Download textbook Medical Computer Vision Algorithms For Big Data International Workshop Mcv 2014 Held In Conjunction With Miccai 2014 Cambridge Ma Usa September 18 2014 Revised Selected Papers 1St Edition Bjoern Menze ebook all chapter pdf

Medical Computer Vision Algorithms
for Big Data International Workshop

MCV 2014 Held in Conjunction with
MICCAI 2014 Cambridge MA USA
September 18 2014 Revised Selected
Papers 1st Edition Bjoern Menze
Visit to download the full and correct content document:
https://textbookfull.com/product/medical-computer-vision-algorithms-for-big-data-inter
national-workshop-mcv-2014-held-in-conjunction-with-miccai-2014-cambridge-ma-us
a-september-18-2014-revised-selected-papers-1st-edition-bjoern-menze/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Spatio temporal Image Analysis for Longitudinal and

Time Series Image Data Third International Workshop
STIA 2014 Held in Conjunction with MICCAI 2014 Boston
MA USA September 18 2014 Revised Selected Papers 1st
Edition Stanley Durrleman
https://textbookfull.com/product/spatio-temporal-image-analysis-
for-longitudinal-and-time-series-image-data-third-international-
workshop-stia-2014-held-in-conjunction-with-miccai-2014-boston-
ma-usa-september-18-2014-revised-selected/
Medical Computer Vision Large Data in Medical Imaging

Third International MICCAI Workshop MCV 2013 Nagoya
Japan September 26 2013 Revised Selected Papers 1st
Edition Bjoern Menze
https://textbookfull.com/product/medical-computer-vision-large-
data-in-medical-imaging-third-international-miccai-workshop-
mcv-2013-nagoya-japan-september-26-2013-revised-selected-
papers-1st-edition-bjoern-menze/
Clinical Image Based Procedures Translational Research

in Medical Imaging Third International Workshop CLIP
2014 Held in Conjunction with MICCAI 2014 Boston MA USA
September 14 2014 Revised Selected Papers 1st Edition
Marius George Linguraru
https://textbookfull.com/product/clinical-image-based-procedures-
translational-research-in-medical-imaging-third-international-
workshop-clip-2014-held-in-conjunction-with-miccai-2014-boston-
ma-usa-september-14-2014-revised-selected-p/
Statistical Atlases and Computational Models of the

Heart Imaging and Modelling Challenges 5th
International Workshop STACOM 2014 Held in Conjunction
with MICCAI 2014 Boston MA USA September 18 2014
Revised Selected Papers 1st Edition Oscar Camara
https://textbookfull.com/product/statistical-atlases-and-
computational-models-of-the-heart-imaging-and-modelling-
challenges-5th-international-workshop-stacom-2014-held-in-
Augmented Environments for Computer Assisted
Interventions 9th International Workshop AE CAI 2014
Held in Conjunction with MICCAI 2014 Boston MA USA
September 14 2014 Proceedings 1st Edition Cristian A.
Linte
https://textbookfull.com/product/augmented-environments-for-
computer-assisted-interventions-9th-international-workshop-ae-
cai-2014-held-in-conjunction-with-miccai-2014-boston-ma-usa-
september-14-2014-proceedings-1st-edition-cristian/
Bayesian and grAphical Models for Biomedical Imaging

First International Workshop BAMBI 2014 Cambridge MA
USA September 18 2014 Revised Selected Papers 1st
Edition M. Jorge Cardoso
https://textbookfull.com/product/bayesian-and-graphical-models-
for-biomedical-imaging-first-international-workshop-
bambi-2014-cambridge-ma-usa-september-18-2014-revised-selected-
papers-1st-edition-m-jorge-cardoso/
Medical Image Computing and Computer Assisted

Intervention MICCAI 2014 17th International Conference
Boston MA USA September 14 18 2014 Proceedings Part II
1st Edition Polina Golland
https://textbookfull.com/product/medical-image-computing-and-
computer-assisted-intervention-miccai-2014-17th-international-
conference-boston-ma-usa-september-14-18-2014-proceedings-part-
ii-1st-edition-polina-golland/
Multimodal Analyses enabling Artificial Agents in Human

Machine Interaction Second International Workshop
MA3HMI 2014 Held in Conjunction with INTERSPEECH 2014
Singapore Singapore September 14 2014 Revised Selected
Papers 1st Edition Ronald Böck
https://textbookfull.com/product/multimodal-analyses-enabling-
artificial-agents-in-human-machine-interaction-second-
international-workshop-ma3hmi-2014-held-in-conjunction-with-
interspeech-2014-singapore-singapore-september-14-2014-re/
Combinatorial Algorithms 25th International Workshop

IWOCA 2014 Duluth MN USA October 15 17 2014 Revised
Selected Papers 1st Edition Kratochvíl Jan
https://textbookfull.com/product/combinatorial-algorithms-25th-
international-workshop-iwoca-2014-duluth-mn-usa-
october-15-17-2014-revised-selected-papers-1st-edition-
Bjoern Menze · Georg Langs
Albert Montillo · Michael Kelm
Henning Müller · Shaoting Zhang
Weidong (Tom) Cai · Dimitris Metaxas (Eds.)
Medical Computer
LNCS 8848
Vision: Algorithms
for Big Data
International Workshop, MCV 2014
Held in Conjunction with MICCAI 2014
Cambridge, MA, USA, September 18, 2014
Revised Selected Papers
123
Lecture Notes in Computer Science 8848
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, Saarbruecken, Germany
More information about this series at http://www.springer.com/series/7412
Bjoern Menze Georg Langs
•
Albert Montillo Michael Kelm

•
Henning Müller Shaoting Zhang

•
Weidong (Tom) Cai Dimitris Metaxas (Eds.)

•
Medical Computer Vision:

Algorithms for Big Data
International Workshop, MCV 2014
Held in Conjunction with MICCAI 2014
Cambridge, MA, USA, September 18, 2014
Revised Selected Papers
123
Editors
Bjoern Menze Henning Müller
Technische Universität München University of Applied Sciences
Munich Sierre
Germany Switzerland
Georg Langs Shaoting Zhang
Medical University of Vienna University of North Carolina
Vienna Charlotte
Austria USA
Albert Montillo Weidong (Tom) Cai
GE Global Research School of Information Technologies
Niskayuna, NY Biomedical and Multimedia Information
USA Technology (BMIT) Research Group,
University of Sydney
Michael Kelm Sydney, NSW
Siemens AG Australia
Erlangen
Germany Dimitris Metaxas
Rutgers University
Piscataway, NJ
USA
ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-319-13971-5 ISBN 978-3-319-13972-2 (eBook)
DOI 10.1007/978-3-319-13972-2
Library of Congress Control Number: 2014958150
Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2014
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 MICCAI 2014 Workshop on Medical Computer Vision: Algorithms for Big Data
(MICCAI-bigMCV 2014) was held in conjunction with the 17th International Con-
ference on Medical Image Computing and Computer Assisted Intervention (MICCAI
2014) on September 18, 2014 in Boston, USA. It succeeds the Workshops on Medical
Computer Vision that were held in September 2010 in conjunction with MICCAI 2010
in Beijing, in June 2012 in conjunction with CVPR 2012 in Providence, in October
2012 in conjunction with MICCAI 2012 in Nice, and in September 2013 in conjunction
with MICCAI 2013 in Nagoya.
With the ever-increasing amount of annotated multimodal medical data consisting of
imaging and textual information, large-scale, data-driven methods provide the promise
of bridging the semantic gap between images and diagnoses. The one-day workshop
aimed at exploring the use of modern computer vision technology and “big data”
algorithms in tasks such as automatic segmentation and registration, localization of
anatomical features, and detection of anomalies. We emphasized questions of har-
vesting, organizing, and learning from large-scale medical imaging datasets and gen-
eral-purpose automatic understanding of medical images. We were especially interested
in modern, scalable, and efficient algorithms that generalize well to previously unseen
images and can be applied to large-scale datasets that are arising, for example, from
studies with significant populations, through the use of wide-field-of-view imaging
sequences at high-spatial resolution, or when compiling hospital-scale databases. In
addition, the bigMCV 2014 hosted a VISCERAL session for presentation and dis-
cussion of methods for anatomical structure segmentation and localization. The session
highlights the results of the VISCERALanatomy2 challenge, and provided a forum to
discuss the individual approaches and their comparative evaluation.
Our call for papers resulted in 29 submissions of up to 12 pages. Each paper
received two to four reviews. Based on these peer reviews, we accepted 13 papers that
were regular oral presentations, and 5 papers for the VISCERAL session.
October 2014 Bjoern Menze

Georg Langs
Albert Montillo
Michael Kelm
Henning Müller
Shaoting Zhang
Weidong (Tom) Cai
Dimitris Metaxas
Organization
Organization Committee
General Co-chairs
Bjoern Menze Technische Universität München, Munich,
Germany
Georg Langs Medical University of Vienna, Vienna, Austria
and MIT, Cambridge, MA, USA
Albert Montillo GE Global Research, USA
Michael Kelm Siemens, Erlangen, Germany
Henning Müller University of Applied Sciences Western
Switzerland (HES-SO) and University Hospitals
and University of Geneva, Switzerland
Shaoting Zhang UNC Charlotte, USA
Weidong (Tom) Cai University of Sydney, Australia and Harvard
Medical School, Boston, MA, USA
Dimitris Metaxas Rutgers University, USA
Publication Chair
Henning Müller Switzerland
International Program Committee

Adrien Depeursinge Stanford University, USA
Allan Hanbury Technical University of Vienna, Austria
Allison Noble University of Oxford, UK
Alvina Goh National University, Singapore
Christian Wachinger MIT, USA
Daniel Rueckert Imperial College London, UK
Darko Zikic Microsoft Research Cambridge, UK
Diana Mateus Technische Universität München, Munich,
Germany
Dinggang Shen UNC Chapel Hill, USA
Ender Konukoglu Harvard Medical School, USA
Ertan Cetingul Siemens Corporate Research, USA
Herve Lombaert Inria, France
Jan Margeta Inria, France
Jurgen Gall University of Bonn, Germany
Kilian Pohl Stanford University, USA
VIII Organization
Koen Van Leemput Technical University of Denmark, Denmark

Le Lu National Institutes of Health, USA
Marius Erdt Fraunhofer/NTU, Germany
Marleen de Bruijne Erasmus MC, The Netherlands
Matthias Schneider ETH Zurich, Switzerland
Matthew Toews Harvard Medical School, USA
Maxime Sermesant Inria, France
Michael Wels Siemens Healthcare, Germany
Milan Sonka University of Iowa, USA
Paul Suetens KU Leuven, Belgium
Philippe Cattin University of Basel, Switzerland
Pingkun Yan Chinese Academy of Sciences, China
Ron Kikinis Harvard Medical School, USA
Sebastian Ourselin University College London, UK
Stefan Wesarg Fraunhofer Gesellschaft, Germany
Stefan Bauer University of Bern, Switzerland
Tammy Riklin-Raviv Ben-Gurion University of the Negev, Israel
Tobias Gass ETH Zurich, Switzerland
Tom Vercauteren Mauna Kea Technologies, France
Xavier Pennec Inria, France
Xinghua Lou Microsoft Research, UK
Yang Song University of Sydney, Australia
Yefeng Zheng Siemens Corporate Research, USA
Sponsors
European Commission 7th Framework Programme, VISCERAL (318068) and

