Expectation
Maximization: Towards a Parameter-Free
Approach

Suyash P. Awate(B) and Thyagarajan Radhakrishnan

Computer Science and Engineering Department, Indian Institute of Technology (IIT)

Bombay, Mumbai, India
suyash@cse.iitb.ac.in

Abstract. In microscopy imaging, colocalization between two biological

entities (e.g., protein-protein or protein-cell) refers to the (stochastic)
dependencies between the spatial locations of the two entities in the
biological specimen. Measuring colocalization between two entities relies
on fluorescence imaging of the specimen using two fluorescent chemicals,
each of which indicates the presence/absence of one of the entities at any
pixel location. State-of-the-art methods for estimating colocalization rely
on post-processing image data using an adhoc sequence of algorithms
with many free parameters that are tuned visually. This leads to loss of
reproducibility of the results. This paper proposes a brand-new frame-
work for estimating the nature and strength of colocalization directly
from corrupted image data by solving a single unified optimization prob-
lem that automatically deals with noise, object labeling, and parame-
ter tuning. The proposed framework relies on probabilistic graphical
image modeling and a novel inference scheme using variational Bayesian
expectation maximization for estimating all model parameters, includ-
ing colocalization, from data. Results on simulated and real-world data
demonstrate improved performance over the state of the art.

Keywords: Microscopy · Colocalization · Probabilistic graphical mod-

els · Expectation maximization · Variational Bayesian inference · Mean
field · Pseudo likelihood

1 Introduction

Colocalization [10,20] between two biological entities refers to the (stochastic)

dependencies between the spatial locations of the two entities in the specimen.
For instance, the presence of a protein at a specific location may be codepen-
dent with the presence of a specific cell type in the spatial neighborhood [14].
We thank funding via the IIT Bombay Seed Grant 14IRCCSG010 and T. Liou for
the data.

c Springer International Publishing Switzerland 2015
S. Ourselin et al. (Eds.): IPMI 2015, LNCS 9123, pp. 3–16, 2015.
DOI: 10.1007/978-3-319-19992-4 1
4 S.P. Awate and T. Radhakrishnan

The degree of codependence in the spatial distributions of two proteins can reveal
details about their dynamic interaction and help understand several biochemi-
cal processes [4,8]. Such studies are crucial in understanding the mechanisms of
several debilitating diseases, including cancer.
The semantics of colocalization has two aspects: (i) co-occurrence, i.e., the
presence of the two entities within a spatial neighborhood, and (ii) correlation or
codependence, where the two entities co-occur and their local quantities exhibit
a functional relationship. This paper focuses on the latter, more general, aspect
and proposes a framework to automatically estimate the nature and strength of
such colocalization from image data.
In microscopy, the measurement of colocalization typically relies on fluores-
cence imaging that involves simultaneous or successive detection of multiple
fluorescent chemicals in the same biological specimen, where each fluorescent
chemical is designed to bind to exactly one of the biological entities. Thus, the
imaging produces multiple images, lying in the same physical coordinate space,
where the intensity at each pixel in an image is indicative of the presence/absence
of the corresponding biological entity at that pixel location. This paper proposes
to estimate a quantitative measure of colocalization directly from noisy image
data involving two fluorescence microscopy images.
The literature describes several different strategies for formulating and esti-
mating quantitative measures of colocalization [10,11,20,23]. However, none of
these measure colocalization directly from image data, but only from a post-
processed version of the data. This post-processing is typically performed man-
ually and in an adhoc manner, i.e., via an adhoc sequence of algorithms with
several free parameters tuned visually. This leads to loss of reproducibility of the
results. In contrast, this paper proposes a brand new framework for colocaliza-
tion estimation directly from unprocessed data, by solving a single optimization
problem with data-driven parameter optimization for all model parameters. This
framework includes (i) a novel quantitative measure of colocalization, incorpo-
rated as a model parameter, and (ii) a novel scheme for statistical inference.

2 Related Work

Quantification of the nature and strength of colocalization is challenging and

continues to be an area of active research [1,6,8,9,11,18,22,23,28]. Conventional
colocalization techniques were typically qualitative and, thereby, highly prone to
random error and bias stemming from inter-rater and intra-rater variability [28].
Colocalization measures involving pixelwise analysis [1,6,8,9,18] typically
rely on (i) cross-correlation (CC) of the intensities in the two images or (ii) clus-
tering of the joint intensity histogram of the two images. Such methods, however,
severely fail to account for long-range spatial dependencies when the two entities
occur in close proximity but (almost) never occur in the same pixel. Further-
more, background pixels artificially inflate CC because both image intensities are
significantly below their average levels [10]. Some methods [10,27] address this
effect by excluding a background region from the analysis, but the background
 Colocalization Estimation Using Graphical Modeling and VBEM 5

