Big Data in Multimodal Medical Imaging

1st Edition Ayman El-Baz (Editor)

Visit to download the full and correct content document:
https://textbookfull.com/product/big-data-in-multimodal-medical-imaging-1st-edition-ay
man-el-baz-editor/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Level Set Method in Medical Imaging Segmentation 1st

Edition Ayman El-Baz (Editor)

https://textbookfull.com/product/level-set-method-in-medical-
imaging-segmentation-1st-edition-ayman-el-baz-editor/

Diabetes and Retinopathy 1st Edition Ayman S. El-Baz

https://textbookfull.com/product/diabetes-and-retinopathy-1st-
edition-ayman-s-el-baz/

Neural Engineering Techniques for Autism Spectrum

Disorder Volume 1 Imaging and Signal Analysis 1st
Edition Ayman S. El-Baz

https://textbookfull.com/product/neural-engineering-techniques-
for-autism-spectrum-disorder-volume-1-imaging-and-signal-
analysis-1st-edition-ayman-s-el-baz/

Diabetes and Retinopathy: Computer-Assisted Diagnosis

Volume 2 Ayman S. El-Baz

https://textbookfull.com/product/diabetes-and-retinopathy-
computer-assisted-diagnosis-volume-2-ayman-s-el-baz/
Multimodal Imaging in Uveitis 1st Edition H. Nida Sen

https://textbookfull.com/product/multimodal-imaging-in-
uveitis-1st-edition-h-nida-sen/

Big Data in Omics and Imaging Association Analysis 1st

Edition Momiao Xiong

https://textbookfull.com/product/big-data-in-omics-and-imaging-
association-analysis-1st-edition-momiao-xiong/

Big Data and Ethics: The Medical Datasphere 1st Edition

Jérôme Béranger

https://textbookfull.com/product/big-data-and-ethics-the-medical-
datasphere-1st-edition-jerome-beranger/

Multiscale Multimodal Medical Imaging First

International Workshop MMMI 2019 Held in Conjunction
with MICCAI 2019 Shenzhen China October 13 2019
Proceedings Quanzheng Li
https://textbookfull.com/product/multiscale-multimodal-medical-
imaging-first-international-workshop-mmmi-2019-held-in-
conjunction-with-miccai-2019-shenzhen-china-
october-13-2019-proceedings-quanzheng-li/

Medical Imaging Systems Andreas Maier

https://textbookfull.com/product/medical-imaging-systems-andreas-
maier/
Big Data in Multimodal
Medical Imaging
Big Data in Multimodal
Medical Imaging

Edited by
Ayman El-Baz and Jasjit S. Suri
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2020 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-138-50453-0 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have
been made to publish reliable data and information, but the author and publisher cannot assume responsibility for
the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace
the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission
to publish in this form has not been obtained. If any copyright material has not been acknowledged please write
and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted,
or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, includ-
ing photocopying, microfilming, and recording, or in any information storage or retrieval system, without written
permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers,
MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety
of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment
has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only
for identification and explanation without intent to infringe.

Library of Congress Cataloging‑in‑Publication Data

Names: El-Baz, Ayman S., editor. | Suri, Jasjit S., editor.

Title: Big data in multimodal medical imaging / edited by Ayman El-Baz and
Jasjit S. Suri.
Description: Boca Raton : CRC Press, [2020] | Includes bibliographical
references and index. | Summary: “There is an urgent need to develop and
integrate new statistical, mathematical, visualization, and
computational models with the ability to analyze Big Data in order to
retrieve useful information to aid clinicians in accurately diagnosing
and treating patients”-- Provided by publisher.
Identifiers: LCCN 2019025184 (print) | LCCN 2019025185 (ebook) | ISBN
9781138504530 (hardback) | ISBN 9781315146102 (ebook)
Subjects: LCSH: Diagnostic imaging--Data processing. | Big data.
Classification: LCC RC78.7.D53 B54 2020 (print) | LCC RC78.7.D53 (ebook)
| DDC 616.07/54--dc23
LC record available at https://lccn.loc.gov/2019025184
LC ebook record available at https://lccn.loc.gov/2019025185

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
 With love and affection to my mother and father,
whose loving spirit sustains me still
Ayman El-Baz
To my late loving parents, immediate family, and children
Jasjit S. Suri
Contents

Preface............................................................................................................................... ix
Editors...............................................................................................................................xi
Contributors....................................................................................................................xiii
Acknowledgments...........................................................................................................xvii

1 Multimodal Imaging Radiomics and Machine Learning............................... 1

Gengbo Liu, Youngho Seo, Debasis Mitra, and Benjamin L. Franc

2 Multimodal Medical Image Fusion in NSCT Domain................................... 23

Gaurav Bhatnagar, Zheng Liu, and Q.M. Jonathan Wu

3 Computer Aided Diagnosis in Pre-Clinical Dementia: From Single-

Modal Metrics to Multi-Modal Fused Methodologies.................................... 47
Yi Lao, Sinchai Tsao, and Natasha Lepore

4 Automated Diagnosis and Prediction in Cardiovascular Diseases

Using Tomographic Imaging............................................................................... 71
Lisa Duff and Charalampos Tsoumpas

5 Big Data in Computational Health Informatics........................................... 103

Ruogu Fang, Yao Xiao, Jianqiao Tian, Samira Pouyanfar, Yimin Yang,
Shu-Ching Chen, and S. S. Iyengar

6 Fast Dual Optimization for Medical Image Segmentation........................ 143

Jing Yuan, Ismail Ben Ayed, and Aaron Fenster

7 Non-Parametric Bayesian Estimation of Rigid Registration for

Multi-Contrast Data in Big Data Analysis..................................................... 169
Stathis Hadjidemetriou, and Ismini Papageorgiou

8 Multimodal Analysis in Biomedicine.............................................................. 193

Mohammad-Parsa Hosseini, Aaron Lau, Kost Elisevich,
and Hamid Soltanian-Zadeh

vii
viii Contents

9 Towards Big Data in Acute Renal Rejection.................................................205

Mohamed Shehata, Ahmed Shalaby, Ali Mahmoud,
Mohammed Ghazal, Hassan Hajjdiab, Mohammed A. Badawy,
Mohamed Abou El-Ghar, Ashraf M. Bakr, Amy C. Dwyer, Robert Keynton,
Adel Elmaghraby, and Ayman El-Baz

10 Overview of Deep Learning Algorithms Applied to Medical Images......225

Behnaz Abdollahi, Ayman El-Baz, and Hermann B. Frieboes

11 Big Data in Prostate Cancer..............................................................................239

Islam Reda, Ashraf Khalil, Mohammed Ghazal, Ahmed Shalaby,
Mohammed Elmogy, Ahmed Aboelfetouh, Ali Mahmoud,
Mohamed Abou El-Ghar, and Ayman El-Baz

12 Automatic Detection of Early Signs of Diabetic Retinopathy Based

on Feature Fusion from OCT and OCTA Scans............................................263
Nabila Eladawi, Ahmed ElTanboly, Mohammed Elmogy,
Mohammed Ghazal, Ali Mahmoud, Ahmed Aboelfetouh, Alaa Riad,
Magdi El-Azab, Jasjit S. Suri, Guruprasad Giridharan, and Ayman El-Baz

13 Computer Aided Diagnosis System for Early Detection of Diabetic

Retinopathy Using OCT Images....................................................................... 281
Ahmed ElTanboly, Ahmed Shalaby, Ali Mahmoud, Mohammed Ghazal,
Andrew Switala, Fatma Taher, Jasjit S. Suri, Robert Keynton,
and Ayman El-Baz