Khresmoi (257528) projects.
Contents
Workshop Overview
Overview of the 2014 Workshop on Medical Computer Vision—Algorithms

for Big Data (MCV 2014) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Henning Müller, Bjoern Menze, Georg Langs, Albert Montillo,
Michael Kelm, Shaoting Zhang, Weidong (Tom) Cai, and Dimitris Metaxas
Segmentation of Big Medical Data
Joint Segmentation and Registration for Infant Brain Images . . . . . . . . . . . . 13

Guorong Wu, Li Wang, John Gilmore, Weili Lin, and Dinggang Shen
LINKS: Learning-Based Multi-source IntegratioN FrameworK for Segmentation

of Infant Brain Images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Li Wang, Yaozong Gao, Feng Shi, Gang Li, John H. Gilmore, Weili Lin,
and Dinggang Shen
Pectoralis Muscle Segmentation on CT Images Based on Bayesian Graph Cuts

with a Subject-Tailored Atlas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Rola Harmouche, James C. Ross, George R. Washko,
and Raúl San José Estépar
Advanced Feature Extraction
Learning Features for Tissue Classification with the Classification

Restricted Boltzmann Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Gijs van Tulder and Marleen de Bruijne
Dementia-Related Features in Longitudinal MRI: Tracking Keypoints

over Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Elisabeth Stühler and Michael R. Berthold
Object Classification in an Ultrasound Video Using LP-SIFT Features. . . . . . 71

Mohammad Ali Maraci, Raffaele Napolitano, Aris Papageorghiou,
and J. Allison Noble
Unsupervised Pre-training Across Image Domains Improves Lung

Tissue Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Thomas Schlegl, Joachim Ofner, and Georg Langs
X Contents
Multi-atlas and Beyond
Atlas-Guided Multi-channel Forest Learning for Human Brain Labeling. . . . . 97

Guangkai Ma, Yaozong Gao, Guorong Wu, Ligang Wu, and Dinggang Shen
Fast Multiatlas Selection Using Composition of Transformations

for Radiation Therapy Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
David Rivest-Hénault, Soumya Ghose, Josien P.W. Pluim, Peter B. Greer,
Jurgen Fripp, and Jason A. Dowling
Classifier-Based Multi-atlas Label Propagation with Test-Specific Atlas

Weighting for Correspondence-Free Scenarios . . . . . . . . . . . . . . . . . . . . . . 116
Darko Zikic, Ben Glocker, and Antonio Criminisi
Translational Medical Computer Vision
CT Prostate Deformable Segmentation by Boundary Regression . . . . . . . . . . 127

Yeqin Shao, Yaozong Gao, Xin Yang, and Dinggang Shen
Precise Lumen Segmentation in Coronary Computed

Tomography Angiography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Felix Lugauer, Yefeng Zheng, Joachim Hornegger, and B. Michael Kelm
Confidence-Based Training for Clinical Data Uncertainty

in Image-Based Prediction of Cardiac Ablation Targets . . . . . . . . . . . . . . . . 148
Rocío Cabrera-Lozoya, Jan Margeta, Loïc Le Folgoc, Yuki Komatsu,
Benjamin Berte, Jatin Relan, Hubert Cochet, Michel Haïssaguerre,
Pierre Jaïs, Nicholas Ayache, and Maxime Sermesant
VISCERAL Session
Rule-Based Ventral Cavity Multi-organ Automatic Segmentation

in CT Scans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Assaf B. Spanier and Leo Joskowicz
Multi-atlas Segmentation and Landmark Localization in Images

with Large Field of View . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Tobias Gass, Gabor Szekely, and Orcun Goksel
Automatic Liver Segmentation Using Statistical Prior Models

and Free-form Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Xuhui Li, Cheng Huang, Fucang Jia, Zongmin Li, Chihua Fang,
and Yingfang Fan
Contents XI
Hierarchic Multi–atlas Based Segmentation for Anatomical Structures:

Evaluation in the VISCERAL Anatomy Benchmarks . . . . . . . . . . . . . . . . . . 189
Oscar Alfonso Jiménez del Toro and Henning Müller
Automatic 3D Multiorgan Segmentation via Clustering and Graph Cut

Using Spatial Relations and Hierarchically-Registered Atlases. . . . . . . . . . . . 201
Razmig Kéchichian, Sébastien Valette, Michaël Sdika, and Michel Desvignes
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

Workshop Overview
Overview of the 2014 Workshop on Medical
Computer Vision—Algorithms for Big Data
(MCV 2014)
Henning Müller1,2(B) , Bjoern Menze3,4 , Georg Langs5,6 ,

Albert Montillo7 , Michael Kelm8 , Shaoting Zhang9 , Weidong (Tom) Cai10,11 ,
and Dimitris Metaxas12
1
University of Applied Sciences Western Switzerland (HES–SO), Sierre, Switzerland
henning.mueller@hevs.ch
2
University Hospitals and University of Geneva, Geneva, Switzerland
3
Technical University of Munich, Munich, Germany
4
INRIA, Sophia Antipolis, France
5
Medical University of Vienna, Vienna, Austria
6
MIT, Cambridge, MA, USA
7
GE Global Research, Niskayuna, USA
8
Siemens Healthcare, Erlangen, Germany
9
UNC Charlotte, Charlotte, USA
10
University of Sydney, Sydney, NSW, Australia
11
Harvard Medical School, Boston, USA
12
Rutgers University, Piscataway, USA
Abstract. The 2014 workshop on medical computer vision (MCV):

algorithms for big data took place in Cambridge, MA, USA in connec-
tion with MICCAI (Medical Image Computing for Computer Assisted
Intervention). It is the fourth MICCAI MCV workshop after those held
in 2010, 2012 and 2013 with another edition held at CVPR 2012. This
workshop aims at exploring the use of modern computer vision technol-
ogy in tasks such as automatic segmentation and registration, localisation
of anatomical features and extraction of meaningful visual features. It
emphasises questions of harvesting, organising and learning from large-
scale medical imaging data sets and general-purpose automatic under-
standing of medical images. The workshop is especially interested in
modern, scalable and efficient algorithms which generalise well to pre-
viously unseen images.The strong participation in the workshop of over
80 persons shows the importance of and interest in Medical Computer
Vision. This overview article describes the papers presented in the work-
shop as either oral presentations or short presentations and posters. It
also describes the invited talks and the results of the VISCERAL session
in the workshop on the use of big data in medical imaging.
Keywords: Medical image analysis · Medical computer vision · Seg-

mentation · Detection

c Springer International Publishing Switzerland 2014
B. Menze et al. (Eds.): MCV 2014, LNCS 8848, pp. 3–10, 2014.
DOI: 10.1007/978-3-319-13972-2 1
4 H. Müller et al.
1 Introduction
The Medical Computer Vision workshop (MCV) took place in conjunction with
MICCAI (Medical Image Computing for Computer-Assisted Interventions) on
September 18, 2014 in Cambridge, MA in the USA. This fifth workshop on med-
ical computer vision was organised in connection with MICCAI after the work-
shops in 2010 [12], 2012 [10] and 2013 [11] and with CVPR in 2012. The workshop
received 29 paper submissions of which five were submitted to the VISCERAL
session. All papers were reviewed by at least three external reviewers of the sci-
entific committee of the workshop. Then, all borderline papers were reviewed
in addition by at least one member of the workshop organisers. The 13 best
papers were presented as oral presentations and authors had the possibility to
also present a poster on their techniques for discussions during the lunch break.
With the increasing importance of large datasets (and the addition of big
data in the workshop title) it was also decided to again add a session on an
evaluation campaign called VISCERAL1 (VISual Concept ExtRaction challenge
in RAdioLogy) in 2014. The VISCERAL project [5] is creating large amounts
of manually annotated 3D medical data, and is making them available to the
research community in four benchmark challenges. The first two benchmarks
were focussed on the automatic detection of organs in the body and include
annotations of over 20 organs and 50 landmarks in multiple modalities. The
third benchmark is on lesion detection and the fourth on the retrieval of similar
cases in very large data sets.
This text also gives an overview of the most important discussions that took
place during the medical computer vision workshop and the challenges that
were identified in the field. Participants gave very good feedback and all agreed
to again organize the workshop during future MICCAI conferences.
2 Papers Presented at the Workshop
The oral presentations were separated into four topic areas, papers on segmen-
tation, feature extraction, multi-atlas techniques and the last session on trans-
lational medical computer vision.
2.1 Segmentation of Big Medical Data
Wu et al. [21] addressed the problem of segmentation and registration of infant

brains from subjects at different ages. They estimated tissue probability maps
separately using only training at the respective age and used the probability
maps as a good initialization to guide the level set segmentation.
Then, Wang et al. [20] presented a random forest based approach for infant
brain image segmentation that fuses multi-contrast MRI and tissue probability
maps. Next, Harmouche et al. [3] proposed a method to segment the pectoralis
1
http://visceral.eu/.
Overview of the 2014 Workshop on Medical Computer Vision 5
muscle in CT. Their approach constructs a likelihood using a multivariate distri-

bution of pairwise registered similar training subjects while the posterior tissue
map probability is used to drive a graph cuts segmentation.
2.2 Advanced Feature Extraction
In the first paper in this section, van Tulder and de Bruijne [19] adapted a
convolutional classification restricted Boltzmann machine to learn features well
suited for discriminative feature learning and apply it for texture-based tissue
classification on two lung CT problems. Then, Stühler [17] argued that for large
scale longitudinal key point tracking in brain MRI of dementia studies, time-
consuming non-rigid registration could be avoided by employing local invariant
features that are independent of image scale and orientation. Maraci et al. [9]
then showed how they combined techniques from the computer vision and med-
ical imaging communities to increase the degree of automation in ultrasound
acquisition. They introduced new symmetric SIFT features and used them to
represent the acquired image for classification of fetal image anatomical struc-
tures. Next Schlegl et al. [14] addressed the need to learn from data collected
across multiple hospitals with heterogeneous medical imaging equipment. Using
unsupervised pre-training of convolutional neural networks they inject informa-
tion from hospitals or image classes for which no annotations are available and
they show how this can lead to improved classification accuracy in the classifi-
cation of lung tissue.
2.3 Multi Atlas and Beyond
Ma et al. [8] kicked off our Multi-Atlas section by presenting a hybrid approach