detection uses an adhoc sequence of semi-supervised algorithms, thereby making

it highly inconsistent across experts.
On the other hand, objectwise analysis [11,22,23] involves a two-step process
that first detects and delineates entities/objects represented in the images and
then, independently, quantifies colocalization by measuring overlaps or nearest-
neighbor distances. However, these methods (i) rely on hard segmentations,
(ii) use segmentation methods involving several disconnected algorithms and
free parameters (e.g., thresholds, smoothing levels, filter parameters, etc.) that
are difficult to tune, and (iii) fail to improve segmentation by feeding back knowl-
edge of the estimated colocalization. Lack of data-driven parameter tuning typ-
ically leads to reported results that are inconsistent and irreproducible [28]. On
the other hand, the proposed framework in this paper (i) treats segmentation
labels as hidden random variables and, thereby, accounts for the uncertainty in
the optimal segmentation, (ii) automatically estimates all crucial parameters in
a data-driven manner, and (iii) integrates the tasks of segmentation and colo-
calization estimation as a single unified estimation problem given noisy data.
The noise model underlying fluorescence microscopy involves a complicated
combination of several statistical distributions. While the intrinsic noise obeys
the Poisson process, several sources of extrinsic noise are characterized by
Poisson, Gaussian, and uniform distributions [25]. Indeed, denoising fluorescence
microscopy images is an active research area [12,17]. Many state-of-the-art meth-
ods approximate the noise model by an independent and identically-distributed
(i.i.d.) additive Gaussian model to achieve numerical tractability without signif-
icant loss of efficacy [25]. This approximation hold good when the signal-to-noise
ratio is high. We follow the same strategy.
In broader contexts, some segmentation methods involve estimation with
(hidden) Markov random field (MRF) modeling and expectation maximization
(EM) [2,3,7,13,16,19,21,24,26]. However, virtually all such works manually tune
the MRF smoothness parameter, unlike the proposed approach that estimates
all MRF parameters from the data. Moreover, while some methods [26] employ
adhoc gross approximations for evaluating the E step, others [13] evaluate the E
step via Gibbs sampling that is computationally expensive and that poses well-
known theoretical problems of selecting optimal burn-in periods and detecting
sampler convergence. In contrast, the proposed approach employs principled and
efficient variational inference to evaluate the E step.
This paper makes several contributions. First, it proposes a brand new frame-
work for estimating colocalization directly from image data by solving a single
optimization problem that automatically deals with noise, object labeling, and
parameter tuning. Consequently, unlike previous works, the proposed framework
eliminates (i) adhoc pipelines of disconnected algorithms and (ii) adhoc tuning of
parameters, thereby enabling reproducible results. Second, the proposed frame-
work relies on (i) a novel probabilistic graphical model for image data and (ii) a
novel efficient inference scheme using variational Bayesian expectation maximiza-
tion (VBEM) [5] for data-driven estimation of all model parameters. Specifically,
it derives a novel combination of mean-field theory and pseudo-likelihood estima-
tion for parameter inference. Third, the results on simulated and real-world data
6 S.P. Awate and T. Radhakrishnan

demonstrate that the proposed method estimates colocalization more accurately

and consistently, compared to the state of the art.

3 Methods

We first formulate the problem of quantifying colocalization strength and nature

(positive/negative codependence) as a Bayesian estimation problem. Then, we
describe a novel VBEM inference algorithm that estimates all model parameters
from data.

3.1 Bayesian Random-Field Image Modeling

We present a generative model for a pair of fluorescence microscopy images to

measure colocalization of two objects, relying on probabilistic graphical modeling
(Fig. 1).

Fig. 1. (a) Real-world fluorescence microscopy images z 1 (top) and z 2 (bottom),

depict the presence of protein and cell nuclei, respectively. (b) Proposed Proba-
bilistic Graphical Model for Colocalization Estimation. Filled circles denote
observed image data {z 1 , z 2 }. Label images {L1 , L2 } are hidden random variables. We
estimate all parameters θ from data. The nature and strength of colocalization is mod-
eled by parameter ρ. Parameters β 1 , β 2 model spatial smoothness of the labels in l1 , l2 .
In this paper, for simplicity, we use a single β = β 1 = β 2 . Means and standard devia-
tions μ·· , σ ·· model image intensities of objects and backgrounds in z 1 , z 2 . (c) MAP
label images l1∗ , l2∗ , indicating object presence /absence, computed after optimal para-
meters θ∗ are found by VBEM. Note: this paper focuses on estimating colocalization ρ
using VBEM, which is crucial to several clinical and scientific studies; the MAP label
images are unused by VBEM and are shown only to provide additional insights into
the proposed approach.
 Colocalization Estimation Using Graphical Modeling and VBEM 7

Let L := {L1 , L2 } denote a random field modeling the pixel labels of the two
objects. Specifically, let L1 be a binary label image with I pixels {L1i }Ii=1 with
possible label values ∈ {+1, −1}, where li1 = +1 indicates the object’s presence
at pixel i in l1 ; similarly, li1 = −1 indicates the object’s absence at pixel i in l1 .
We follow similar notations and semantics for L2 . We assume that the spatial
distributions of objects are (i) smooth within each image and (ii) codependent
between images. To model these properties, each pixel has intra-image and inter-
image neighbors.
We introduce a neighborhood system N := {Niself , Niother }Ii=1 , where Niself
and Niother denote sets of neighbors of pixel i in the same image (to which pixel
i belongs) and the other image, respectively. To have isotropic neighborhoods,
we employ Gaussian-weighted masks w where the ij-th element wij := exp(− 
xj − xi 2 /α2 ), where xi and xj are physical coordinates of pixels i and j,
respectively. We use two such masks, i.e., wself (for intra-image smoothness)
and wother (for inter-image colocalization) with underlying parameters αself and
αother , respectively. We set αself := 2/3 pixels to restrict the neighborhood to 2
neighbors along each dimension (we restrict the neighborhood so that wij > 10−3
for neighbors i, j); we found that this level of smoothness suffices for the data
used in this paper. We set αother to be large enough to be able to model the
longest-range direct spatial interactions between the biological entities; this is
application dependent and informed by prior knowledge, but the results are
fairly robust to the choice of this parameter. In this paper, αother := 2 pixels.
We model the prior probability mass function (PMF) of the label image pair
L as
⎛ ⎞
 I  I 
1 11 
P (l) := exp ⎝ β li lj + li2 lj2 wij
self
+ other ⎠
ρli1 lj2 wij ,
Z(ρ, β) i=1 self i=1 other
j∈Ni j∈Ni
(1)