Index............................................................................................................................... 301
Preface

This book covers the state-of-the-art approaches for big data in the field of medical
image analysis and healthcare. Big data refers to the amalgamation and processing of
huge datasets that are composed of different data types (e.g. clinical, genomic, imag-
ing, pathological, etc.) and have rapidly become more massive and complex, particu-
larly with the advent of new technologies. Broadly speaking, we are considering data of
high volume that are obtained from a variety of sources and processed at high speed.
Currently, big data is a very active area of research and is significantly dominating in
the medical research field, where it achieved very promising results with medical imag-
ing, content-based image retrieval, healthcare automation, diagnosis, and prediction of
many diseases.
Moreover, big data analysis is now trending in many commercial applications in
healthcare, which include, but are not limited to, designing computer aided diagno-
sis systems that could be used in helping physicians in diagnosing and selecting the
optimal treatment plan for many diseases. Various studies are discussed in this book
to demonstrate how powerful big data is in the medical field. This involves radiomics,
multimodal image fusion, image segmentation, and registration, as well as applications
to computational health information, pre-clinical dementia, cardiovascular diseases,
acute renal rejection, prostate cancer, and diabetic retinopathy,
In summary, the main aim of this book is to help advance scientific research within
the broad field of big data. The book focuses on major trends and challenges in this
area, and it presents work aimed at identifying the new achievements and their use in
biomedical analysis.

Ayman El-Baz
Jasjit S. Suri

ix
Editors

Ayman El-Baz is a Professor, University Scholar,

and Chair of the Bioengineering Department at
the University of Louisville, Kentucky. Dr. El-Baz
earned his B.Sc. and M.Sc. degrees in electri-
cal engineering in 1997 and 2001, respectively. He
earned his Ph.D. in electrical engineering from the
University of Louisville in 2006. In 2009, Dr. El-Baz
was named a Coulter Fellow for his contributions to
the field of biomedical translational research. Dr.
El-Baz has 17 years of hands-on experience in the
fields of bio-imaging modeling and non-invasive com-
puter-assisted diagnosis systems. He has authored
or coauthored more than 500 technical articles (133
journals, 25 books, 57 book chapters, 212 refereed-
conference papers, 143 abstracts, and 27 US patents
and disclosures).

Jasjit S. Suri is an innovator, scientist, visionary,

industrialist, and an internationally known world
leader in biomedical engineering. Dr. Suri has
spent over 25 years in the field of biomedical engi-
neering/devices and its management. He received
his Ph.D. from the University of Washington,
Seattle, and his Business Management Sciences
degree from Weatherhead, Case Western Reserve
University, Cleveland, Ohio. Dr. Suri was awarded
the President’s Gold medal in 1980 and made Fellow
of the American Institute of Medical and Biological
Engineering for his outstanding contributions.
In 2018, he was awarded the Marquis Life Time
Achievement Award for his outstanding contribu-
tions and dedication to medical imaging and its
management.

xi
Contributors

Behnaz Abdollahi Shu-Ching Chen

Department of Electrical and Computer School of Computing & Info. Sciences
Engineering Florida International University
University of Louisville Miami, Florida
Louisville, Kentucky
Lisa Duff
Mohamed Abou El-Ghar Biomedical Imaging Science Department,
Radiology Department, Urology and Leeds Institute of Cardiovascular and
Nephrology Center Metabolic Medicine
Mansoura University University of Leeds
Mansoura, Egypt Leeds, United Kingdom

Ahmed Aboelfetouh Amy C. Dwyer

Information Systems Department, Kidney Transplantation–Kidney Disease
Faculty of Computers and Information Center
Mansoura University University of Louisville
Mansoura, Egypt Louisville, Kentucky

Ismail Ben Ayed Nabila Eladawi

École de technologie supérieure Bioengineering Department
Montreal, Québec University of Louisville
Louisville, Kentucky
Mohammed A. Badawy
Magdi El-Azab
Radiology Department, Urology and
Mathematics and Physical Engineering
Nephrology Center
Department
Mansoura University
Mansoura University
Mansoura, Egypt
Mansoura, Egypt

Ashraf M. Bakr Kost Elisevich

Pediatric Nephrology Unit, Mansoura
University Children’s Hospital
Mansoura University
Mansoura, Egypt
Gaurav Bhatnagar
Department of Mathematics
Indian Institute of Technology Jodhpur
Jodhpur, India
Department of Clinical Neurosciences
Spectrum Health
Grand Rapids, Michigan
Adel Elmaghraby
Computer Engineering and Computer
Science Department
University of Louisville
Louisville, Kentucky

xiii
xiv Contributors

Mohammed Elmogy Guruprasad Giridharan

Bioengineering Department Bioengineering Department
University of Louisville University of Louisville
Louisville, Kentucky Louisville, Kentucky

Ahmed ElTanboly Stathis Hadjidemetriou

Bioengineering Department Department of Electrical Engineering
University of Louisville Cyprus University of Technology
Louisville, Kentucky Limassol, Cyprus
and
Mathematics Department
Hassan Hajjdiab
Mansoura University
Electrical and Computer Engineering
Mansoura, Egypt
Department
Abu Dhabi University
Ruogu Fang
Abu Dhabi, United Arab Emirates
Department of Biomedical Engineering
University of Florida
Gainesville, Florida Mohammad-Parsa Hosseini
Bioengineering Department
Aaron Fenster School of Engineering
Robarts Research Institute Santa Clara University
Western University Santa Clara, California
London, Ontario
S. S. Iyengar
Benjamin L. Franc School of Computing & Info. Sciences
Department of Radiology Florida International University
Stanford University Miami, Florida
Stanford, California
Robert Keynton
Hermann B. Frieboes
Bioengineering Department
Department of Electrical and Computer
University of Louisville
Engineering
Louisville, Kentucky
University of Louisville
Louisville, Kentucky
and Ashraf Khalil
Department of Bioengineering Electrical and Computer Engineering
University of Louisville Department
Louisville, Kentucky Abu Dhabi University
and Abu Dhabi, United Arab Emirates
James Graham Brown Cancer Center
University of Louisville Yi Lao
Louisville, Kentucky Department of Radiation Oncology
Cedars Sinai Medical Center
Mohammed Ghazal Los Angeles, California
Electrical and Computer Engineering and
Department Department of Radiology
Abu Dhabi University Children’s Hospital Los Angeles
Abu Dhabi, United Arab Emirates Los Angeles, California
 Contributors xv