for brain anatomy segmentation that combines multi-atlas and learning based
methods. Different from traditional learning-based labelling methods, their atlas-
guided multi-channel forest learning method utilized information from both the
target image and the aligned atlas for a voxel-wise labelling. Next the task
of reducing registration cost for radiation therapy planning was addressed by
Rivest-Henault et al. [13]. Their approach finds a proxy that can be used to hop
from a given image A to a target image B with minimal distortion and they also
defined both a clustering scheme and the transitivity error function. Last, Zikic
et al. [22] adapted the Atlas Forest approach for the case when target and test
brain images lack correspondences such as the case when there is a tumor in one.
By training on only atlases similar to the test, they managed to overcome the
inherent overtraining problem as shown in the results they presented on BraTS
2013 (Multimodal Brain Tumor Segmentation Challenge).
2.4 Translational Medical Computer Vision
The translation of concepts from computer vision to applications in the medical

imaging domain was well represented in our workshop. Shao et al. [15] let off
this section by describing a prostate boundary delineation method that forms an

estimate of voxel boundary likelihood using votes cast by a regression forest and
then form a discrete segmentation by fitting a deformable model. Next Lugauer
et al. [7] proposed a model-guided segmentation approach to segment the lumen
in coronary computed tomography angiography. Their method builds a Markov
Random Field model with convex priors to ensure tubular solutions, which they
optimize through a graph-cut based approach. Finally, a method for identifying
local image characteristics capable of predicting the presence of local abnor-
mal ventricular activities in the heart was proposed by Cabrera Lozoya et al. [1].
While determining the optimum intensity and texture-based local image features
using a random forest, they developed an approach for integrating uncertain-
ties due to errors in the training set and describe how this improves algorithm
performance.
3 Invited Speakers
3.1 Xiang Sean Zhou
The first invited speaker was Xiang Sean Zhou, Head of Innovations at Siemens
Medical Solutions in Malvern, PA, USA who presented his experiences in Med-
ical Imaging research from a large company perspective. He described main
lines of his research approach that has lead him to develop rapid and robust
anatomy localization. This accomplishment has in part lead him to recently be
awarded Inventor of the Year at Siemens. Emphasizing the principle of “robust-
ness through redundancy”, he argued that the three keys to achieve high robust-
ness in medical image analysis are “redundancy, redundancy, redundancy”,
(which reminded us of the saying that “the three key aspects in buying a house
are ’location, location, location”). The perspective on implementing or exploit-
ing redundancy was also inspired by the space and aeronautics industries that
are well known for pioneering work in fault tolerant design. Dr. Zhou described
multiple ways in which he has successfully designed redundancy into his solu-
tions. Redundancy through ensemble learning was an approach he adopted early
on which increases reliability through aggregating multiple machine learning
models, while redundancy through modality leverages multiple modalities when
providing a clinical interpretation such as was prevalent in the Health-e-Child
project. Redundancy through algorithm fusion entails using all of the best meth-
ods for image interpretation including detection, registration and segmentation.
These as well as several other redundancies have formed the hallmarks of his
research which lead to stimulating discussions about what academia might pro-
vide for industrial research. Dr. Zhou described how while many algorithms
might achieve success on 80–90 % of the cases, to find truly robust solutions
that can work on 98–99 % of the cases requires quantitative evaluation on large-
scale standardized datasets, such as those represented in the large challenges
now becoming popular at MICCAI or other events.
3.2 Eric G. Learned-Miller

The second invited speaker was Eric G. Learned-Miller, Associate Professor of
Computer Science at the University of Massachusetts in Amherst, MA, USA
who presented a talk entitled “Experience with Big Data: A Decade of Research
in Face Recognition”. In it he both presented an overview of computer vision
techniques for face recognition and provided a historical perspective. A cen-
tral theme in his research has been in the reduction of the face recognition
problem to its salient components. Towards this aim he has created a widely
used public resource for the community in the form of a curated face database,
Labeled Faces in the Wild. This has over 13,000 faces collected from the web,
each labeled with the name of the person pictured. Every face was identified
with the Viola-Jones face detector making detection less of a concern. Addition-
ally he provides subsets of the database in which the faces have been cropped,
scaled and aligned to a standard reference frame, leaving only the core recog-
nition task. As a nice complement to the discussion we had with Dr. Zhou,
Dr. Learned-Miller described how he sees a primary benefit coming from main-
taining the database online, so that researchers can at any time benchmark
their approaches to the state of the art. The system maintains a ranking of all
methods submitted. This has enabled Dr. Learned-Miller to uncover trends in
the approaches being applied. For example he has observed that, while at first
computer vision researchers submitted the best methods, the recent trend has
been that researchers from the machine learning community have attained top
scores. Additionally, multiple ranking strategies are employed. In one, methods
are allowed to train only from data in the database, while in another training
data can come from any additional source, including Facebook which several
methods employed successfully. This has enabled further insights including an
understanding of the relative value in training on more data versus the develop-
ment of new methodological models and approaches. This provided an excellent
tutorial example to complement the MICCAI debate session “Signal Processing
or Machine learning: What’s right for MICCAI?”.
4 VISCERAL Session
The VISCERAL session started with an overview of the challenges in multi-
organ detection and the data that were annotated and made available in the
VISCERAL project. An overview of the results was presented without detailing
the various techniques of the participants. This included a description of the
cloud-based evaluation infrastructure that avoids to physically distribute the
data. The session also included five presentations of participants on the tech-
niques employed in the benchmark.
Spanier et al. [16] presented their approach for multi-organ segmentation
starting with the simplest organs and then going towards harder organs in
the process. The process includes identification of the region of interest for
each organ, thresholding, seed point identification and then slice growing. Gass
et al. [2] used a multi-atlas approach for the segmentation of multiple organs and
also the identification of landmarks. The goal was a data-driven and modality-
independent approach for multi-organ segmentation.
Only liver segmentation is done by Li et al. in [6]. The approach uses mul-
tiple prior knowledge models and an Adaboost classifier, reaching good results
on the liver. Jimenez del Toro et al. [18] presented an approach to multi-organ
segmentation that is entirely data driven and does not use any organ-specific
optimizations. It uses first a global registration and then successive local regis-
trations. For large organs with much contrast a single local registration is used
and then for small organs with less contrast the registration of the larger organs
is refined in a successive manner. Segmentation reached best results in several
of the organs. Kechichian et al. [4] employed multiple graph cut optimization
for multi-organ segmentation. Spatial relationships of organs are modelled and
registration was done using SURF key points to reach good segmentation results.
5 Discussions at the Workshop

The large number of over 80 participants at the workshop also led to a large
number of very interesting questions during the discussions after the talks and
also the lunch and coffee breaks. Many comments after the two invited talks
highlighted both the importance of data availability and systematic testing. For
commercial applications it is clear that robustness is much more important than
only pure performance on very specific data sets as pointed out by Sean Zhou.
The case of face recognitions also highlights the importance that standardised
and publicly available data sets have as well as standardized performance com-
parisons on the development of algorithms. This can really show advances over
the years and it was highlighted that popular believe on best techniques often
does not correspond to the reality of systematic evaluations of it.
The discussions also made clear that theoretical novelty is not necessarily
the main point when building real applications in medical imaging as stabil-
ity is important or redundancy as Erik Learned-Miller emphasised. Several of
the approaches show that clinical impact and importance gain in importance in
the field and that computer vision and machine learning approaches can now well
be applied to large and heterogeneous data sets in medical imaging.Registration
and segmentation remain very important underlying techniques that can help
clinical applications. The session on translational medical imaging also highlights
that there are many potential application areas with a potential real impact. At
this point we would also like to thank the speakers and the workshop participants
for the many discussions and exchange of ideas.
6 Conclusions
The fourth edition of the workshop on medical computer vision at MICCAI
was a clear success. High quality papers and posters were presented and many
discussions on challenges and techniques in medical imaging emerged at the
workshop. The workshop gives a forum for exchange at the crossing of medical
imaging, computer vision, machine learning and techniques to manage large

data sets of heterogeneous nature. Based on the positive experience we foresee
to again hold similar workshops at MICCAI in the coming years to follow up on
developments in this quickly changing research area.
Acknowledgments. This work was supported by the EU in the FP7 through the
VISCERAL (318068) and Khresmoi (257528) projects.
References
1. Cabrera Lozoya, R., et al.: Confidence-based training for clinical data uncertainty
in image-based prediction of cardiac ablation targets. In: Menze, B., et al. (eds.)
MCV 2014. LNCS, vol. 8848, pp. 148–159. Springer, Heidelberg (2014)
2. Gass, T., Szekely, G., Goksul, O.: Multi-atlas segmentation and landmark local-
ization in images with large field of view. In: Menze, B., et al. (eds.) MCV 2014.
LNCS, vol. 8848, pp. 171–180. Springer, Heidelberg (2014)
3. Harmouche, R., Ross, J., Washko, G., San Jose Estepar, R.: Pectoralis muscle
segmentation on CT images based on bayesian graph cuts with a subject-tailored
atlas. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848, pp. 34–44. Springer,
Heidelberg (2014)
4. Kéchichian, R., Valette, S., Sdika, M., Desvignes, M.: Automatic 3D multiorgan
segmentation via clustering and graph cut using spatial relations and hierarchically-
registered atlases. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848, pp.
201–209. Springer, Heidelberg (2014)
5. Langs, G., Müller, H., Menze, B., Hanbury, A.: Visceral: towards large data in
medical imaging - challenges and directions. In: Medical Content-based Retrieval
for Clinical Decision Support, MCBR-CDS 2012, October 2012
6. Li, X., Huang, C., Jia, F., Li, Z., Fang, C., Fan, Y.: Automatic liver segmentation
using statistical prior models and free-form deformation. In: Menze, B., et al. (eds.)
MCV 2014. LNCS, vol. 8848, pp. 181–188. Springer, Heidelberg (2014)
7. Lugauer, F., Zheng, Y., Hornegger, J., Kelm, B.M.: Precise lumen segmentation
in coronary computed tomography angiography. In: Menze, B., et al. (eds.) MCV
2014. LNCS, vol. 8848, pp. 137–147. Springer, Heidelberg (2014)
8. Ma, G., Gao, Y., Wu, G., Wu, L., Shen, D.: Atlas-guided multi-channel forest
learning for human brain labeling. In: Menze, B., et al. (eds.) MCV 2014. LNCS,
vol. 8848, pp. 97–104. Springer, Heidelberg (2014)
9. Maraci, M.A., Napolitano, R., Papageorghiou, A., Noble, J.A.: Object classification
in an ultrasound video using LP-SIFT features. In: Menze, B., et al. (eds.) MCV
2014. LNCS, vol. 8848, pp. 71–81. Springer, Heidelberg (2014)
10. Menze, B., Langs, G., Lu, L., Montillo, A., Tu, Z., Criminisi, A. (eds.): MCV 2012.
LNCS, vol. 7766. Springer, Heidelberg (2013)
11. Menze, B., Langs, G., Montillo, A., Kelm, M., Müller, H., Tu, Z. (eds.): MCV 2013.
12. Menze, B., Langs, G., Tu, Z., Criminisi, A. (eds.): MICCAI 2010 Workshop MCV.
13. Rivest-Hénault, D., Ghose, S., Pluim, J.P.W., Greer, P.B., Fripp, J., Dowling,
J.A.: Fast multiatlas selection using composition of transformations for radiation
therapy planning. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848, pp.
14. Schlegl, T., Ofner, J., Langs, G.: Unsupervised pre-training across image domains
improves lung tissue classification. In: Menze, B., et al. (eds.) MCV 2014. LNCS,
vol. 8848, pp. 82–93. Springer, Heidelberg (2014)
15. Shao, Y., Gao, Y., Yang, X., Shen, D.: CT prostate Deformable segmentation by
boundary regression. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848, pp.
16. Spanier, A.B., Joskowicz, L.: Rule-based ventral cavity multi-organ automatic seg-
mentation in CT scans. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848,
pp. 163–170. Springer, Heidelberg (2014)
17. Stühler, E., Berthold, M.R.: Dementia-related features in longitudinal MRI: track-
ing keypoints over time. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848,
pp. 59–70. Springer, Heidelberg (2014)
18. Jimenez del Toro, O.A., Müller, H.: Hierarchic multi-atlas based segmentation
for anatomical structures: evaluation in the VISCERAL anatomy benchmarks.
In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848, pp. 189–200. Springer,
Heidelberg (2014)
19. van Tulder, G., de Bruijne, M.: Learning features for tissue classification with the
classification restricted boltzmann machine. In: Menze, B., et al. (eds.) MCV 2014.
LNCS, vol. 8848, pp. 47–58. Springer, Heidelberg (2014)
20. Wang, L., Gao, Y., Shi, F., Li, G., Gilmore, J.H., Lin, W., Shen, D.: LINKS:
learning-based multi-source integration framework for segmentation of infant brain
images. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol. 8848, pp. 22–33.
Springer, Heidelberg (2014)
21. Wu, G., Wang, L., Gilmore, J., Lin, W., Shen, D.: Joint segmentation and regis-
tration for infant brain images. In: Menze, B., et al. (eds.) MCV 2014. LNCS, vol.
8848, pp. 13–21. Springer, Heidelberg (2014)
22. Zikic, D., Glocker, B., Criminisi, A.: Classifier-based multi-atlas label propagation
with test-specific atlas weighting for correspondence-free scenarios. In: Menze, B.,
et al. (eds.) MCV 2014. LNCS, vol. 8848, pp. 127–136. Springer, Heidelberg (2014)
Segmentation of Big Medical Data
Joint Segmentation and Registration
for Infant Brain Images
Guorong Wu1, Li Wang1, John Gilmore2, Weili Lin1,