where the normalization constant (partition function) Z(ρ, β) is a function of

the smoothness parameter β ∈ R+ and the colocalization parameter ρ ∈ R. The
motivation for this model is as follows. The terms involving β are derived from
a standard Ising model of smoothness on the label images. The term involving
ρ is novel and carefully designed as follows. First, ρ = 0 decouples the two label
images such that P (l) can be written as P (l1 )P (l2 ); this indicates absence of
colocalization. Second, because the labels are designed to be ±1, a positive ρ  0
forces the interacting labels in the two images l1 , l2 (i.e., labels in neighborhoods
of each other) towards being equal (i.e., both +1 or both −1) to produce a
high label image probability. This implies that the neighboring labels in the
two images both likely indicate the presence of the objects or both indicate the
absence of the objects. Third, similarly, a negative ρ  0 forces the interacting
labels in the two images to be negative of each other, i.e., the presence of the
object on one image indicates the absence of the other object in the other image.
Let Z := {Z 1 , Z 2 } denote a random field modeling the intensities of the two
objects and the backgrounds. We assume that the intensities of each object are
8 S.P. Awate and T. Radhakrishnan

Gaussian distributed as a result of natural intensity variation, partial-volume

effects, minor shading artifacts, etc. We assume that the data is corrupted by
i.i.d. additive Gaussian noise. Let the intensities of the object and the back-
ground in the observed image z k , where k ∈ {1, 2}, be Gaussian distributed
with parameters {μk1 , σ k1 } and {μk2 , σ k2 }, respectively. Then, the likelihood of
observing data z, given labels l, is
I I
1 1 2 2
P (z|l) := G(zi1 |μ1ti , σ 1ti ) G(zi2 |μ2ti , σ 2ti ), (2)
i=1 i=1

where t1i maps the label value li1 to the object number, i.e., t1i := 1 if li1 = −1
and t1i := 2 if li1 = +1; similarly for t2i .
Let the set of parameters be θ := θP ∪ θL , where the prior parameters are
θP := {ρ, β} and the likelihood parameters are θL := {μ11 , σ 11 , μ12 , σ 12 , μ21 ,
σ 21 , μ22 , σ 22 }. We propose to formulate colocalization estimation as the following
maximum-a-posteriori Bayesian estimation problem over all parameters θ:
 
arg max max P (z|θ) , where P (z|θ) = P (z, l|θ) = P (z|l, θL )P (l|θP ).
ρ θ−{ρ}
l l
(3)
A novelty in our approach is that we estimate ρ (as well as β) automatically
from the data. We observe that the optimization of the parameters ρ and β is
non-trivial because of their involvement in the partition function Z(ρ, β) that is
intractable to evaluate. Furthermore, we optimize all parameters underlying the
model directly from the data.

3.2 Expectation Maximization for Parameter Estimation

We treat the labels L as hidden variables and solve the parameter-estimation
problem using EM. EM is an iterative optimization algorithm, each iteration
comprising an E step and an M step. Consider that after iteration n, the para-
meter estimate is θn . Then, in iteration n + 1, the updated parameter estimate
θn+1 is obtained as follows.
The E step involves constructing the Q(·) function
Q(θ; θn ) := EP (L|z,θn ) [log P (z, L|θ)] (4)
that is intractable to evaluate analytically. Furthermore, a Monte-Carlo
simulation-based approximation of the expectation leads to challenges in
reliable and efficient sampling; e.g., while Gibbs sampling of the label fields
is computationally expensive, determining an appropriate burn-in period and
detecting convergence of the Gibbs sampler pose serious theoretical challenges.
Thus, we opt for variational inference and choose to approximate the posterior
PMF P (L|z, θn ) by a factorizable analytical model
I
n n
P (L|z, θ ) ≈ F (L|z, θ ) := Fi1 (L1i )Fi2 (L2i ), (5)
i=1
 Colocalization Estimation Using Graphical Modeling and VBEM 9

where we optimize the per-pixel factors, i.e., PMFs, Fi1 (L1i ), Fi2 (L2i ) to best fit
the posterior. This factorized approximation for the posterior makes the Q(θ; θn )
function more tractable. We describe this strategy in the next section.
The M step subsequently maximizes the Q(θ; θn ) function over parameters θ.
The E and M steps are repeated until convergence; we observe that a few itera-
tions are suffice.