Aaron Lau Samira Pouyanfar

Bioengineering Department School of Computing and Information
School of Engineering Sciences
Santa Clara University Florida International University
Santa Clara, California Miami, Florida
Natasha Lepore Islam Reda
Department of Radiology Bioengineering Department
Children’s Hospital Los Angeles University of Louisville
Los Angeles, California Louisville, Kentucky
and
Department of Biomedical Engineering Alaa Riad
University of Southern California Information Systems Department,
Los Angeles, California Faculty of Computers and Information
Mansoura University
Gengbo Liu Mansoura, Egypt
Department of Computer Engineering
and Sciences Ahmed Shalaby
Florida Institute of Technology BioImaging Laboratory
Melbourne, Florida Bioengineering Department
University of Louisville
Zheng Liu Louisville, Kentucky
School of Electrical Engineering and
Computer Science Mohamed Shehata
University of Ottawa BioImaging Laboratory
Ottawa, Ontario Bioengineering Department
and University of Louisville
National Research Council Canada Louisville, Kentucky
Ottawa, Ontario
Youngho Seo
Ali Mahmoud Department of Radiology and Biomedical
BioImaging Laboratory Imaging
Bioengineering Department University of California, San Francisco
University of Louisville San Francisco, California
Louisville, Kentucky
Hamid Soltanian-Zadeh
Debasis Mitra Control and Intelligent Processing Center
Department of Computer Engineering of Excellence (CIPCE)
and Sciences School of Electrical and Computer
Florida Institute of Technology Engineering
Melbourne, Florida University of Tehran
Tehran, Iran
Ismini Papageorgiou and
Institute of Diagnostic and Interventional Radiology and Research Administration
Radiology Henry Ford Health System
University Hospital of Jena Detroit, Michigan
Jena, Germany
and Andrew Switala
Institute of Radiology Bioengineering Department
Südharz Hospital Nordhausen University of Louisville
Nordhausen, Germany Louisville, Kentucky
xvi Contributors

Fatma Taher Q.M. Jonathan Wu

Computer Engineering Department of Electrical and Computer
Zayed University Engineering
Dubai, United Arab Emirates University of Windsor
Windsor, Ontario
Jianqiao Tian
Department of Biomedical Engineering Yao Xiao
University of Florida Department of Biomedical Engineering
Gainesville, Florida University of Florida
Gainesville, Florida
Sinchai Tsao
Department of Radiology Yimin Yang
Children’s Hospital Los Angeles School of Computing and Information
Los Angeles, California Sciences
and Florida International University
Department of Biomedical Engineering Miami, Florida
University of Southern California
Los Angeles, California Jing Yuan
School of Mathematics and Statistics
Charalampos Tsoumpas Xidian University
Biomedical Imaging Science Department Xi’an, China
Leeds Institute of Cardiovascular and
Metabolic Medicine
University of Leeds
Leeds, United Kingdom
Acknowledgments

The completion of this book could not have been possible without the participation
and assistance of so many people whose names may not all be enumerated. Their con-
tributions are sincerely appreciated and gratefully acknowledged. However, the edi-
tors would like to express their deep appreciation and indebtedness particularly to Dr.
Ali H. Mahmoud, Yaser ElNakieb, Omar Dekhil, and Mohamed Ali for their endless
support.

Ayman El-Baz
Jasjit S. Suri

xvii
Chapter 1
Multimodal Imaging Radiomics
and Machine Learning
Gengbo Liu, Youngho Seo, Debasis Mitra, and Benjamin L. Franc

1.1 Introduction................................................................................................................ 1
1.2 High-Quality Standardized Imaging Data Acquisition........................................... 2
1.3 Segmentation of the Region of Interest from the Image.......................................... 3
1.4 High-Throughput Feature Extraction....................................................................... 3
1.4.1 First-Order Features with Image Statistics.................................................. 3
1.4.2 Shape- and Size-Based Features.................................................................... 4
1.4.3 Second-Order Features................................................................................... 4
1.4.4 Wavelet-Based Features................................................................................. 5
1.4.5 Feature Extraction Software.......................................................................... 5
1.5 Clinical Predictive Modeling with Machine Learning............................................. 6
1.5.1 Feature Selection Methods............................................................................. 6
1.5.2 Radiomic Features Classifying Outcomes..................................................... 7
1.5.2.1 Clustering Algorithms....................................................................... 7
1.5.2.2 Regression Model............................................................................... 7
1.5.2.3 Support Vector Machine.................................................................... 8
1.6 Big Data and Machine Learning.............................................................................. 9
1.7 Conclusions............................................................................................................... 10
References........................................................................................................................ 11
Appendix........................................................................................................................... 14
A. First-Order Statistics Features....................................................................... 14
B. Shape- and Size-Based Features..................................................................... 16
C. GLCM Features................................................................................................ 16
D. GLRLM Features............................................................................................. 20
E. Neighborhood Gray Tone Difference Matrix (NGTDM) Features................. 21

1.1 Introduction
According to Gerlinger et al. [1], single tumor biopsy samples can cause underesti-
mation of intratumor heterogeneity. The heterogeneity in the gene, cell and tissue of
the solid tumor limited the accuracy and representation of invasive detection results. In
order to develop the concept of personalized medicine and attain a better understanding
of tumor heterogeneity noninvasively, the concept of radiomics was proposed by Lambin
et al. [2] in 2012. A suffix of “omics”, such as genomics, refers to the objects of study of
a certain field and its relationships to biology and medicine. As genomics attempts to
relate genes to phenotype, radiomics studies relate medical images to phenotypes of

1
2 Big Data in Multimodal Medical Imaging

disease. Much of radiomics literature has focused on applications in oncologic medical

imaging which presents a noninvasive, often quantitative observation of the overall
shape of the tumor, the tumor development process and treatment response monitor-
ing, providing a reliable mechanism to quantify tumor heterogeneity. Lambin et al. [2]
hypothesized that microscopic-level gene or protein pattern changes can be expressed
in macroscopic imaging features acquired during medical imaging. Therefore, Lambin
et al. proposed that the high-throughput extraction of large amounts of image features
from radiographic images would be able to capture intratumoral heterogeneity in a non-
invasive way. Compared to traditional radiology practice, which is primarily based on
visual interpretation and simple quantitative measurements, radiomics can dig deeper
into data contained within medical images and potentially provide further objective
support for clinical decisions. Due to the large number of image features generated from
radiomics, many machine learning algorithms are ideal tools to explore the relation-
ship between radiomics and tumor diagnosis, treatment and prognosis [3, 4, 5].
Four general steps of an ideal machine learning model over radiomic features have
been described in the literature [3, 6] and are summarized in Figure 1.1: (a) high-qual-
ity standardized imaging data acquisition; (b) region of interest image segmentation; (c)
high-throughput feature extraction from images; (d) clinical predictive model establish-
ment. Machine learning algorithms were mainly used in the fourth step to predict the
patient outcomes. In the following sections, the first three steps will be described briefly
and the fourth step will be described in detail.

1.2 High-Quality Standardized Imaging Data Acquisition

The radiomics approach is based on machine learning a large amount of data. However,
maintaining the same standard of image quality in diverse datasets is a problem. High-
quality standardized image data acquisition is near impossible. This is owing to the fact
that the parameter settings during image acquisition and reconstruction may be differ-
ent between exams as well as sites, and there is no uniform standard. Even on the same
equipment, contrast agent dose, image pixel size and slice thickness, filters used in the

FIGURE 1.1: Schematic diagram of workflow.

 Big-Data Radiomic Features in Medical Imaging 3

image reconstruction algorithm, etc., influence the quality of images. For this reason,
many standardized image datasets are being made available for the research community.
For example, the National Institutes of Health (NIH) and the National Cancer Institute
(NCI) have established the Lung Image Database Consortium (LIDC) [7], The Cancer
Genome Atlas (TCGA) [8] and The Cancer Imaging Archive (TCIA) [9], etc., covering
medical images of the lung, brain, breast, prostate and other organs.