and Dinggang Shen1(&)
1
Department of Radiology and BRIC, University of North Carolina,
Chapel Hill, NC 27599, USA
dgshen@med.unc.edu
2
Department of Psychiatry, University of North Carolina,
Chapel Hill, NC 27599, USA
Abstract. The first year of life is the most dynamic and perhaps the most
critical phase of postnatal brain development. The ability to accurately charac-
terize structure changes is very critical in early brain development studies, which
highly relies on the performance of image segmentation and registration tech-
niques. However, either infant image segmentation or registration, if deployed
independently, encounters much more challenges than the adult brains due to
dynamic appearance change with rapid brain development. Fortunately, image
segmentation and registration of infant images can assist each other to overcome
the above difficulties by harnessing the growth trajectories (temporal corre-
spondences) learned from a large set of training subjects with complete longi-
tudinal data. To this end, we propose a joint segmentation and registration
algorithm for infant brain images. Promising segmentation and registration
results have been achieved for infant brain MR images aged from 2-week-old to
1-year-old, indicating the applicability of our joint segmentation and registration
method in early brain development studies.
1 Introduction
Human brain undergoes rapid physical growth and functional development from birth
to 1 year old. The ability to accurately measure the structural changes from MR
(Magnetic Resonance) images at this period is indispensable for shedding new light on
the exploration of brain development and also the early detection of neurodevelop-
mental disorder. For example, infants with autism were found to have 5 %–10 %
abnormal enlargement in total brain volume at early development stage [1].
However, the automatic image segmentation and registration tools for processing a
large amount of infant brain MR images lag behind the demands from ongoing neu-
roscience/clinical studies. Both image segmentation and registration are challenged by:
(1) the dynamic appearance changes of brain tissues from birth to 1 year old [1], and (2)
the fast and spatially-varied developments of brain anatomy and size, especially in the
first year of life [1]. Consequently, either segmentation or registration, if deployed
independently, is difficult to handle the above challenges.
© Springer International Publishing Switzerland 2014

B. Menze et al. (Eds.): MCV 2014, LNCS 8848, pp. 13–21, 2014.
DOI: 10.1007/978-3-319-13972-2_2
14 G. Wu et al.
Since the imaging-based study on early brain development becomes more and more
popular, a sufficient number of longitudinal infant brain images have been collected in
the past years. Many subjects with complete longitudinal images have been well
segmented by either human interactions or automatic methods with multi-modality
information [2]. As demonstrated by many literatures, appropriate joint segmentation
and registration could significantly improve both of their performances [3]. Motivated
by this, we aim to develop joint segmentation and registration method to overcome the
above challenges. The leverage to achieve this goal is the availability of a large number
of longitudinal infant data (scanned at 2 weeks, 3 months, 6 months, 9 months, and
1 year of age) and their respective segmentation results. It is worth noting that these
valuable data are often ignored when performing segmentation/registration for the new
infant images by current methods.
In this paper, we present an accurate and robust approach for joint segmentation
and registration of any two given infant brain images by using the knowledge learned
from the training infant subjects with the complete longitudinal data and segmentation
results. Specifically, we first establish accurate temporal correspondences for each
training subject with complete longitudinal images, in order to learn the subject-specific
growth trajectories [4]. Then, to deal with the potential large age gap between two new
infant images, we first segment them separately with a sparse patch-based level set
approach that allows each patch in the new infant image to look for similar patches in
the respective training subjects with similar age and further combine the labels of
matched patches in the training subjects to provide a good initialization for the level set
segmentation. Afterwards, deformations between two new infant images can be esti-
mated by deformable registration upon their segmented images, thus avoiding the
challenge of directly registering two new infant images with large appearance differ-
ence. The refined registration allows all matched patches of longitudinal training
images (regardless of their ages) to improve the segmentation of each new infant
image. By alternating these segmentation and registration steps, we can iteratively
refine both segmentation and registration results of two new infant images. The
advantages of our method include: (1) avoiding the direct registration between two new
infant images with dynamic appearance changes, and (2) improving both segmentation
and registration performances by fully using the available information from a number
of training infant subjects with complete longitudinal data and their subject-specific
growth models. In experiments, improved segmentation and registration results have
been achieved for the infant images aged from 2-week-old to 1-year-old.
2 Methods
Our goal here is to register a moving infant brain image Mt2 with a fixed infant brain
image Ft1 and also simultaneously determine tissue maps for Ft1 and Mt2 , where t1 and
t2 are two different ages, each of which could be as young as 2-week-old or as old as
1-year-old. Assume that we have N training subjects I s (s = 1, …, N) with longitudinal
data I s ¼ Its jt ¼ 1; . . .; Ts . For each image sequence I s , we can apply state-of-the-art
4D segmentation method [2] to segment 3D image at each time-point into WM (white
matter), GM (gray matter), and CSF (cerebral-spinal fluid), which can be denoted as
Joint Segmentation and Registration for Infant Brain Images 15

Ls ¼ Lst jt ¼ 1; . . .; Ts . With some human inspection, we can regard these segmen-
tation results of longitudinal training images as the ground truth.
Next, we first estimate the growth trajectory (Sect. 2.1) to determine temporal
correspondences for each point in the longitudinal data Is, as designated by the purple
dash curves in Fig. 1. Second, we use sparse patch-based label fusion method to
calculate the tissue probability maps for Ft1 and Mt2 separately, by using only the
training images at the respective time-point as the atlas images. For
n example, the label
o
fusion on Ft1 selects only the image patches from training images Its1 js ¼ 1; . . .; N at
time-point t1 (as shown in the dash pink box of Fig. 1) to form the dictionary. The
obtained tissue probability maps can be used as a good initialization for level-set
approach for tissue segmentation (Sect. 2.2). In this way, we can just register the two
segmented images, thus avoiding the difficulty of directly registering the original two
images with different appearances (Sect. 2.3). Given the spatial correspondences
between Ft1 and Mt2 , we can further improve the segmentation accuracy by augmenting
the dictionary with additional image patches from training images at all other time-
points, not simply the similar time-point(s) (Sect. 2.4). By alternating these segmenta-
tion and registration steps, we can iteratively refine both segmentation and registration
results for Ft1 and Mt2 .
Fig. 1. The overview of joint segmentation and registration for infant brain images.
2.1 Learning Subject-Specific Growth Trajectories

For each training subject with complete longitudinal data I s and segmentations Ls , a
conventional way is to independently register all follow-up images to the baseline
image (first time-point). However, such independent image registration may tear down
the coherence of temporal correspondences in each longitudinal data. Hence, we go one
step further to apply a 4D image registration method [5] for jointly estimating the
deformation fields Us = {ust |t = 1, …, Ts} that can bring the image at each time-point to
the latent common space. Thus, the temporal deformations ust!t0 from time-point t to
16 G. Wu et al.
time-point t′ can be efficiently computed as the composition of inverse deformation