3.3 E Step: Variational Inference of the Factorized Posterior

Within each EM iteration, we use variational inference to optimize the factors
underlying F (L|z, θn ). To simplify notation, we omit θ where it is obvious. We
rewrite

log P (Z|θ) = L(F (L), θ) + KL F (L)  P (L|z, θ) , where (6)


L(F (L), θ) := F (l) log P (z, l|θ)/F (l) (7)
l

and KL(A  B) ≥ 0 denotes the Kullback-Leibler (KL) divergence between

distributions A and B. Our goal is to find the maximal lower bound L(F (L)) of
the data log-probability log P (Z), under the factorization constraint on F (L).
We do this by maximizing L(F (L)), because KL(·) ≥ 0 guarantees L(F (L)) to be
a lower bound. We perform this functional optimization by iterative optimization
of each of the factor functions; this is possible because the factors are designed
to be independent of each other. We now incorporate the factorized form of
F (L) and, without loss of generality, separate terms involving Fj1 from the other
terms. This yields
 
L(F (L)) = Fj1 (lj1 ) log P(z, lj1 ) − Fj1 (lj1 ) log Fj1 (lj1 ) + constant, (8)
lj1 lj1

where log P(z, lj1 ) := log P (z, l) Fi1 (li1 ) Fi2 (li2 ) + constA, (9)
1
li:i =j
l2 i:i=j i

where (i) the first term on the right hand side is an expectation of the log
complete-data probability log P (z, l) over all the factors in F (L) except Fj1 (·)
and (ii) constA is the normalization factor for P (z, L1j ). We see that L(F (L)) is
the negative KL divergence between the distributions Fj1 (L1j ) and log P(z, L1j ).
Thus, fixing the set of functions {Fi1=j (·), Fi2 (·)}Ii=1 , the optimal function Fj1 (·)
that maximizes L(F (L)) is
⎛ ⎞
1 
Fj1 (lj1 ) := P(z, lj1 ) = exp ⎝ log P (z, l) Fi1 (li1 ) Fi2 (li2 )⎠, (10)
η i
1 2 li:i=j l i:i=j

where η is the normalization constant. The expectation inside the exponent on

the right hand side is easy to evaluate for the following reasons. First, because
10 S.P. Awate and T. Radhakrishnan

we designed P (z, l) to be in the exponential class, log P (z, l) gives us the terms
in its exponent. Second, because we are estimating a PMF solely on the label lj1 ,
we only need to consider terms involving lj1 . Third, by our design, all terms in
log P (z, l) involving the label lj1 are linear functions of other labels li:i
1
=j and li .
2

Thus, the optimal Fj (·) is similar to the conditional PMF of label lj given (i) its
1 1

neighboring labels in both images and (ii) data zj1 , but a crucial difference is that
the contributions of the neighboring labels are through their expectations under
the corresponding factors, instead of their values themselves. This variational
inference scheme is also called mean-field approximation [15].
Within the n-th EM iteration, we propose the following algorithm for varia-
tional optimization of the factors:
1. Input: Observed data z and the current parameter estimates θn .
2. Initialize the expectations of all labels {li1 , li2 }Ii=1 . This can be taken to be
the (i) maximum-a-posteriori estimate given θn and z or (ii) empirically-
computed expectation using Gibbs sampling of the labels from the posterior
PMF. We have found the former strategy to be effective as well as computa-
tionally efficient.
3. Iterate over all pixels (order column by column) in both the label images. At
each pixel, perform the next step.
4. Without loss of generality, consider pixel j in the first image. Analytically
evaluate the optimal PMF Fj1 (·), given the expected values of the interacting
labels and data; as described previously in this section. Analytically evaluate
the expectation of lj1 over this optimal PMF Fj1 (·) and update it.
5. Repeat the last two steps, until convergence of the expected label images and
thereby the factorized approximation F (L|z, θn ).
6. Output: The factorized posterior F (L|z, θn ) in terms of the factors
{Fi1 (·), Fi2 (·)}Ii=1 .
After obtaining the factorized posterior F (L|z, θn ), we evaluate the Q(θ; θn )
function efficiently as follows.
1. Input: Observed data z, the current parameter estimates θn , and the optimal
factorized posterior F (L|z, θn ).
2. Sample a large number S of label image pairs {ls := (ls1 , ls2 )}Ss=1 from
F (L|z, θn ) by independent sampling at each pixel from the label distribu-
tion. This is straightforward (unlike Gibbs sampling) and computationally
cheap because sampling at each pixel is sampling from a binomial (factor)
distribution. Thus, we can easily generate a large sample size to ensure a
high-quality approximation.
3. Closely approximate the Q(θ; θn ) function as follows:
S
 θn ) := 1
Q(θ; θn ) ≈ Q(θ; log P (z, ls |θ) (11)
S s=1

 θn ).
4. Output: The function Q(θ;
 θn ).
The next section describes the M step for optimizing parameters θ using Q(θ;
 Colocalization Estimation Using Graphical Modeling and VBEM 11