1.3 Segmentation of the Region of

Interest from the Image
Image segmentation directly determines the accuracy of the extracted feature data.
Image segmentation methods can be classified as manual, semi-automatic and fully
automatic methods. The first two segmentation methods are applied widely, and the last
method is still in the development stage. Manual segmentation has the advantage of
high accuracy for objects with clear boundaries and is also suitable for irregular objects
with blurred boundaries but has the disadvantage of being time-consuming and ineffi-
cient. This disadvantage does not meet the requirement of high-throughput data extrac-
tion from mass datasets, needed for a Big-Data radiomics approach. Semi-automatic
and automatic segmentation are able to provide high-throughput data processing and
have a higher reproducibility and consistency. Semi-automatic segmentation refers to
the process whereby automatic segmentation is aided by manual interventions at dif-
ferent stages of segmentation. The semiautomatic segmentation algorithms include
region-growing methods, volumetric CT-based segmentation (e.g., 3D-Slicer) [10], graph
cuts-based method [11] and active contours (snake) algorithms [12]. And fully automatic
segmentation refers to the process whereby segment boundaries of objects are assigned
fully automatically by software. The automatic segmentation applications include white
matter hyperintensities automated segmentation algorithm [13], brain tumor image
analysis [14] and automatic MRI bone segmentation [15], etc.

1.4 High-Throughput Feature Extraction

Many authors have classified radiomic features in different ways [16, 17]. According
to Aerts [18], radiomic features can be grouped into four groups: (1) first-order statistics;
(2) shape- and size-based features; (3) second-order (long range) features; and (4) wavelet
processing-based features. Here, we follow the groups classified by Aerts et al. [18].

1.4.1 First-Order Features with Image Statistics

First-order statistics features (formulas in Appendix A.1), also called image inten-
sity features, describe information of an image related to the pixel values distribution
in gray level [17]. Pixel values may be actual intensities on image or estimated.
First-order statistics features are not texture features because they ignore the spa-
tial interaction between image pixels, whereas the texture features calculation consid-
ers the relationships between neighborhood pixels at a distance. Different first-order
4 Big Data in Multimodal Medical Imaging

statistics features have been widely applied in estimating tumor heterogeneity. In a

PET/CT study [19] evaluating changes in the heterogeneity of conventional parame-
ters and first-order statistics of rectal tumors on PET/CT before and after treatment,
Bundschuh et al. found that coefficient of variation (COV, a first-order feature) pre-
dicted response more accurately than conventional parameters. In a dynamic contrast-
enhanced magnetic resonance imaging (DCE-MRI) study [20], first-order statistics
features on images at different time points were evaluated on lung tumors in terms of
temporal change, optimal time of imaging for analysis and prognostic potential. The
results showed that standard deviation and entropy plateaued at 30–60 seconds after
contrast administration, but skewness and kurtosis decreased very fast. Standard
deviation and entropy showed high predictive ability of survival on the univariate Cox
regression model, and entropy showed significant predictive ability of 2 years of pro-
gression-free survival on the multivariate Cox regression model.

1.4.2 Shape- and Size-Based Features

Shape and size of region of interest have always been used by radiologists to predict
tumor phenotypes. Shape- and size-based features (formulas in Appendix A.2) are sig-
nificant in some specific types of cancer. In one FDG-PET study, Naqa et al. [21] evalu-
ated shape and texture features from PET images to predict cervix cancer and head
and neck cancer patients’ responses to treatment. The results showed that shape-based
metrics had a higher failure risk related to the categorical prediction of treatment in
head and neck cases than that from commonly used standardized uptake value (SUV)
descriptive statistics. And the texture features are more relevant in the cervix cancer.
Shape- and size-based features are also used to quantify the spatial changes of patients
under Parkinson’s disease on PET images and to classify patients between healthy
controls and Parkinson’s disease [22]. In another PET/CT study [23], the shape-based
feature value of squamous cell lung carcinoma (SQCLC) patients who experienced
recurrence was significantly higher than the rest of the patients.

1.4.3 Second-Order Features
Second-order features indicate long range interaction of pixel or voxel values in
an image. These include gray level cooccurrence matrix (GLCM), gray level size zone
matrix (GLSZM) and gray level run length matrix (GLRLM). As one of the important
second-order features, the mathematical meaning and application of GLCM are dis-
cussed in this section.
The GLCM (formulas in Appendix A.3) is based on the second-order statistical com-
bined conditional probability density of the image that considers the spatial relationship
of pixels. Haralick et al. [24] first proposed this matrix and 14 corresponding statistics
and showed those statistics reflect the information regarding the relative position of the
various gray levels over the images. Although GLCM features have a good ability to dis-
tinguish different textures, GLCM computation is time consuming and computationally
expensive because it needs to generate GLCM first and then extract 14 texture features.
Therefore, many researchers attempt to improve the computation time in several direc-
tions. Ulaby et al. [25, 26] reported that many features in GLCM are correlated with
each other and selected only four independent texture features, namely, energy, entropy,
contrast and correlation. In another study [27], the results show that two parameters,
 Big-Data Radiomic Features in Medical Imaging 5

energy and contrast, are considered to be the most efficient for discriminating different
textures among the most relevant of 14 originally proposed parameters. In another study,
to simplify the computation of GLCM, Bo et al. [28] proved that GLCM tends to a con-
stant for different distances and angles when the distance is large enough. As one of the
improvements on GLCM, gray level cooccurrence linked list (GLCLL) [29], which stores
only the nonzero cooccurring probabilities, has better computational speed [30].
Many researchers have utilized GLCM features to analyze multimodal medical
images. In one study, Hatt et al. [31] investigated the association between the meta-
bolically active tumor volume (MATV) of esophageal and non-small cell lung cancer
(NSCLC) through texture features generated by GLCM. In an ultrasound study [32],
the authors investigated the features’ differences between the normal and postradio-
therapy parotid glands by extracting eight GLCM features at distance of 4 pixels and
all angles in two dimensional images. The result suggested that significant differences
in GLCM features were observed between the normal and postradiotherapy parotid
glands. Petkovska et al. [33] reported that the GLCM features extracted from contrast-
enhanced CT have a better identification ability to differentiate malignant from benign
nodules than visual inspection by three experienced radiologists.

1.4.4 Wavelet-Based Features
Wavelets are small oscillatory waveforms which can be stretched into different sizes
and convolved against a signal for extracting their coefficients. The coefficients are, thus,
scale and translation invariant. Wavelet coefficients decompose input signal into differ-
ent components, and then study each component with a resolution matched to its scale.
The wavelet transform essentially splits the signal into a low pass sub-band (abbreviation
notation is “L”, also called approximation level) and a high pass sub-band (abbreviation
notation is “H”, also called detail level) at each scale. A wavelet transformed-2D image
is shown in Figure 1.2 as an example. The wavelet transform-based method processes
an image at multiple levels of resolution and returns a series of gray level edge maps at
different resolutions. After the generation of different level of coefficients (that are 2D
images themselves of varying resolutions) by wavelet transform, all the radiomic fea-
tures described before, such as first-order features, shape- and size-based features and
second-order features, can be extracted on the high pass and low pass sub-band images
again. These could generate thousands of radiomic features [34]. Many studies have used
wavelet-based features to reveal image features of medical images. Yang et al. [35, 36]
used Go ́mes et al.’s informative classifying features and developed a feature extraction
scheme using wavelet transform that further improved the diagnostic performance.

1.4.5 Feature Extraction Software

To extract the radiomic features listed above, many software packages are avail-
able. In the supplementary of Hatt et al.’s review paper on 3D-Slicer [17], many pieces
of feature extraction software were listed. HeterogeneityCAD, a module in 3D-Slicer,
is a radiomics extraction graphical user interface toolbox to quantify the radiomic
features of a region under a segmentation mask. Pyradiomics [37] is an open-source
Python script library for quantifying different radiomics. Wong et al. also provided a
list of many open-source software programs for feature extraction on different modality
images for head and neck cancer [38].
6 Big Data in Multimodal Medical Imaging

FIGURE 1.2: Wavelet transformed images, upper-left is original image and right are
three different levels of wavelet coefficients. And horizontal direction goes along high-
pass filtering, and vertical direction goes along low-pass filtering.