(ust )−1 (from Ist toward common space) and the forward deformation ust0 (from common
1 s
space to Its0 ): ust!t0 ¼ ust ut0 , where ∘ denote for the deformation composition. In
this way, we can use the estimated temporal deformation fields to form the growth
trajectories of each training subject I s , as indicated by the purple dash curves in Fig. 1.
2.2 Sparse Patch-Based Level Set Segmentation
Estimation of Tissue Probability Maps: The initial tissue probability maps for each
new infant image are very important to initialize the level set approach for achieving
accurate tissue segmentation. Here, we use a patch-based label fusion method to
estimate the tissue probability maps for each new infant image by considering the
training images with same age as the atlases.
Let’s take Ft1 as example, where we consider all training images at t1 time-point,
fðIts1 ; Lst1 Þjs ¼ 1; . . .; Ng, as the atlases. We first affine register all atlases to Ft1 and then
apply deformable registration method to deform all atlases to Ft1 image space. To
determine the tissue type (WM, GM, or CSF) for each image point x in Ft1 , we extract a
referent patch Qðx; t1 Þ Ft1 centered at image point x. Next, we collect a number of
atlas patches Ps ðv; t1 Þ Its1 across all training infant images Its1 at the same time-point
t1, with the center point v sitting within a search neighborhood n(x). Thus, all of these
atlas patches form an over-complete dictionary D(x, t1) = {Ps(v, t1)|v ∊ n
(x), s = 1, …, N}. Since Ft1 and Its1 are at the same time-point t1, the appearances of
these image patches are very similar. For clarity, we vectorize the reference patch Q
(x, t1) into a column vector b. Also, we arrange each atlas patch Ps(v, t1) into a column
vector ap and then assemble them into a matrix A ¼ ap p¼1;...;g , where p = (v, s) is a
bivariate index for the particular atlas patch Ps(v, t1) and g ¼ N jnðxÞj denotes the total
number of atlas patches.
Inspired by the power of sparse representation, we further look for a sparse coef-
ficient vector w to represent the reference patch b by the dictionary matrix A, i.e.,
b Aw, where each element in w indicates the contribution of a particular atlas patch
ap in representing the reference patch b. Thus, the estimation of w falls to the classic
LASSO (Least Absolute Shrinkage and Selection Operator) problem [6–8]:
b ¼ minw kb Awk2 þkkwk1 ; s:t:w [ 0;

w ð1Þ
where λ controls the sparsity of the coefficient vector w. Here, we specifically use
e t1 Þ to denote the set of selected image patches in D(x, t1) with the sparse coeffi-
Dðx;
cient wp > 0. Since the tissue type for each ap is known, we can calculate the tissue
probability w.r.t. WM, GM, and CSF, respectively. After repeating this procedure for
every point in the fixed image Ft1 , we can obtain the tissue probability maps (as shown
by a pink box in Fig. 1) to initialize the level set algorithm for segmenting Ft1 .
Level Set Segmentation: In level set algorithm, we employ three level sets, with their
zero-level surfaces, respectively, denoting for interfaces of WM/GM, GM/CSF, and
CSF/background. The tissue probability maps can be integrated as prior knowledge

into the coupled level set segmentation algorithm to improve the segmentation accu-
racy. Similarly, we can repeat the above procedure to segment the moving image Mt2 ,
except building the patch dictionary from the training images at time-point t2. The
tentative segmentation results of Ft1 and Mt2 are shown in the blue boxes of Fig. 1.
2.3 Symmetric Feature-Based Image Registration

Given segmentations for the fixed and moving images, we can deploy the state-of-the-
art registration method, i.e., HAMMER [9], to register the two segmented images.
Since geometric invariant moment (GMI) features are extracted from the segmented
images, image registration is free of dynamic appearance changes in the original
intensity images. Here, we further improve HAMMER by using the symmetric
deformation estimation strategy, where we simultaneously estimate the deformation
pathways ϕ1 (from fixed image) and ϕ2 (from moving image). It is worth noting that the
deformed fixed image w.r.t. ϕ1 and the deformed moving image w.r.t. ϕ2 should be
similar in the end of registration.
Since ϕ1 and ϕ2 are iteratively refined during registration, we use k (k = 0,.., K) to
denote the iteration. In the beginning of registration (k = 0), F ð0Þ ¼ Ft1 and M ð0Þ ¼ Mt2 ,
ð0Þ ð0Þ
along with the identity deformation pathways /1 and /2 : In the following, we adopt
the hierarchical deformation mechanism in HAMMER for establishing the corre-
ðkÞ ðkÞ
spondence between the deformed fixed image F ¼ Ft1 /1 and the deformed

ðkÞ
moving image M ðkÞ ¼ Mt2 /2 , instead of always using F(0) and M(0) in HAM-
MER. Also, only a small number of key points with distinctive features are selected
from F(k) and the M(k) to establish anatomical correspondences by matching the GMI
ðkÞ
features. The entire deformation pathways /1 and ϕ(k) 2 are steered by the correspon-
dences on these key points by requiring all other non-key points following the
deformations on the nearby key points. With progress of registration, more and more
ðkÞ
key points are selected to refine the deformation pathways /1 and ϕ(k)2 regarding to F
(k)
and M , which is repeated until F and M become very similar in the end of
(k) (k) (k)
registration. Finally, the deformation pathway from fixed image to moving image can
be calculated by ϕ = ϕK1 ∘ (ϕK2 )−1, where ‘∘’ denotes the composition of deformation
1
pathway ϕK1 and the inverse deformation pathway /K2 .
2.4 Joint Segmentation and Registration by Using Growth Trajectories

Given the tentatively estimated deformations, we can further refine the tissue seg-
mentation and then continue to improve the registration results with more accurate
segmentation results. The key to achieve this goal is the augmented dictionary in tissue
segmentation step since the refined registration will allow all matched image patches in
other time-points to assist the segmentation, while these additional image patches are
not included in the initial dictionary.
18 G. Wu et al.
Taking Ft1 as example, the initial dictionary used to segment each image point
x 2 Ft1 uses only the image patches in time-point t1 of training subjects (the dash pink
box in Fig. 1). After registration, we assume the tentative corresponding location of x in
the moving image Mt2 is ϕ(x). Since we know the growth trajectory in each training
image, we can construct an augmented dictionary D*(x, t1) for image point x, by now
including: (1) image patches at time-point t1: the training image patches from the initial
dictionary D(x, t1) (at time-point t1); (2) image patches at time-point t2: the selected
image patches in D e ð/ð xÞ; t2 Þ that is used to represent the moving image patch Q(ϕ
(x), t2), and (3) image patches at all other time points: all temporally corresponded
image patches derived from D e ð/ð xÞ; t2 Þ. Thus, we also know the appearance infor-
mation at all other time points for each image patch in D e ð/ð xÞ; t2 Þ by traversing the
growth trajectory learned in Sect. 2.1.
As shown in Fig. 1, in our joint segmentation and registration framework, seg-
menting infant image at a particular time-point can now utilize the information in the
entire training images (pink and red boxes in Fig. 1). Since such additional subject-
specific image patches are more pertinent to the image point x than the training image
patches from the entire population, the augmented dictionary can provide more useful
information to guide tissue segmentation. By alternating the segmentation and regis-
tration steps, we can achieve state-of-the-art performance for both segmentation and
registration for the infant brain images.
3 Experiments
In the training stage, we collect 24 infant subjects with complete longitudinal data as
the training subjects, where each subject has T1- and T2-weighted images at 2 weeks,
3 months, 6 months, 9 months, and 12 months of age. All images were acquired from a
Siemens head-only 3T MR scanner. T1-weighted images were acquired with 144
sagittal slices at resolution of 1 × 1 × 1 mm3. T2-weighted images with 64 axis slices
were obtained at resolution of 1.25 × 1.25 × 1.95 mm3. For each subject, the
T2-weighted image is aligned to the T1-weighted image at the same time-point and
then further resampled to 1 × 1 × 1 mm3. Note that we use T2-weighted images for
segmenting 2-week-old and 3-month-old infant images, while use T1-weighted images
to segment 6-, 9-, and 12-month-old infant images, considering the strong tissue
contrast in the respective time-points for the respective MR images.
Since most of early brain development studies use the infant image at 1-year-old or
older age as the atlas to discover structure development, we apply our joint segmen-
tation and registration method to align 27 images at 2-week-old and 28 images at
3-month-old as moving images onto a 1-year-old infant image that is used as the fixed
image (top right of Fig. 2). In the following, we evaluate the segmentation and reg-
istration performances one by one.
3.1 Evaluation of Segmentation Results

For each testing infant image, we have the manual segmentation results of WM, GM,
and CSF. Here, we use Dice ratio to quantitatively measure the overlap degree between
manual segmentations (used as ground truth) and our estimated tissue segmentations.
The sparse patch-based level set segmentation algorithm (without using registration to
refine) is used as the baseline method for comparison. The Dice ratios on each tissue
type for 2-week-old and 3-month-old infant brain images are listed in Tables 1 and 2,
respectively. After joint segmentation and registration, our method achieves overall
2.38 % and 1.95 % improvement in segmenting 2-week-old and 3-month-old infant
brain images, respectively. Some typical improvements on 2-week-old infant brain
image (top left of Fig. 2) are displayed in Fig. 2. It is clear that the initial mis-
segmentations (in the blue boxes of Fig. 2) have been successfully corrected based on
more and more accurate image registration (in the red boxes of Fig. 2).
Table 1. The Dice ratios of WM, GM, and CSF on 2-week-old infant brain images
WM GM CSF Overall
Sparse-level set 82.04 ± 5.17 81.02 ± 4.10 73.02 ± 5.77 78.69 ± 3.33
Our method 84.12 ± 2.38 83.53 ± 1.97 76.48 ± 2.98 81.07 ± 1.49
Table 2. The Dice ratios of WM, GM, and CSF on 3-month-old infant brain images
WM GM CSF Overall
Sparse-level set 82.53 ± 4.61 81.25 ± 3.53 75.14 ± 3.38 79.77 ± 2.68
Our method 84.35 ± 1.76 82.44 ± 2.11 77.00 ± 2.12 81.94 ± 1.27
3.2 Evaluation on Registration Results