3.4 M Step: Parameter Updates

The M step in the n-th EM iteration updates the parameter estimates θ as
S

 θn ) = arg max
θn+1 := arg max Q(θ; log P (z, ls |θ). (12)
θ θ
s=1

We perform alternating minimization over the set of parameter estimates as

follows.
The optimal likelihood parameters θL maximize the data likelihood given
labels:
S
 S

arg max log P (z, ls |θ) = arg max log P (z|ls , θL ). (13)
θL θL
s=1 s=1

These updates, for the means and variances of the object-specific and
background-specific Gaussians, are obtained in closed form.
The optimal prior parameters θP maximize the label prior probability:
S
 S

arg max log P (z, ls |θ) = arg max log P (ls |θP ). (14)
θP θP
s=1 s=1

The optimization of the prior parameters θP is more difficult because they appear
as part of the intractable partition function Z(θP ). Nevertheless, we have found
an effective strategy for estimating θP , which relies on the principles underly-
ing maximum pseudo-likelihood (MPL) estimation [15] that obtains parameter
estimates by maximizing a pseudo-likelihood function instead of the true likeli-
hood function (Note: in the context of MPL, the “likelihood” is the probability
of observing a label image given the MRF model and underlying parameters).
The key idea is that the pseudo-likelihood function (i) is simpler to optimize
because it eliminates the partition function and, (ii) for sufficiently large data,
mimics the true likelihood function well.
The classic MPL strategy for estimation of MRF parameters relies on a single
observed (label) image, where the MPL estimator is known to be consistent [15].
In our case, we have a set of sampled label images {ls }Ss=1 . This image set can be
considered similar to data obtained by cutting out images from a much larger
image. In a statistical sense, observing one large image drawn from a MRF
model is equivalent to observing many smaller images drawn from the same
model (i.e., with same potential functions and parameters). Thus, we propose to
apply the MPL strategy to the entire set of images by replacing each P (ls |θP )
by its pseudo-likelihood analog to optimize θP as:
S

max log P (ls |θP ) ≈
θP
s=1
S
 I I


max log P (lis1 |lN
s1 s2
self , lN other , θP ) P (lis2 |lN
s2 s1
self , lN other , θP ) =
θP i i i i
s=1 i=1 i=1
Another random document with
no related content on Scribd:
The Project Gutenberg eBook of History of the
United States of America, Volume 2 (of 9)
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.

Title: History of the United States of America, Volume 2 (of 9)

During the first administration of Thomas Jefferson

Author: Henry Adams

Release date: December 24, 2023 [eBook #72499]

Language: English

Original publication: New York: Charles Scribner's Sons, 1889

Credits: Richard Hulse, Karin Spence and the Online Distributed

Proofreading Team at https://www.pgdp.net (This file was
produced from images generously made available by The
Internet Archive/American Libraries.)

*** START OF THE PROJECT GUTENBERG EBOOK HISTORY OF

THE UNITED STATES OF AMERICA, VOLUME 2 (OF 9) ***
THE FIRST ADMINISTRATION
OF

THOMAS JEFFERSON
1801–1805
 HISTORY
OF THE

UNITED STATES OF AMERICA

DURING THE FIRST ADMINISTRATION OF

THOMAS JEFFERSON

By HENRY ADAMS

Vol. II.

NEW YORK
CHARLES SCRIBNER’S SONS
1889
 Copyright, 1889,
By Charles Scribner’s Sons.

University Press:
John Wilson and Son, Cambridge.
 CONTENTS OF VOL. II.
CHAPTER PAGE
I. Rupture of the Peace of Amiens 1
II. The Louisiana Treaty 25
III. Claim to West Florida 51
IV. Constitutional Difficulties 74
V. The Louisiana Debate 94
VI. Louisiana Legislation 116
VII. Impeachments 135
VIII. Conspiracy 160
IX. The Yazoo Claims 192
X. Trial of Justice Chase 218
XI. Quarrel with Yrujo 245
XII. Pinckney’s Diplomacy 264
XIII. Monroe and Talleyrand 288
XIV. Relations with England 316
XV. Cordiality with England 342
XVI. Anthony Merry 360
XVII. Jefferson’s Enemies 389
XVIII. England and Tripoli 410