1.5 Clinical Predictive Modeling with Machine Learning

The use of a large number of radiomic features for classification and prediction under
limited samples not only takes a long time to compute but may not be optimal. After a
large number of high-throughput imaging features are extracted, the feature selection
method is used to obtain the best-performing feature sets which are then used in turn
as input to a machine learning algorithm or a statistical modeling algorithm to classify
one or more target diagnostic variables. In this section, we will discuss methods for
selecting features and then, for classification used in radiomics recently.

1.5.1 Feature Selection Methods

The feature selection process aims to find the most informative features among all
extracted radiomic features. To guide the practical application, selected features need
to have good stability and strong discrimination ability.
In much research, radiomic features are reported to be unstable within differ-
ent imaging methods or at different longitudinal time points [39]. Feature stability
in radiomics means feature consistency of each individual in the same cohort cluster.
 Big-Data Radiomic Features in Medical Imaging 7

Features with better stability showed higher prognostic performance, and that phenom-
enon may be due to reduced noise in the stable features. In one study, Aerts et al. [18]
selected top features in four feature groups out of 100 most stable features which were
determined by two testing datasets. According to the stability rank of test and retest
CT scan dataset, they selected the four most significant radiomic features which were
statistical energy, shape compactness, gray level nonuniformity in the original image
and gray level nonuniformity in HLH wavelets filtered images. Their results showed
that radiomic features have prognostic meaning in both lung and head and neck cancer.
Leijenaar et al. [40] assessed every extracted feature between test-retest and inter-
observer stability of features by the intraclass correlation coefficient (ICC), which is
calculated by Kruskal–Wallis one-way analysis of variance (ANOVA) method and pro-
vides an indication of feature measurements reliability. Similarity between test-retest
and interobserver stability rankings of features was evaluated by Spearman’s rank
correlation coefficient.
In addition to being useful in improving the predictive performance of machine learn-
ing algorithms, feature selection also reduces the dimensionality of radiomics feature-
space and, thus, improves computational efficiency. Feature selection methods can be
classified as univariate and multivariate feature selection methods [41]. Some research
investigated the role of feature selection methods on predictive models [43]. Parmar et
al. [41, 42, 44] analyzed 14 feature selection methods on the predictive performance of
patient survival and found that the prediction accuracy of radiomic features is influ-
enced by the number of features, feature selection methods and pattern recognition
classifiers. Among 14 feature selection algorithms, Wilcoxon test-based feature selec-
tion method (WLCX), one of the univariate methods, showed the highest performance
in most of the selected features numbers.

1.5.2 Radiomic Features Classifying Outcomes

1.5.2.1 Clustering Algorithms
Clustering algorithms were used to reveal the cluster patterns of radiomic features
without knowing prior patient outcome knowledge. In one recent widely cited publica-
tion, Aerts et al. [18] applied a hierarchical clustering method to reveal the clusters of
patients in similar expression patterns among 440 features. By comparing with the
clinical parameters, the primary tumor stage, overall stage and histology of tumors
were observed to have a significant association with the cluster. In another study
[42], Parmar used a consensus clustering method and showed the correlation between
radiomic features and clinical parameters of two types of cancer. The term “consensus
clustering algorithm” refers to finding a single (consensus) clustering which is the best
fit with input dataset from the existing clusters. After that overlap, features between
the feature clusters of lung and head and neck cancer cohorts were assessed by the
Jaccard index matrix. In another study, a hierarchical clustering algorithm was used
to identify nonredundant features, and the method is reproducible across multiple CT
machines [45].

1.5.2.2 Regression Model
Linear regression is a model that provides the predictive relationship between
a dependent variable and one or more independent variables. The linear regression
model has been used to analyze the temporal relationship between undergoing therapy
8 Big Data in Multimodal Medical Imaging

and radiomic features. For example, Yang et al. [46] compared tumor texture features
of fluorodeoxyglucose (FDG) heterogeneity with tumor SUV measures on patients’
PET images and classified their chemoradiotherapy responses at different time points.
Temporal changes were analyzed by the linear regression model. The temporal behav-
iors in the early phase of treatment show that texture features are a better predictor
for the outcome of therapeutic intervention than SUV measures. In another study
[47], a linear regression model was also used to assess the relationship between the
radiation dose and radiomic features. Li et al. [48] demonstrated significant associa-
tions between radiomic features and multigene assay recurrence scores using multiple
linear regression models.
Unlike linear regression where a dependent variable is normally continuous, a logis-
tic regression model is used when a dependent variable has only a limited number of
possible or categorical values. In King et al.’s study [49], radiomic features were corre-
lated for tumor local failure and control at 2 years. A logistic regression model was used
to predict local failure by univariate and multivariate analyses of the apparent diffu-
sion coefficients parameters, tumor stage and tumor volume. The poorly responding
tumors showed a significant increase in skewness and kurtosis values than that shown
by well-controlled tumors. El et al. [50] employed radiomic features to predict treatment
outcomes in patients with cervical cancers and head and neck cancers. Association
between different extracted radiomic features and nine patients’ overall survival was
investigated by a logistic regression model. The result showed that shape-based fea-
tures had the highest categorical predictive capability of failure-risk, while commonly
used SUV descriptive statistics had the lowest predictive ability.
The Cox regression model, referred to as the Cox proportional hazards regression
model, was applied mainly for the prognosis of survival-time outcomes of tumors and
other chronic diseases on one or more predictors. Cook et al. [51] measured tumor
radiomic features like coarseness, contrast, busyness and complexity which were derived
from a PET dataset of 53 patients with NSCLC treated with chemoradiotherapy. Cox
regression was used to examine the effects of the radiomic features on the survival
outcomes. Therapy responders showed lower coarseness and higher contrast and busy-
ness than non-responders. However, none of the SUV parameters predicted treatment
response accurately. Those results indicate that coarseness, busyness, and contrast are
more closely associated with differentiating between responders and non-responders to
chemoradiotherapy than SUV measures, and coarseness was an independent predictor
of overall patient survival. In another study [52], the Cox model was trained to evaluate
radiomic features’ capability to predict distant metastasis.

1.5.2.3 Support Vector Machine

Support vector machine (SVM) is another machine learning algorithm for classi-
fication problems. SVM transforms images by different types of kernel functions and
then finds an optimal boundary between the predictive outputs. SVM is also applied
to patient outcome prediction. Some studies have used SVM for predicting patients’
response to different therapies. Zhang et al. [53] studied 20 esophageal cancer patients’
18F-FDG PET/CT scans both before and after trimodality therapy (chemoradiotherapy

plus surgery). SVM and logistic regression (LR) models were used to predict pathologic
tumor response to treatment. Compared to different models and feature groups, they
found that SVM models constructed by radiomic features combined with conventional
SUV measures and clinical parameters significantly improved the pathologic response
 Big-Data Radiomic Features in Medical Imaging 9

prediction. Some studies [54, 55] have shown that many image features are able to dif-
ferentiate tumor stages using different medical imaging modalities. In one study, Mu
et al. [56] classified cervical cancer patients into early stage and advanced stage using
54 PET-based image features. One of the radiomic features, run percentage (RP), which
is selected by the SVM classifier, was the most discriminative index with higher accu-
racy and ability to differentiate early stage and advanced stage with high accuracy. In
another tumor stage study, Dong et al. [57] assessed radiomic features on PET images
of 40 patients with esophageal squamous cell carcinoma. Correlations were observed
between SUVmax, GLCM-entropy and GLCM-energy on the one hand and T and N
stages of patients on the other. Xu et al. [58] developed a computer-aided diagnosis
(CAD) system to differentiate malignant and benign bone and soft-tissue lesions on
FDG PET/CT images. After selection of a subset of the most optimal radiomic features
parameters, an SVM classifier was used to automatically differentiate different bone
tumor tissues. The result showed that the CAD method with the combined PET and
CT optimal radiomic features has a better ability in differential diagnosis than that
by each of PET or CT texture analysis alone. Compared with the histological diagnosis
of the lesions, the diagnostic method based on classification results of their radiomic
features achieved a relatively good accuracy.