A typical result of registering 2-week-old image to 1-year-old image is shown in the
bottom right of Fig. 2. Since we have the manually labeled hippocampus region for
both fixed image (1-year-old image as shown in the top right of Fig. 2) and eight
2-week-old infant brain images, we can further quantitatively evaluate the registration
accuracy on 2-week-old images, by measuring the overlap degree between manual
ground-truth and our estimated hippocampus (by deforming the hippocampal region of
the fixed image to the image space of each individual 2-week-old image). In order to
demonstrate the power of joint segmentation and registration, we apply HAMMER [9]
on the baseline segmentations, obtained by using the sparse patch-based level set only
(without refinement by joint registration). The average and standard deviation of Dice
ratios on hippocampus is 70.13 ± 4.69 by HAMMER (based on the baseline seg-
mentations) and 73.48 ± 2.05 by our joint segmentation and registration method,
achieving almost 4.7 % increase in Dice ratio by our proposed method.
20 G. Wu et al.
Fig. 2. Segmentation improvements on 2-week-old infant brain images, from results in blue
boxes to the results in red boxes (Colour figure online).
4 Conclusion
In this paper, we propose a novel joint segmentation and registration method for infant
brain images by using the growth trajectories learned from a large number of training
subjects with complete longitudinal data. Specifically, image segmentation assists the
registration by providing accurate tissue segmentations, which avoid the challenge of
directly registering the two infant brain images with large appearance changes. In
return, the refined image registration can bring more useful information to provide
better tissue probability maps for guiding the level set based segmentation. Promising
results for both segmentation and registration have been achieved, indicating the
potential applicability of our method for early brain development study.
References
1. Knickmeyer, R., Gouttard, S., Kang, C., Evans, D., Wilber, K., Smith, J., Hamer, R., Lin, W.,
Gerig, G., Gilmore, J.: A structural MRI study of human brain development from birth to 2
years. J. Neurosci. 28, 12176–12182 (2008)
2. Wang, L., Shi, F., Yap, P.-T., Gilmore, J.H., Lin, W., Shen, D.: 4D Multi-Modality Tissue
Segmentation of Serial Infant Images. PLoS ONE 7, e44596 (2012)
3. Pohl, K., Fisher, J., Grimson, W.E.L., Kikinis, R., Wells, W.M.: A bayesian model for joint
segmentation and registration. Neuroimage 31, 228–239 (2006)
4. Miller, M.I.: Computational anatomy: shape, growth, and atrophy comparison via diffeo-
morphisms. NeuroImage 23, S19–S33 (2004)
5. Shen, D., Davatzikos, C.: Measuring temporal morphological changes robustly in brain MR
images via 4-dimensional template warping. NeuroImage 21, 1508–1517 (2004)
6. Tibshirani, R.: Regression shrinkage and selection via the lasso. J. R. Stat. Soc. Ser. B (Stat.
Methodol.) 58, 267–288 (1996)
Another random document with
no related content on Scribd:
CHAPTER VI
PURSUED BY AN ALLIGATOR
Bomba was hurled headlong into the green depths of the river. This
in itself would not have bothered him. He could swim like a fish and
was almost as much at home in the water as on land.
But a thrill of terror passed through him as he realized that not far
away was a group of monsters that could swim faster than he could,
and whose terrible jaws, once clamped on him, could bite him in half.
His mind worked with lightning swiftness. He must remain under
water as long as possible. His instinct for direction told him where
the land lay.
So instead of permitting himself to shoot up to the surface, he
remained under and struck out toward the shore.
For more than a minute he shot ahead at a rapid pace. And only
when it seemed as though his lungs would burst did he at last rise to
the surface.
He shook the water from his eyes and looked behind him. He could
see at least half a dozen caymen tearing at the body of the tapir and
others swimming about greedy for their share.
His only hope lay in the possibility that the brutes would be so busy
fighting each other for a portion of the spoils that they would not
notice him at the distance that he had already attained.
But his chase for the tapir had carried him far out into the stream and
the shore seemed still a terrible distance away.
He summoned up all his resolution and struck out for land, keeping
himself as low in the water as possible and moving with scarcely a
ripple.
For a time it looked as though his hope might be realized. Twice he
looked behind him, and each time the brutes still seemed to be
fighting about the body of the tapir.
But the third time he looked back he saw something that filled him
with consternation.
A huge alligator had detached itself from the group and was making
toward him at full speed. He could see the long, scaly body, the fiery,
red eyes and the hideous jaws with their rows of glistening teeth.
Bomba turned and swam for his life.
As the lad swam he measured the distance to the shore. He knew
that the alligator was coming at a faster pace than his own, but he
hoped that the start he had gained might yet enable him to reach the
shore before those terrible jaws should close on him.
It was a close calculation, but he decided that he had a chance. He
put redoubled power into his strokes and swam as he had never
swum before.
But his heart sank as he noted how steep the bank was at the
nearest point. Had the shore been shelving, he might have been
able to find a footing when the water grew shallower and rush up on
land before the cayman could reach him.
But the bank was three or four feet high and ran straight up and
down. He would have to reach up, grasp the edge and pull himself
up. And he realized only too well that before he could do this the
alligator might have caught him.
In his heart Bomba felt that he was doomed.
But his indomitable spirit refused to give up. He called to his aid all
his reserves of force and swam on. As a last resort he had his knife.
He could stab with it, perhaps hold off for a moment the vicious
rushes of his foe. But as against a knife, of course the alligator would
conquer in the end.
One last glance Bomba cast behind. The cayman had made up
much of the distance between them and was now fearfully close. A
minute more perhaps and then——
Bomba cast one farewell look at the sky and felt for his knife.
But as he looked upward, something brushed across his face. It was
the branch of a tree that hung far out over the water.
A ray of hope cut through the blackness of the boy’s despair. With
the quickness of a cat, he grasped the branch and swung himself up
on it out of the water. Even as he did so, he heard beneath him the
snap of the alligator’s jaws.
But those awful jaws missed the boy by a hair’s breadth. The speed
at which the brute was going carried it several yards farther before it
could turn. And by that time, spurred on by dreadful necessity,
Bomba had climbed farther up on the heavy bough and was out of
reach.
The rage of the baffled man-eater was fearful to witness. It churned
the water into foam and emitted frightful bellowings as it leaped half
out of the water, trying to reach the prey of which it had been
cheated. It gnashed its teeth and its red eyes gleamed with fury.
For a full half hour Bomba lay extended on the branch, feeling with
unspeakable relief and delight the strength come back into his
muscles and vitality creep into his exhausted frame.
The alligator still kept up its watch, and Bomba wondered why. Did it
not know that its quarry had escaped? Then he realized that the
brute was not so stupid after all.
For the alligator is a land as well as a water animal, and though its
prey had baffled it for a while, it knew that the boy, soon or late,
would have to descend the tree. And when Bomba should slide
down the trunk to the ground, his enemy meant to clamber up the
bank and be waiting at the foot of the tree to receive him.
Luckily Bomba’s bow was still slung over his shoulder and his arrows
were at his belt. His immersion had been so sudden that he had not
had time to discard them. They had hindered him in his swimming,
but now he blessed the fact that they were with him.
His position was a difficult one to shoot from, for he would have to
use both hands for his bow and depend on his sense of equilibrium
to keep him secure on the branch. If he should lose his balance, he
would topple into the water, and he shuddered to think what would
happen after that.
He solved his problem in part by wrapping both legs closely around
the bough. Then he fitted an arrow to his string and took careful aim.
He knew that about the only vulnerable part of the alligator was the
eye. An arrow striking anywhere else would simply rebound from the
tough hide without inflicting any material damage.
It was some time before Bomba could get a good view of his target.
The cayman kept swimming about with its head half submerged in
the turbid water.
But at last the alligator lifted its head and glared at the boy. At the
same instant Bomba shot.
The arrow went straight to its mark, pierced the creature’s eye and
penetrated to the brain.
There was an awful bellow and a tremendous thrashing about for a
few moments. Then the alligator slowly sank below the surface of the
river.
Bomba waited for some time, but the body did not reappear. The
arrow had done its work well.
Once fully convinced of this, the boy made his way to the trunk of the
tree and slid down it to the ground.
It was good to feel the earth again beneath his feet. His heart
swelled with gratitude. True, he had been cheated. The alligators
had robbed him of his tapir. But they had not robbed him of his life,
and compared with that fact everything else seemed insignificant.
The sun was near its setting now, and his hunting would have to be
deferred to the next day. Soon the four-footed hunters of the jungle
would themselves be abroad, and it behooved Bomba to make
himself safe for the night.
Ordinarily he would have collected brush and made a fire, trusting to
the flames to keep wild beasts at bay. But now, owing to the possible
presence of the dreaded headhunters in the jungle, he dared do
nothing that would betray his presence to the marauders. He had not
detected any signs of them so far, but a sort of sixth sense told him
that they had not yet returned to their homes above the Giant
Cataract.
He gathered some jaboty eggs which he ate raw, and which, with
some cured meat he had brought with him from the cabin of Pipina,
made a simple but satisfying meal.
Then he cast about for some place in which he could pass the night.
He found it in the center of a dense thorn thicket, into which he
penetrated slowly and with great care, pushing the thorny stalks
aside so that they would not wound his flesh. Once in the center, he
beat down enough of the brush to serve him as a bed, and covered it
with bunches of soft moss that he had gathered near the river’s
edge.
Here he was safe. Even if his scent betrayed him, no wild beast was
likely to venture through the thorns. And if, perchance, some
prowling brute, more daring or hungry than the rest, should try it, the
noise made would awaken the boy at once and he could trust to his
weapons for the rest.
As he lay there, waiting for sleep to come, his thoughts were
tinctured somewhat with bitterness. Why should his life be in
constant peril? Why should he be doomed to be hunted by beast and
reptile?
It was not as though he were a native of the jungle. Then he might
have accepted his lot as the decree of fate and borne whatever
came to him with stolidity, if not with resignation.
But his real place was not there. He was white. He was heir to all the
instincts, traditions and ambitions of his race. He belonged
elsewhere. Then why was he here? Why were his aspirations and
longings doomed to be thwarted? What had he done to deserve
such a fate?
His thoughts turned to Frank Parkhurst. What a difference there was
between their lots! No doubt by this time Frank had reached one of
the cities he had talked about and whose wonders had so deeply
stirred the jungle boy. To-night Frank would be sleeping safely in a
soft bed. He would have abundance of good food. He would be
laughing and talking with others of his kind. And his mother, the
woman with the golden hair, would print a good-night kiss upon his
lips.
But Bomba had no mother near him to kiss him good-night. He had
no friends to talk to, to clap on the shoulder in jovial fashion, as Gillis
had done to Dorn. Monkeys and parrots were his only intimates,
except poor demented Casson and the squaw, Pipina.
As to laughter—when had Bomba last laughed? He could not
remember.
In this melancholy mood he at last fell asleep.
But his depression vanished when he awoke the next morning. His
sleep had been undisturbed. Weariness had departed. The current
of his blood ran swiftly through his young veins. The skies were
azure. It was good to be alive.
And his optimism was increased by the good fortune that for the next
two days attended his hunting. Tapirs, deer, agouti and other game
fell before his arrows, until he had accumulated enough to supply the
cabin for weeks to come.
A good deal of his time was consumed in skinning the animals. He
could not carry them home bodily, as he had no means of
conveyance except his sturdy shoulders. So he cut off the choicest
parts, wrapped them in great leaves tied with bush cord, and on the
third day after he had left the cabin set out on his return.
All this time he had kept a sharp lookout for the headhunters. But
during the whole of the hunting trip he had come across no signs of
them.
He would have felt easier in his mind, however, if he had
occasionally met some of the friendly natives that ordinarily dwelt in
that region. They, too, were conspicuous by their absence. If the
headhunters had really gone, why had not the natives returned to
their usual haunts?
The answer was not far to seek. Those dreaded invaders were
probably still lurking somewhere in the district, and if Bomba had not
crossed their trail, it was simply a bit of good fortune.
A sense of impending calamity grew upon him as he neared his
destination. The burden on his shoulders was heavy, but it was
matched by the burden on his heart.
He tried to throw off his depression, assuring himself that it was due
to weariness. Soon he would reach the cabin, be once again with
Casson and Pipina, and all would be well. So he communed with
himself, though not with much conviction.
The journey back to Pipina’s hut was made in rapid time, the jungle
boy being spurred on by his anxiety.
He turned the bend of the stream from which he could see the hut,
and an exclamation of relief escaped him as he saw that it was still
standing. He had half-feared that he would find it burned or
demolished.
But it was there, and everything about it seemed placid and serene.
No puma this time ranged before the door. It seemed an abode of
peace.
“Bomba was a fool,” he told himself, as he hurried forward.
He gave the loud halloo with which he was accustomed to announce
his coming. Usually this brought either Casson or Pipina to the door
at once.
This time his call evoked no answer. No figure appeared at the
doorway, nor could he hear any stirring within.
He called again, this time in louder tones. But again there came no
answer.
His uneasiness returning at this unusual circumstance, he hastily ran
up to the hut.
The door stood open and a hurried glance into the outer room
showed that it was empty.
He rushed in and examined the farther room. No one was there.
“Casson!” he called in a frenzy of anxiety. “Casson! Pipina! Where
are you?”
His only answer was the echo of his voice.
Casson was gone! Pipina was gone!
CHAPTER VII
CARRIED INTO CAPTIVITY
For a moment Bomba stood stupefied with dread. Then he ran out
into the open.
He beat the bushes about the hut. He dashed down to the edge of
the ygapo, his quick eye scanning the expanse for some sign of the
passing of Casson and Pipina.
Nothing anywhere. No footprints, no trampling of the bushes, no clue
to guide him in a search for the missing ones.
To all appearances no one had trodden that deserted spot since
Bomba had returned from his journey to the Moving Mountain.
Yet Casson was gone. Pipina was gone.
Bomba retraced his steps to the hut, his mind in a ferment of
bewilderment and grief.
Indians! Only Indians could traverse the jungle with the silence and
stealth of ghosts, leaving no trace behind.
“Nascanora!” The word hissed between the boy’s clenched teeth.
“This is your work! If you have killed Casson, Bomba will not rest
until he has found your heart with an arrow, a bullet, or a knife!”
The boy reëntered the empty hut with a sharp pain stabbing at his
heart. He would search the cabin more minutely now for some sign,
some clue, to the whereabouts of the absent ones. And if he could
not find it there, he would call into play all his skill in woodcraft to find
and follow their trail. For trail there must be somewhere. They could
not have vanished into thin air.
Inch by inch, he scrutinized the walls, the floor, even every crevice of
the crude and meagre furniture, thinking he might find some
message from Casson. It was almost a forlorn hope, but it was all
that he had at the moment.
He had nearly abandoned even this hope, however, when he
discovered a faint scrawl on the wall in the darkest corner of the hut.
He bent closer, and his brows drew together in a scowl as he tried to
decipher the writing.
Then suddenly a hoarse cry of rage escaped him. His eyes blazed in
the shadowy hut like those of an angry puma.
For this was what he read:
“Nascanora is taking away Casson, Pipina, Hondura to
camp near Giant Cataract. Come. Help.”