Index to Vols. I. and II. 439

 THE COAST OF
WEST FLORIDA
AND
LOUISIANA
(From Jeffery’s American Atlas. London, 1800.)
HISTORY OF THE UNITED STATES.
 CHAPTER I.
Congress expired; Monroe set sail March 8, 1803; Washington
relapsed into silence; and the President and his Cabinet waited
alone in the empty village, triumphing for the moment over their
difficulties. Although a French prefect was actually in New Orleans,
and the delivery of Louisiana to Bonaparte might from day to day be
expected, not an additional soldier stood on the banks of the
Mississippi, and the States of Kentucky and Tennessee were as
quiet as though their flat-boats still floated down to New Orleans. A
month passed before Madison or Jefferson again moved. Then the
President asked his Cabinet[1] what Monroe should do in case
France, as he expressed it, “refused our rights.” He proposed an
alliance with England, and suggested three inducements which
might be offered to Great Britain: “1. Not to make a separate peace.
2. To let her take Louisiana. 3. Commercial privileges.” The Cabinet
unanimously rejected the second and third concessions, but
Dearborn and Lincoln were alone in opposing the first; and a majority
agreed to instruct Monroe and Livingston, “as soon as they find that
no arrangements can be made with France, to use all possible
procrastination with them, and in the mean time enter into
conferences with the British government, through their ambassador
at Paris, to fix principles of alliance, and leave us in peace till
Congress meets; and prevent war till next spring.”
Madison wrote the instructions. If the French government, he
said,[2] should meditate hostilities against the United States, or force
a war by closing the Mississippi, the two envoys were to invite
England to an alliance, and were to negotiate a treaty stipulating that
neither party should make peace or truce without consent of the
other. Should France deny the right of deposit without disputing the
navigation, the envoys were to make no positive engagement, but
should let Congress decide between immediate war or further
procrastination.
 At no time in Talleyrand’s negotiations had the idea of war
against the United States been suggested. Of his intentions in this
respect alone he had given positive assurances.[3] Above all things
both he and the First Consul feared a war with the United States.
They had nothing to gain by it. Madison’s instructions therefore
rested on an idea which had no foundation, and which in face of the
latest news from Europe was not worth considering; yet even if
intended only for use at home, the instructions were startling enough
to warrant Virginians in doubting their authenticity. The late
Administration, British in feeling as it was supposed to be, had never
thought an alliance with England necessary even during actual
hostilities with France, and had not hesitated to risk the chances of
independent action. Had either of Jefferson’s predecessors
instructed American ministers abroad, in case of war with France, to
bind the United States to make no peace without England’s consent,
the consequence would have been an impeachment of the
President, or direct steps by Virginia, Kentucky, and North Carolina,
as in 1798, tending to a dissolution of the Union. Such an alliance,
offensive and defensive, with England contradicted every principle
established by President Washington in power or professed by
Jefferson in opposition. If it was not finesse, it was an act such as
the Republicans of 1798 would have charged as a crime.
While Madison was writing these instructions, he was interrupted
by the Marquis of Casa Yrujo,[4] who came in triumph to say that his
Government had sent out a brigantine especially to tell the President
that the right of deposit would be restored and continued till another
agreement or equivalent place could be fixed upon.[5] Yrujo was
instructed to thank the President for his friendly, prudent, and
moderate conduct during the excitement. He sent to New Orleans
the positive order of King Charles IV. to the Intendant Morales, that
the right of deposit should be immediately restored; the western
people were told that their produce might go down the river as
before, and thus the last vestige of anxiety was removed. In face of
this action by Godoy, and of the war evidently at hand between
France and England, the success of the peace policy was assured.
These events in some degree explained the extraordinary nature of
the new instructions of April, 1803.
Monroe was then already at Paris. In order to make clear the
situation in which he found himself, the sequence of events in
Europe needs to be understood.
Bonaparte’s expedition to Louisiana was to have sailed at the
end of September, 1802.[6] A general of division, three generals of
brigade, five battalions of infantry, two companies of artillery, sixteen
pieces of cannon, and three thousand muskets were to be collected
at Dunkirk for shipment; but as fast as regiments could be named
they were consumed by the fiery furnace of St. Domingo.
Nevertheless, all the orders and arrangements were gradually made.
Victor was to command the forces in Louisiana; Laussat was to be
prefect, charged with the civil administration. Both received elaborate
written instructions; and although Victor could not sail without ships
or troops, Laussat was sent on his way.
These instructions, which were never published, had extreme
value for the decision of disputes which were to perturb American
politics for the next twenty years. Although Victor was forced to wait
in Holland for the expedition he commanded, a copy of his
instructions was given to Laussat, and served to regulate his conduct
as long as he remained in office. Decrès, the Minister of Marine, was
the author of this paper, which unfolded the purpose that had guided
France in recovering, and was to control her in administering, this
vast possession. Nothing could be simpler, clearer, or more
consistent with French policy than this document, which embodied
so large a part of Talleyrand’s political system.
The instructions began, as was natural, by a careful definition of
the new province. After reciting the terms of the retrocession
according to the Third Article of Berthier’s Treaty, Decrès fixed the
boundaries of the territory which Victor, on the part of the French
republic, was to receive from the Marquis of Somoruelos, the
Captain-General of Cuba.[7]
“The extent of Louisiana,” he said, “is well determined on the south
by the Gulf of Mexico. But bounded on the west by the river called Rio
 Bravo from its mouth to about the 30° parallel, the line of demarcation
stops after reaching this point, and there seems never to have been
any agreement in regard to this part of the frontier. The farther we go
northward, the more undecided is the boundary. This part of America
contains little more than uninhabited forests or Indian tribes, and the
necessity of fixing a boundary has never yet been felt there. There
also exists none between Louisiana and Canada.”