1.6 Big Data and Machine Learning

Machine Learning: The power of machine learning, both supervised and unsuper-
vised, may be best harnessed when the training sample set is sufficiently large. Due
to privacy and expense, obtaining a large number of human datasets is very challeng-
ing in medicine. However, across different laboratories and clinical facilities, data are
being constantly gathered. If one can share these datasets, the use of machine learn-
ing may be as powerful as it is in many of its other application areas, such as com-
puter vision, natural language processing, etc. This is the direction where we see that
medical imaging research is making some pioneering steps [7, 8, 9]. Most recently,
DeepLesion, released by NIH, contains nearly 10,600 publicly available CT scans
to support the development of machine learning algorithms for medical applications
[59]. Natural challenges, of course, are many, e.g., legally sharing data, standardizing
data for uniform utilization by machine learning algorithms (as we have discussed
in Section 1.2) and designing appropriate research questions over varieties of avail-
able data. All these challenges relate to the classical definition of Big Data, namely,
volume, veracity and velocity, which pertain to the large size of data, diversity of data
and rapid change of data, respectively. In the context of medical imaging these three
characteristics respectively pose their own problems: (1) managing large data size over
many imaging studies; (2) homogenizing data (to be accessed by individual algorithms
or data processing procedures) over multiple studies acquired for different purposes
and hypotheses, over different machines, possibly with multiple modalities involved;
and (3) continuously increasing the size of the image database as new data are con-
stantly added by contributors. Computer science is actively seeking solutions to the
large data size problem. For example, new languages are being designed for Big Data
(e.g., PlinyCompute [60]) and new architecture for deep learning as model paralleliza-
tion [61]. We have discussed in this article how the radiology community is addressing
10 Big Data in Multimodal Medical Imaging

the veracity of diversely acquired data. Finally, we believe these two solutions will
have to adjust to the fact that the data will continuously grow and no static model for a
fixed dataset will be sufficient. However, that is for the near future. In this context, the
newly emerging field of topological data analysis (TDA) [62] may help in understanding
the nature of a dataset. As an unsupervised machine learning model, this technique
studies the topology of data points (images) in the underlying feature space. Instead of
clustering the data points, TDA rather investigates more complex structures like con-
nectivity between data. Such topological understanding may provide much better guid-
ance to the supervised machine learning algorithms [63] or may even be embedded in
the latter for a robust algorithm [64] to address the Big Data challenges.
Computing Hardware: Deep learning algorithms’ miraculous success has not only
dazzled all machine learning applications but also medical imaging communities.
However, these algorithms are not just data hungry but also computing-resources hun-
gry. Most of these algorithms need multiple high-performance computing nodes such
as computing-focused graphics processing units (GPUs) with sufficient localized mem-
ory closer to computational unit. For example, Nvidia (Santa Clara, CA) provides a
dedicated web page for medical imaging applications on their GPU platforms [65] and
announced Project Clara in March 2018 [66]. Specialized hardware is being developed
for deep learning algorithms by many other vendors as well.
Image Data Mart: Another sector where we see that Big Data in medical imaging
needs to grow is toward Data Mart development. There are many dimensions of such
image databases, e.g., imaging modalities, disease models, metadata involved with
study-questions or protocols, etc. Currently, many such databases are being made
available, e.g., the recent release of 100,000 chest X-ray images for 30,000 patients
by NIH for lung cancer research [67]. The project will release torso images of the
same type for research on abdominal diseases according to the announcement [68].
Similar other lung cancer datasets are available on the web. Eventually such large
data release projects need to be consolidated, when they are too similar, or easily
referred from one to another to pose higher-level research questions, e.g., for connect-
ing genotype and phenotype of a disease. The National Cancer Informatics Program
provides a step toward such an integrated view of data. It took more than 30 years for
the genomics databases to provide a disciplined view to researchers for data access.
We expect that experience will be harnessed to allow comprehensive access to imag-
ing databases for scientists.

1.7 Conclusions
The machine learning models built upon unique and important radiomic features
provide a new way not only to study tumor heterogeneity but also to quantitatively
analyze the changes in micro-level gene or protein patterns that are hidden behind
medical images. Many studies have applied radiomic features and reached promising
results in precision medicine. As a new research field, radiomics still needs further
exploration. With the standardization of medical imaging data, and the rapid develop-
ment of various methods of image segmentation, feature extraction, feature selection
and machine learning algorithms, the application of radiomic features will have far-
reaching effects and lead to tremendous changes in radiology. As larger datasets are
Another random document with
no related content on Scribd:
 10 qu’il volsissent priier au conte Derbi qu’il les
volsist prendre à merci, salve leurs corps et leurs
biens, et en avant il se metteroient en l’obeissance
dou roy d’Engleterre. Li contes de Pennebruch et li
contes de Kenfort respondirent qu’il en parleroient
15 volentiers; et puis demandèrent il où li contes de
[Lille] et li aultre baron estoient. Il respondirent:
«Certainnement nous ne savons, car il cargièrent et
toursèrent très le mienuit tout le leur et se partirent,
mès point ne nous disent quel part il se trairoient.»
20 Sus ces parolles, se partirent li doi conte dessus
nommet, et vinrent au conte Derbi, qui n’estoit mies
loing de là, et li moustrèrent tout ce que les gens de
Bregerach voloient faire. Li dis contes Derbi, qui fu
moult nobles et très gentilz de coer, respondi: «Qui
25 merci prie, merci doit avoir. Dittes leur qu’il oevrent
leur ville et nous laissent ens: nous les assegurons
de nous et des nostres.» Adonc retournèrent li doi
signeur dessus dit, et recordèrent à chiaus de Bregerach
tout ce que vous avés oy: dont il furent moult
30 joiant, quant il veirent qu’il pooient venir à pais. Si
vinrent ou le place et sonnèrent les sains, et se
[53] assamblèrent tout, hommes et femmes, et fisent ouvrir
leurs portes, et vinrent à grant pourcession et moult
humlement contre le conte Derbi et ses gens, et le
menèrent à le grant eglise. Et là li jurèrent il feaulté
5 et hommage et le recogneurent à signeur, ou nom
dou roy d’Engleterre, par le vertu de le procuration
qu’il en portoit. Ensi eut en ce temps li contes Derbi
le bonne ville de Bregerach, qui se tint toutdis depuis
englesce. Or, parlerons nous des signeurs de Gascongne
10 qui estoient retrait en le ville et ou chastiel
de le Riolle, et vous conterons comment il se
maintinrent.