So the headhunters had achieved their end at last! They had
captured the helpless old man whom their superstition had led them
to regard as a Man of Evil, a magician whose spells had brought
blight on their crops and sickness to their people. Poor old Casson,
whose one desire was to help rather than hurt!
They would torture him. They would make him a sacrifice to their
gods. And when flesh and blood could no longer stand their
torments, they would kill him and place his head on the wigwam of
their chief.
Bomba’s rage was terrific, and it would have fared ill with the savage
chief if at that moment he had come within reach of the boy’s knife.
The boy read the scrawl again. So they had taken Hondura too, the
friendly chief of the Araos tribe, the father of the pretty little girl,
Pirah, who had once saved Bomba’s life!
Why had Nascanora made Hondura captive with Casson and the
squaw? Was it because Hondura and his tribe had been on friendly
terms with Bomba and the old naturalist? Or did Nascanora think
that the Araos tribe was becoming too powerful, and had he hoped
by depriving them of their chief to render his people helpless and
throw confusion and panic into their hearts?
But that problem could be left till later for solution. The pressing thing
was to plan for the rescue of the captives before they were so far
away that their recovery would be impossible.
When his first fury had exhausted itself, Bomba left the writing on the
wall and sat down on the threshold of the hut, the better to think out
a course of action.
“Near the Giant Cataract” said that rude scrawl penciled by poor
Casson, when the attention of his captors was momentarily diverted.
Bomba had no clear idea as to how far away that was or what
direction would have to be followed to reach it. But from what he had
gathered from the natives from time to time, the place was at a great
distance, and he knew that a long and arduous journey lay before
him.
That is, if he had to go that far in order to catch up with the raiding
party. There was nothing in the hut on which he could base any
judgment as to the time the capture had taken place.
It might have occurred the very day that he had left on his hunting
trip. In that case, they would have had a three days’ start of him. Or
again it might have happened yesterday. In that event, his task of
overtaking them would be that much easier.
No matter how much they wanted to hurry, no large party could
proceed very fast, encumbered as they were with prisoners and
probably laden with spoils. Bomba could cover as much ground in
one day as they could in two or possibly in three. So he had little
doubt of his ability to overtake them before they could get very far.
What he would do if he should come up with them, he did not pause
to consider. No plan formed now would be of any service. He must
be guided by circumstances as they developed. But he had enough
confidence in his own quick wit to believe that he would be able to
shape such circumstances to meet his ends.
But time pressed, and he could spend no more time in reflection. He
rose to his feet and secured some cured meat from Pipina’s stores to
feed him on the way. He would have no time to spend in hunting
food when there were human enemies requiring his attention.
He tested his bow, put a new string on it, and replenished his store of
arrows. His revolver and machete were already at his belt. He took
one more look about the cabin to make sure he had overlooked
nothing and plunged into the jungle.
Like a hound when trying to pick up a trail, he described a long
circle, scanning the ground narrowly for every sign that might
indicate the direction in which the party had gone.
For a long time his search was unsuccessful. At last his eyes
brightened, for they rested on a little strip of cloth fluttering on a thorn
bush. He examined it carefully and recognized it as a strip torn from
a loose dress such as the Indian women wear. Probably it was torn
from Pipina’s garment.
This was something, but not enough. He must find a second strip, if
possible, and the line formed by the two would give him the direction
in which the party were traveling.
Before long his eager search was rewarded. Now he could shape his
course, and he hurried forward with redoubled speed.
The bits of cloth also told the jungle boy another story. The fact that
they had been torn off at all showed that the savages had been
hurrying their captives along at great speed, and so roughly that they
took no care to avoid the thorn bushes that tore the clothes and
probably the skin.
Bomba’s heart burned within him as he pictured poor, weak Casson
driven along, perhaps flogged to make him hasten. How long could
he endure such treatment in his feeble condition? Perhaps even now
he had succumbed to the hardships of the journey! Bomba gritted his
teeth and his eyes flamed with fury.
He had not gone far when his jungle instinct warned him of danger in
the immediate vicinity. Motions as vague as shadows, faint rustlings
that could not have been detected by an untrained ear, told him that
something or someone was trailing him, keeping step with him,
moving as swiftly and silently as he. From being the hunter, he had
become the hunted.
When Bomba paused the slight rustlings stopped. When he moved
on they were resumed.
Still he continued on his way. Whether his pursuer were beast or
human he could not tell. But the jungle lad knew that, whether beast
or human, the surest way to provoke attack was to betray a
knowledge of his danger.
To keep on would be at least to delay attack and perhaps derange
the plans of his pursuer. But when the attack at last came—if it
should come—he would be ready.
Suddenly he became conscious that he was encircled. The faint
sound, which had been behind him, was echoed now on the right
and the left and in the front. His enemies, whoever they were, were
closing in upon him.
There was nothing that his eyes could tell him. Not a leaf stirred nor
was there any movement in the brush. There was only that ghostly
rustling that to Bomba’s sensitive ears was as plainly perceptible as
the rumbling of distant thunder.
Then something shaped itself, vague and dim behind a thicket. Like
a beast at bay, Bomba crouched, pulled his bow from his shoulder
and plucked an arrow from his pouch.
Before he could fit the arrow to the string a hideous chorus of shouts
rent the air, and like magic the jungle was filled with men, men with
the ferocious faces of demons, who rushed upon him, shouting and
brandishing their sharp, murderous knives above their heads.
Bomba had no time to turn and flee. He dropped his bow, whipped
out his machete and backed toward a tree, only to feel himself
seized from behind and borne helpless to the ground.
A score of natives bent above him, their faces menacing, their knives
pointed at his throat.
“Hondura?” grunted one of them. “Where is he?”
A light dawned upon Bomba. These were braves of the Araos tribe
seeking their leader. They blamed him, Bomba, for the
disappearance of their chief—Bomba, who at that very moment was
on his way to rescue that chief from captivity.
“Let me up and I will tell you,” he said, refusing to quail before the
fierce eyes directed upon him.
They hesitated, evidently suspicious of some trick. But finally the
leader, a strapping native named Lodo, whom Bomba remembered
having seen on his recent visit to their village, ordered that a close
ring be formed about the lad. Then he ordered him to stand up.
Bomba did so, and Lodo advanced toward him, knife in hand, his
gaze lowering.
“Where is Hondura, our chief?” he demanded. “You hide him.”
“The bad chief, Nascanora, take good chief, Hondura, prisoner,”
Bomba replied, his brown eyes holding the little, shifting ones of
Lodo with great earnestness. “He take, too, my friend Casson and
Pipina, the squaw. Come with me to the hut, and I will show you the
writing on the wall.”
It was plain that the Indians considered this a subterfuge or a trap,
and there was considerable parleying before Lodo finally announced
that they would take him to the hut.
“But you fool us, Bomba,” threatened Lodo with a suggestive twist of
his knife, “and I cut out your heart—so!”
CHAPTER VIII
THE MAN WITH THE SPLIT NOSE
Bomba raged within himself at this enforced delay in his journey. But
resistance against such odds would be nothing less than suicide.
And apprehension was in his heart as he moved along with his
captors. He was by no means sure that he would be able to prove to
the natives that he spoke the truth concerning their chief.
After all, the only proof he had was that writing on the wall, and if
they thought that he was trying to deceive them they might regard
the writing as part of his plan. There was no likelihood that any of
them would be able to decipher it for themselves.
So it was with no great confidence as to the ultimate outcome that he
made his way back to the cabin, surrounded by the lowering bucks
of the Araos tribe and feeling the suspicious gaze of Lodo boring into
his back.
They traveled swiftly, as the way was familiar, and it was not long
before they reached the deserted hut. Bomba led them into it and
pointed out the faint scrawl on the wall.
Lodo could make nothing of it, and looked from the writing back to
Bomba with a fierce scowl and a tightened grasp upon his knife.
“You read,” he said, and the cruel point of the knife pricked Bomba’s
bare shoulder and brought a tiny trickle of blood. “No fool Lodo.”
Bomba read the words twice under Lodo’s direction, and still the
giant was unconvinced. The other bucks patterned their conduct on
his and crowded around Bomba, muttering and growling like beasts
of the jungle about to close in on a helpless prey.
“You lie!” The point of Lodo’s knife pricked again at Bomba’s
shoulder, deeper this time, and a red stream followed it.
Still Bomba did not flinch, giving the sullen Indian look for look
without a sign of fear.
“You lie!” again shouted Lodo, working himself into a frenzy of fury.
“Braves see Hondura near the hut of Casson, the white medicine
man. Hondura not come back to his people. Bomba hide Hondura.
Bomba must die!”
His knife was upraised in menace. A shudder of anguish passed
through Bomba, but he said no word. A dozen hands reached out to
seize him, a dozen knives were pointed at his throat——
“Wait!” A guttural voice came from the doorway of the hut, and Grico,
he of the one eye and the split nose, forced his way through the ring
of angry Indians. “Let me see writing on the wall. I tell you if Bomba
lie.”
Grico was a native of tremendous physical strength. He had for a
while lived in one of the towns on the coast, and as a boy had been
taken under the care of an English missionary school. Here he had
been taught the rudiments of education. But the call of his jungle
blood had proved too strong to be resisted, and he had run away
and thrown in his lot with the Araos tribe.
There, when he reached manhood, he became known as the
swiftest runner and the greatest hunter of them all. He had lost one
eye and acquired his split nose in a battle with jaguars, in which he
had shown almost superhuman strength and courage.
He had become therefore a person of great influence in the tribe, not
only because of his prowess as a hunter but also because of his
knowledge of the world outside and his education, which gave him
great superiority over his ignorant and simple-minded mates. And he
was cunning enough to make this count for all that it was worth and
considerably more, considering that, after all, he had gained only the
merest smattering of learning.
Bomba knew something of the history of Grico, and hope sprang in
him anew as the giant caboclo made his way through the sullen
group and peered at that faint scrawl upon the wall. Slowly he read
the words aloud:
“Nascanora is taking away Casson, Pipina, Hondura to
camp near Giant Cataract. Come. Help.”
Then he turned to the Indians, his one eye gleaming at them in a
contemptuous manner.
“Bomba speaks truth,” he said. “Those are the words he said, and
that is the writing on the wall. Take the point of your knife away,
Lodo. The boy is right. Chief Hondura has been taken away by the
wicked Nascanora. We will capture Nascanora and tie him to a tree
and pile the brushwood up about his knees and with flint and stone
make red flames that will lick at his flesh and bones. Ayah! Ayah!”
The cry of vengeance, indescribably weird and savage, was taken
up by the Indians and filled the jungle with a long wailing shriek that
chilled the blood of Bomba as he thought of what might have
happened to him had it not been for Grico’s timely appearance.
Once their enmity was turned from him, the Indians became as
friendly and pleasant as they had been savage a few moments
before.
One of them found a bit of native cloth in the hut and bound it about
the wound in Bomba’s shoulder. Then they squatted outside the hut
to hold a council and decide upon plans for the rescue of their chief.
While the Indians talked in their guttural language, sometimes sitting
for many minutes of silence between their laconic sentences, Bomba
fretted and fumed, eager to be once more on the trail of the
headhunters.
However, he could not risk offending Lodo and his braves, who might
prove valuable helpers in his quest, by going off too abruptly. So he
waited, answering as best he could the questions that Lodo and
Grico put to him from time to time.
He told them of the previous visits of the headhunters to Casson’s
cabin, how he had beaten off their attacks, how he had wounded
Nascanora himself, how he had captured Ruspak, their medicine
man, and how in the jungle he had overcome the braves of Tocarora.
“I think there are two bands of headhunters,” he said, when asked for
his opinion. “Nascanora is at the head of one, and Tocarora, his half-
brother, leads the other. They will join each other somewhere in the
jungle. Then Nascanora will be strong enough to fight the Araos if
they come looking for their chief.”
“But they did not think of Grico!” said he of the split nose, his one eye
gleaming balefully. “Grico will get other braves who will fight with the
Araos to get back Hondura. Hondura has been good to Grico. Grico
will show that he is grateful.”
Bomba knew that this was no idle boast. Grico’s courage was so well
established that he would have no trouble in rallying many bucks of
other tribes under his leadership. If he could do this, the Araos and
their auxiliaries would prove formidable enemies, even to the
redoubtable headhunters.
But Bomba knew that this would take time and be the subject of
innumerable powwows before the avengers would get fairly started.
And with the knowledge he had of Casson’s danger he was in no
mood for delay. He must go on. They could follow later.
He knew the risks he ran in going alone. He would be only one
against the hordes of Nascanora. But his wit had served him so often
where mere force would have failed that he was willing and eager to
trust to it again.
So when a favorable opportunity presented, he broached his plan to
his new-found allies.
“Bomba will go first and try to find the trail while Lodo and Grico are
getting their braves together,” he said. “If Bomba finds them near by,
he will come back and tell you, so that you may come and fight them.
If he finds they are far off, he will leave a trail of his own so that you
may move fast through the jungle. Does Bomba speak well?”
Lodo and Grico consulted.
“Bomba speaks well—” Lodo was beginning when there was a
sudden commotion in the surrounding jungle. A moment later a
strange company broke into the clearing.
Bomba saw that they were squaws of the Araos tribe. The faces of
the women, usually so stolid, wore the ghastly gray of terror. They
had come from afar and swiftly, for their flesh was torn by thorns and
spiked vines and their breath came in gasps.
One of them, who was the squaw of Lodo and seemed to be the
leader of the women, came over to her husband and stood before
him, striving to regain breath enough to speak.
“What is it, woman?” cried Lodo. “What has happened?”
“The maloca,” she got out at last. “Bucks come. They burn our
houses. They carry off the women. They take the children. All gone.”
CHAPTER IX
THE SAVAGE RAIDERS
The news fell with stunning effect upon the assembled bucks already
wrought up to a high pitch of fury because of the capture of their
chief. There was a hubbub of exclamations of grief and rage.
“Pirah gone, too,” declared the woman, with a sweeping gesture of
her hands.
“Pirah!” This was Bomba’s voice, harsh and unlike his own, as he
pushed through the group of scowling Indians. “You say Pirah taken
too? Who were the bucks?”
“No know.” The squaw shook her head. She had recovered her
breath now, and, with the other women, was regaining her habitual
stolid look. “They come—many come.” She extended her ten fingers,
closing and opening her hands several times to indicate an indefinite
number. “Take women. Take children. We run away. Hide in bushes
till they go. Then we come here.”
Her story roused the Indians even more than the loss of their chief.
That their native village, or maloca, should have been invaded, their
women and children carried off, was a crime that by their code
merited only one punishment, and that torture and death.
The powwow broke up at once and the braves hurried away to their
maloca, taking their women with them.
Bomba accompanied them for some distance, and then took leave of
them, promising to pick up the trail at once and coöperate with them
in every way possible in the rescue of their chief and kidnapped
families, as well as of Casson and Pipina.
Then he left them to return with heavy and angry hearts to their
despoiled maloca while he again sought out the path he had been