In this state of things the captain-general would have to relieve
the most remote Spanish garrisons, in order to establish possession;
in other respects he would be guided only by political and military
interests. The western and northern boundary was of less
consequence than the little strip which separated New Orleans from
Mobile; and to this point the instructions specially called Victor’s
attention. Quoting the treaty of 1763 between Spain, Great Britain,
and France, when Florida was to become a British possession,
Decrès fixed its terms as still binding upon all the interested parties.
“‘It is agreed,’” said the seventh article of this treaty, “‘that in future
the boundaries between the States of his Most Christian Majesty and
those of his Britannic Majesty shall be irrevocably fixed by a line
drawn down the middle of the Mississippi River from its source to the
River Iberville, and from there by a line down the middle of that river
and of the lakes Maurepas and Pontchartrain to the sea. New Orleans
and the island on which it stands shall belong to France.’ Such is still
to-day the eastern limit of Louisiana. All to the east and north of this
limit makes part of the United States or of West Florida.”
Nothing could be clearer. Louisiana stretched from the Iberville to
the Rio Bravo; West Florida from the Iberville to the Appalachicola.
The retrocession of Louisiana by Spain to France could restore only
what France had ceded to Spain in 1762. West Florida had nothing
to do with the cession of 1762 or the retrocession of 1800, and being
Spanish by a wholly different title could not even be brought in
question by the First Consul, much as he wanted Baton Rouge,
Mobile, and Pensacola. Victor’s orders were emphatic:—
“There is therefore no obscurity as to our boundary on this side
any more than as to that of our allies; and although Florida belongs to
Spain, Spain’s right of property in this quarter will have as much
interest for the Captain-General of Louisiana as though Florida were a
French possession.”
 After thus establishing the boundary, as far as possible, in every
direction, the minister treated at some length of the English claim to
navigation on the Mississippi, and at last reached the general subject
of the relation between Louisiana and the world about it,—the
subject in which Jefferson would have found acute interest:—
“The system of this, as of all our other colonies, should be to
concentrate its commerce in the national commerce; it should have in
particular the aim of establishing its relations with our Antilles, so as to
take the place, in these colonies, of the American commerce for all the
objects whose import and export is permitted to them. The captain-
general should especially abstain from every innovation favorable to
strangers, who should be restricted to such communications as are
absolutely indispensable to the prosperity of Louisiana and to such as
are explicitly determined by the treaties.”
Commercial relations with the Spanish colonies were to be
encouraged and extended as much as possible, while the utmost
caution was to be observed toward the United States:—
“From what has been said of Louisiana and the adjacent States, it
is clear that the republic of France, being master of both banks at the
mouth of the Mississippi, holds the key to its navigation. This
navigation is nevertheless a matter of the highest importance for the
western States of the Federal Government.... This is enough to show
with what jealousy the Federal Government will see us take
possession of Louisiana. Whatever may be the events which this new
part of the continent has to expect, the arrival of the French forces
should be marked there by the expression of sentiments of great
benevolence for these new neighbors.”
Expression of benevolent sentiments was a pleasing duty; but it
was not to interfere with practical measures, both defensive and
offensive:—
“The greatest circumspection will be required in directing the
colonial administration. A little local experience will soon enable you to
discern the sentiments of the western provinces of the Federal
Government. It will be well to maintain sources of intelligence in that
country, whose numerous, warlike, and sober population may present
you a redoubtable enemy. The inhabitants of Kentucky especially
should fix the attention of the captain-general.... He must also fortify
himself against them by alliance with the Indian nations scattered to
the east of the river. The Chibackas, Choctaws, Alabamas, Creeks,
 etc., are represented as being entirely devoted to us.... He will not
forget that the French government wishes peace; but that if war takes
place, Louisiana will certainly become the theatre of hostilities.... The
intention of the First Consul is to raise Louisiana to a degree of
strength which will allow him in time of war to abandon it to its own
resources without anxiety; so that enemies may be forced to the
greatest sacrifices merely in attempting to attack it.”
In these instructions not a word could be found which clashed
with Jefferson’s pacific views; and partly for that reason they were
more dangerous to the United States than if they had ordered Victor
to seize American property on the Mississippi and occupy Natchez
with his three thousand men. Victor was instructed, in effect, to
tamper with every adventurer from Pittsburg to Natchez; buy up
every Indian tribe in the Georgia and Northwestern Territory; fortify
every bluff on the western bank from St. Louis to New Orleans; and
in a few years create a series of French settlements which would
realize Madison’s “sound policy” of discouraging the United States
from colonizing the west bank.
Fortified by these instructions, the Citizen Laussat set sail Jan.
12, 1803, and in due time arrived at New Orleans. Victor labored in
Holland to put his ships and supplies in a condition to follow. As
Laussat sailed, another step was taken by the French government.
General Bernadotte, a very distinguished republican officer, brother-
in-law of Joseph Bonaparte, was appointed minister at Washington.
[8] The First Consul had his own reasons for wishing to remove
Bernadotte, as he meant to remove Moreau; and Washington was a
place of indirect banishment for a kinsman whose character was to
be feared. Bernadotte’s instructions[9] were signed by Talleyrand
Jan. 14, 1803, the day after Monroe was confirmed as special envoy
to France by the Senate at Washington, and while Laussat was still
on the French coast. Although Bonaparte had been obliged to
withdraw a part of Victor’s force, he still intended that the expedition
should start at once with two thousand men;[10] and its departure
was to be so timed that Bernadotte should reach Washington as
Victor and his troops reached New Orleans. Their instructions were
on one point identical. News of the closure of the Mississippi by
Morales had reached Paris, and had already caused an official
protest by Livingston, when Talleyrand drew up the instructions to
Bernadotte:—