§ 212. Celle propre nuitie et le journée ensievant

 que li contes de [Lille] et li baron et chevalier, qui
15 avoecques lui estoient, furent venu en le Riolle, il
regardèrent et avisèrent li un par l’autre qu’il se departiroient
et se trairoient ens ès garnisons, et guerrieroient
par forterèces, et metteroient sus les camps
entre quatre cens ou cinq cens combatans dont il
20 feroient frontière: si en seroient chief et meneur li
contes de Comignes et li viscontes de Quarmaing.
Adonc se departirent il. Si se traist li contes de
Pieregorth en [Perigueux[292]], li seneschaus de Thoulouse
à Montalben, li viscontes de Villemur à Auberoce,
25 messires Bertrans des Prés à Pellagrue, messires Phelippes
de Dyon à Montagrée, li sires de Montbrandon
à Maudurant, Ernaus de Dyon à [Lamongis[293]],
Bobers de Malemort à Byaumont en [Lillois], messires
[54] Charles de Poitiers à Pennes en Aginois, et ensi
les chevaliers de garnison en garnison. Si se departirent
tout li un de l’autre, et li contes de [Lille] demora
en la Riolle. Et fist remparer et rabillier le
5 ville et le forterèce, telement que elle n’avoit garde
d’assaut que on y fesist sus un mois ne deux. Or,
retourrons nous au conte Derbi qui estoit en
Bregerach.

§ 213. Quant li contes Derbi eut pris le possession

10 et le saisine de le ville de Bregerach, et il s’i fu
rafrescis par deux jours, il demanda au senescal de
Bourdiaus quel part il se [trairoit[294]], car mies ne voloit
sejourner. Li seneschaus respondi que ce seroit
bon d’aler devers Pieregort et en le haute Gascongne.
15 Dont fist li dis contes Derbi ordonner toutes
ses besongnes et traire au chemin par devers [Perigueux[295]],
et laissa en Bregerach un chevalier à chapitainne,
qui s’appelloit messires Jehans de la Souce.
Ensi que li Englès chevauçoient, il trouvèrent en
20 leur chemin un chastiel qui s’appelle Lango, dont li
 vigiers de Thoulouse estoit chapitaine, une moult
aperte armeure de fier. Il s’arrestèrent là, et disent
qu’il ne lairoient pas che chastiel derrière. Si le commença
li bataille des mareschaus à assallir, et y furent
25 un jour tout entier. Et là eut mervilleusement dur
et fort assaut, car li Englès assalloient de grant volenté,
et cil dou fort se deffendoient moult vassaument.
Nequedent, ce premier jour il n’i fisent riens.
[55] A l’endemain, priès que toute li hos fu devant. Et
recommencièrent à assallir telement et par si forte
manière, avoech ce que on avoit jetté ens ès fossés
grant fuison de bois et de velourdes par quoi on
5 pooit bien aler jusques as murs sans dangier, que cil
de dedens se commencièrent à esbahir. Adonc leur
fu demandé de messire Franke de Halle, se il se renderoient,
et que il y poroient bien tant mettre qu’il
n’i venroient point à temps. Il requisent à avoir conseil
10 de respondre. Ce leur fu acordé. Il se consillièrent;
et me samble qu’il se partirent de le forterèce
et riens n’enportèrent, et s’en alèrent devers Montsach
qui se tenoit françoise. Ensi eurent li Englès le
chastiel de Lango. Si y establi li contes Derbi un escuier
15 à chapitainne, qui s’appelloit Aymon Lyon, et
laissa laiens avoecques lui jusques à trente arciers.
Si se partirent de Lango, et cheminèrent devers une
ville qui s’appelle le Lach.

§ 214. Quant cil de le ville dou Lach sentirent

20 les Englès venir si efforciement, et qu’il avoient pris
Bregerach et le chastiel de Lango, si en furent si
effraet qu’il n’eurent mies conseil d’yaus tenir. Et
s’en vinrent au devant dou conte Derbi, et li aportèrent
les clefs de le ville, et le recogneurent à
25 signeur, ou nom dou roy d’Engleterre. Li contes
Derbi prist le feaulté d’yaus, et puis passa oultre, et
s’en vint à Maudurant, et le gaegna d’assaut sus
 chiaus dou pays. Et y fu pris uns chevaliers dou pays,
qui [s’appelloit[296]] li sires de Montbrandon. Si laissa li
[56] dis contes Derbi gens d’armes dedens le forterèce; et
puis passèrent oultre, et vinrent devant le chastiel de
[Lamongis[297]]. Si l’assallirent, et le prisent d’assaut tantost,
et le chevalier qui estoit dedens, et l’envoiièrent
5 tenir prison à Bourdiaus. Puis chevaucièrent devers
Pinac et le conquisent, et en apriès le ville et le
chastiel de Laliene, et s’i rafreschirent par trois jours.
Au quatrime, il s’en partirent et vinrent à Forsach,
et le gaegnièrent assés legierement, et en apriès le
10 tour de Prudaire. Et puis chevaucièrent devers une
bonne ville et grosse, que on appelle Byaumont en
[Lillois], qui se tenoit liegement dou conte de [Lille].
Si furent li Englès trois jours par devant, et y fisent
tamaint grant assaut, car elle estoit bien pourveue
15 de gens d’armes et d’artillerie, qui le deffendirent
tant qu’il peurent durer. Finablement elle fu prise,
et y eut grant occision de chiaus qui dedens furent
trouvet. Et le rafreschi li contes Derbi de nouvelles
gens d’armes, et puis chevauça oultre, et vint devant
20 Montagrée. Si le prist ossi d’assaut et le chevalier
qui estoit dedens, et l’envoia tenir prison à
Bourdiaus.
Et ne cessèrent li Englès de chevaucier; si vinrent
devant [Lille], la souverainne ville dou conte, dont
25 messires Phelippes de Dyon et messires Ernaulz de
Dyon estoient chapitainne. Quant li contes Derbi et
ses gens furent venu par devant, si l’avironnèrent et
regardèrent que elle estoit bien prendable. Si fisent
traire leurs archiers avant, et approcier jusques as
30 barrières. Cil commencièrent à traire si fortement
[57] que nulz de chiaus de le ville n’osoient apparoir
pour deffendre. Et conquisent li Englès ce premier
jour les bailles et tout jusques à le porte, et sus le
soir il se retraisent. Quant ce vint au matin, de rechief
 5 il commencièrent à assallir fortement et despertement
en pluiseurs lieus, et ensonniièrent si
chiaus de le ville qu’il ne savoient au quel entendre.
Li bourgois de le ville, qui doubtèrent le leur à
perdre, leurs biens, leurs femmes et leurs enfans,
10 regardèrent que à le longe il ne se poroient tenir. Si
priièrent as deux chevaliers qui là estoient, qu’il
trettiassent à ces signeurs d’Engleterre, par quoi il
demorassent à pais et que li leurs fust sauvés. Li
chevalier, qui assés bien veoient le peril où il estoient,
15 s’i acordèrent assés legierement. Et envoiièrent un
hiraut de par yaus au conte Derbi, pour avoir respit
un jour tant seulement et parlement de composition.
Li hiraus vint devers le dit conte qu’il trouva sus les
camps assés priès de le ville, et li remoustra ce pour
20 quoi il estoit là envoiiés. Li contes s’i acorda et fist
retraire ses gens, et s’en vint jusques as barrières
parlementer à chiaus de le ville, dalés lui le baron de
Stanfort et le signeur de Mauni. Là furent il en grant
parlement ensamble, car li contes Derbi voloit qu’il
25 se rendesissent simplement, et il ne l’euissent jamais
fait. Toutes fois, accors se porta que cil de [Lille] se
metteroient en l’obeissance dou roy d’Engleterre. Et
de ce envoieroient il douze de leurs plus honnourables
hommes en le bonne cité de Bourdiaus, en nom
30 de crant. Et sur ce li chevalier et li escuier françois,
qui là estoient, se pooient partir et traire quel part
qu’il voloient. Ensi eut li contes Derbi en ce temps
[58] le ville de [Lille] en Gascongne; et se partirent sus
son saufconduit les gens d’armes qui dedens estoient,
et s’en alèrent devers le Riolle.