Download textbook Medical Computer Vision Algorithms For Big Data International Workshop Mcv 2014 Held In Conjunction With Miccai 2014 Cambridge Ma Usa September 18 2014 Revised Selected Papers 1St Edition Bjoern Menze ebook all chapter pdf

Uploaded by

Copyright:

Available Formats

You might also like

Download textbook Medical Computer Vision Algorithms For Big Data International Workshop Mcv 2014 Held In Conjunction With Miccai 2014 Cambridge Ma Usa September 18 2014 Revised Selected Papers 1St Edition Bjoern Menze ebook all chapter pdf

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Download textbook Medical Computer Vision Algorithms For Big Data International Workshop Mcv 2014 Held In Conjunction With Miccai 2014 Cambridge Ma Usa September 18 2014 Revised Selected Papers 1St Edition Bjoern Menze ebook all chapter pdf

Uploaded by

Copyright:

Available Formats

Medical Computer Vision Algorithms

for Big Data International Workshop

Spatio temporal Image Analysis for Longitudinal and

Medical Computer Vision Large Data in Medical Imaging

Clinical Image Based Procedures Translational Research

Statistical Atlases and Computational Models of the

Bayesian and grAphical Models for Biomedical Imaging

Medical Image Computing and Computer Assisted

Multimodal Analyses enabling Artificial Agents in Human

Combinatorial Algorithms 25th International Workshop

Albert Montillo Michael Kelm

Henning Müller Shaoting Zhang

Weidong (Tom) Cai Dimitris Metaxas (Eds.)

Medical Computer Vision:

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Library of Congress Control Number: 2014958150

Springer Cham Heidelberg New York Dordrecht London

Printed on acid-free paper

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

October 2014 Bjoern Menze

International Program Committee

Koen Van Leemput Technical University of Denmark, Denmark

European Commission 7th Framework Programme, VISCERAL (318068) and

Overview of the 2014 Workshop on Medical Computer Vision—Algorithms

Segmentation of Big Medical Data

Joint Segmentation and Registration for Infant Brain Images . . . . . . . . . . . . 13

LINKS: Learning-Based Multi-source IntegratioN FrameworK for Segmentation

Pectoralis Muscle Segmentation on CT Images Based on Bayesian Graph Cuts

Advanced Feature Extraction

Learning Features for Tissue Classification with the Classification

Dementia-Related Features in Longitudinal MRI: Tracking Keypoints

Object Classification in an Ultrasound Video Using LP-SIFT Features. . . . . . 71

Unsupervised Pre-training Across Image Domains Improves Lung

Multi-atlas and Beyond

Atlas-Guided Multi-channel Forest Learning for Human Brain Labeling. . . . . 97

Fast Multiatlas Selection Using Composition of Transformations

Classifier-Based Multi-atlas Label Propagation with Test-Specific Atlas

Translational Medical Computer Vision

CT Prostate Deformable Segmentation by Boundary Regression . . . . . . . . . . 127

Precise Lumen Segmentation in Coronary Computed

Confidence-Based Training for Clinical Data Uncertainty

Rule-Based Ventral Cavity Multi-organ Automatic Segmentation

Multi-atlas Segmentation and Landmark Localization in Images

Automatic Liver Segmentation Using Statistical Prior Models

Hierarchic Multi–atlas Based Segmentation for Anatomical Structures:

Automatic 3D Multiorgan Segmentation via Clustering and Graph Cut

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

Henning Müller1,2(B) , Bjoern Menze3,4 , Georg Langs5,6 ,

Abstract. The 2014 workshop on medical computer vision (MCV):

Keywords: Medical image analysis · Medical computer vision · Seg-

2 Papers Presented at the Workshop

2.1 Segmentation of Big Medical Data

Wu et al. [21] addressed the problem of segmentation and registration of infant

muscle in CT. Their approach constructs a likelihood using a multivariate distri-

2.2 Advanced Feature Extraction

2.3 Multi Atlas and Beyond

Ma et al. [8] kicked oﬀ our Multi-Atlas section by presenting a hybrid approach

2.4 Translational Medical Computer Vision

The translation of concepts from computer vision to applications in the medical