“Louisiana being soon to pass into our hands, with all the rights
which have belonged to Spain, we can only with pleasure see that a
special circumstance has obliged the Spanish Administration to
declare formally [constater] its right to grant or to refuse at will to the
Americans the privilege of a commercial entrepôt at New Orleans; the
difficulty of maintaining this position will be less for us than that of
establishing it.... Yet in any discussion that may arise on this subject,
and in every discussion you may have to sustain, the First Consul
wishes you to be informed of his most positive and pronounced desire
to live in good understanding with the American government, to
cultivate and to improve for the advantage of American commerce the
relations of friendship which unite the two peoples. No one in Europe
wishes the prosperity of that people more than he. In accrediting you
to its Government he has given it a peculiar mark of his good
disposition; he doubts not that you will make every effort to bind closer
the ties which exist between the two nations. In consequence of the
firm intention which the First Consul has shown on this subject, I must
recommend you to take every care to avoid whatever might alter our
relations with that nation and its Government. The agents of the
French republic in the United States should forbid themselves
whatever might even remotely lead to a rupture. In ordinary
communication, every step should show the benevolent disposition
and mutual friendship which animate the chiefs and all the members
of the two Governments; and when any unforeseen difficulty rises
which may in any degree whatever compromise their good
understanding, the simplest and most effectual means of preventing
all danger is to refer its solution to the inquiry and direct judgment of
the two Governments.”
Talleyrand’s language was more elaborate, but not clearer, than
that which Bonaparte himself used to Victor.[11]
“I have no need to tell you,” the First Consul wrote, “with what
impatience the Government will wait for news from you in order to
settle its ideas in regard to the pretensions of the United States and
their usurpations over the Spaniards. What the Government may think
proper to do must not be judged in advance until you have rendered
an account of the state of things. Every time you perceive that the
United States are raising pretensions, answer that no one has an idea
 of this at Paris (que l’on n’a aucune idée de cela à Paris); but that you
have written, and that you are expecting orders.”
These were the ideas held by the government of France at the
moment when Jefferson nominated Monroe as a special envoy to
buy New Orleans and West Florida. Jefferson’s hopes of his success
were small; and Livingston, although on the spot and eager to try the
experiment, could only write:[12] “Do not absolutely despair.”
Whatever chance existed of obtaining New Orleans seemed to lie in
the possibility that Addington’s peaceful administration in England
might be driven into some act contrary to its vital interests; and even
this chance was worth little, for so long as Bonaparte wanted peace,
he could always keep it. England was thoroughly weary of war; and
proved it by patiently looking on while Bonaparte, during the year,
committed one arbitrary act after another, which at any previous time
would have been followed by an instant withdrawal of the British
minister from Paris.
On the other hand, the world could see that Bonaparte was
already tired of peace; his rôle of beneficent shopkeeper disgusted
him, and a new war in Europe was only a question of months. In
such a case the blow might fall on the east bank of the Rhine, on
Spain, or on England. Yet Bonaparte was in any case bound to keep
Louisiana, or return it to Spain. Florida was not his to sell. The
chance that Jefferson could buy either of these countries, even in
case of a European war, seemed so small as hardly to be worth
considering; but it existed, because Bonaparte was not a man like
other men, and his action could never be calculated in advance.
The news that Leclerc was dead, that his army was annihilated,
St. Domingo ruined, and the negroes more than ever beyond control,
reached Paris and was printed in the “Moniteur” Jan. 7, 1803, in the
same active week when Bernadette, Laussat, and Victor were
ordered from France to America, and Monroe was ordered from
America to France. Of all the events of the time, Leclerc’s death was
the most decisive. The colonial system of France centred in St.
Domingo. Without that island the system had hands, feet, and even
a head, but no body. Of what use was Louisiana, when France had
clearly lost the main colony which Louisiana was meant to feed and
fortify? The new ruler of France was not unused to failure. More than
once he had suddenly given up his dearest plans and deserted his
oldest companions when their success was hopeless. He had
abandoned Paoli and Corsica with as little compunction as afterward
he abandoned the army and the officers whom he led to Egypt.
Obstinate in pursuing any object which led to his own advancement,
he was quick to see the moment when pursuit became useless; and
the difficulties that rose in his path toward colonial empire were quite
as great as those which had driven him to abandon Corsica and
Egypt. Not only had the island of St. Domingo been ruined by the
war, its plantations destroyed, its labor paralyzed, and its population
reduced to barbarism, so that the task of restoring its commercial
value had become extremely difficult; but other and greater
objections existed to a renewal of the struggle. The army dreaded
service in St. Domingo, where certain death awaited every soldier;
the expense was frightful; a year of war had consumed fifty thousand
men and money in vast amounts, with no other result than to prove
that at least as many men and as much money would be still needed
before any return could be expected for so lavish an expenditure. In
Europe war could be made to support war; in St. Domingo peace
alone could but slowly repair some part of this frightful waste.
Leclerc was succeeded at St. Domingo by General Rochambeau,
a son of the Comte de Rochambeau, who twenty years before had
commanded the French corps which enabled Washington to capture
Cornwallis at Yorktown. A brave officer, but known to be little fit for
administration, Rochambeau was incompetent for the task that fell
on him. Leclerc had warned the Government that in case of his own
retirement he had no officer fit to replace him,—least of all
Rochambeau, who was next in rank. Rochambeau wrote to inform
the First Consul that thirty-five thousand men must be sent to save
the island.[13] Without a new commander-in-chief of the highest
ability, a new army was useless; and meanwhile Rochambeau was
certain to waste the few thousand acclimated soldiers who should
form its nucleus.
The First Consul found himself in a difficult and even dangerous
situation. Probably the colonial scheme had never suited his tastes,