§ 215. Apriès le conquès de [Lille], et que li contes

5 Derbi y eut laissiés gens d’armes et arciers de par
lui, et envoiiés douze bourgois de la ville en ostagerie,
pour plus grant seurté, en le cité de Bourdiaus, il
 chevauca oultre et vint devant Bonival. Là eut grant
assaut et dur, et pluiseurs hommes bleciés dedens et
10 dehors. Finablement, li Englès le prisent et le misent
à merci, et le rafreschirent de gens d’armes et de
chapitainne. Et puis chevaucièrent oultre, et entrèrent
en le conté de Pieregorth, et passèrent devant
Bourdille, mais point n’i assallirent, car il veirent
15 bien qu’il euissent perdu leur painne. Si esploitièrent
tant qu’il vinrent jusques à [Perigueux[298]]. Par dedens
estoient li contes de Pieregorth et messires Rogiers
de Pieregorth ses oncles, et li sires de Duras et bien
six vingt chevaliers et escuiers dou pays, qui tout
20 s’estoient là recueilliet sus le fiance dou fort liu, et
ossi li uns pour l’autre. Quant li contes Derbi et se
route parfurent venu devant, il avisèrent et imaginèrent
moult bien comment, ne par quel voie à leur
avantage il le poroient assallir. Si le veirent forte
25 durement: si ques, tout, consideret, il n’eurent mies
conseil d’emploiier y leurs gens, mais se retraisent
arrière sans riens faire. Et s’en vinrent logier à deux
liewes de là, sus une petite rivière, pour venir devant
le chastiel de Pellagrue.
[59] Ces gens d’armes, qui estoient dedens le chastiel
de Perigueux, parlèrent ce soir ensamble, et disent
ensi: «Cil Englès nous sont venu veoir et aviser de
priès, et puis se sont parti sans riens faire. Ce seroit
5 bon que à nuit nous les alissions resvillier, car il ne
sont pas logiet trop loing de ci.» Tout s’acordèrent
à ceste oppinion, et issirent de Perigueux environ
mienuit bien deux cens lances, montés sus fleur de
chevaus; et chevaucièrent radement, et furent devant
10 le jour ou logeis des Englès. Si se ferirent dedens
baudement; si en occirent et mehagnièrent grant
fuison. Et entrèrent ou logeis le conte de Kenforth,
et le trouvèrent qu’il s’armoit. Si fu assallis vistement
et pris par force; aultrement il euist estet mors, et
 15 ne sçai trois ou quatre chevaliers de son hostel. Puis
se retraisent li Gascon sagement, ançois que li hos
fust trop estourmis, et prisent le chemin de [Perigueux].
Si leur fu bien besoins qu’il trouvassent les
portes ouvertes et apparillies, car il furent sievoit
20 caudement et rebouté durement dedens leurs barrières.
Mais si tretost que li Gascon furent en leur
garde, il descendirent de leurs chevaus et prisent
leurs glaves, et s’en vinrent franchement combatre
main à main as Englès; et tinrent leur pas souffisamment,
25 et fisent tant qu’il n’i eurent point de damage.
Adonc se retraisent li Englès tout merancolieus de
ce que li contes de Kenfort estoit pris, et vinrent à
leur logeis; si se deslogièrent bien matin, et cheminèrent
devers Pellagrue.

30 § 216. Tant chevaucièrent li Englès qu’il vinrent

devant le chastiel de Pellagrue. Si l’environnèrent
[60] de tous lés et le commencièrent à assallir fortement,
et cil de dedens à yaus deffendre comme gens de
grant volenté, car il avoient un bon chevalier à chapitainne
que on appelloit monsigneur Bertran des
5 Prés. Et furent li Englès six jours devant Pellagrue,
et y fisent tamaint assaut. Là en dedens furent tretties
les delivrances dou conte de Renfort et de ses
compagnons, en escange dou visconte de Boskentin,
dou visconte de Chastielbon, dou signeur de [Lescun[299]]
10 et dou signeur de Chastielnuef, parmi tant encores
que la terre dou conte de Pieregorth demorroit trois
ans en pais. Mais bien se pooient armer li chevalier
et li escuier de celle terre, sans fourfait; mès on ne
pooit prendre, ardoir ne exillier, piller ne rober
15 nulle cose, le terme durant, en la ditte conté. Ensi
revinrent li contes de Kenfort et tout li prisonnier
englès qui avoient estet pris de chiaus de Pieregorth.
Et ossi li chevalier de Gascongne furent delivret
 parmi le composition dessus ditte.
20 Et se partirent li Englès de devant Pellagrue, car
la terre est tenue dou conte de Pieregort, et chevaucièrent
devers Auberoce, qui est uns biaus chastiaus
et fors de l’archevesquié de Thoulouse. Si trestost
que li Englès furent venu devant Auberoce, il s’i logièrent
25 ossi faiticement que donc qu’il y vosissent
demorer et sejourner une saison. Et envoiièrent segnefiier
et dire à ceulz qui dedens estoient, qu’il se
rendesissent et mesissent en l’obeissance dou roy
d’Engleterre; ou aultrement, s’il estoient pris par
[61] force, il seroient tout mort sans merci. Cil de le ville
et dou chastiel d’Auberoce doubtèrent leurs biens et
leurs corps à perdre, et veirent qu’il ne leur apparoit
nul confort de nul costé. Si se rendirent, salve leurs
5 corps et leurs biens, et se misent en l’obeissance
dou conte Derbi, et le recogneurent à signeur, ou nom
dou roy d’Engleterre, par le vertu d’une procuration
qu’il en avoit. Adonc s’avisa li contes Derbi qu’il se
retrairoit tout bellement devers le cité de Bourdiaus.
10 Si laissa dedens Auberoce en garnison monsigneur
Franke de Halle, monsigneur Alain de Finefroide et
monsigneur Jehan de Lindehalle.
Puis s’en vint li contes à Liebrone, une bonne
ville et grosse, en son chemin de Bourdiaus, à douze
15 liewes d’illuech; et l’assega, et dist bien à tous chiaus
qui oïr le vorrent, qu’il ne s’en partiroit si l’aroit.
Quant cil de Liebrone veirent assegie leur ville, et le
grant effort que li contes Derbi menoit, et comment
tous li pays se rendoit à lui, si alèrent li homme de
20 le ville ensamble en conseil à savoir comment il se
maintenroient. Tout consideret [et yaulx consilliet[300]],
et peset le bien contre le mal, il se rendirent et ouvrirent
leurs portes, ne point ne se fisent assallir ne
heriier. Et jurèrent feaulté et hommage au conte
25 Derbi, ou nom dou roy d’Engleterre, et à demorer
 bons Englès de ce jour en avant. Ensi entra li contes
Derbi dedens Liebrone, et y fu quatre jours, et là
ordonna il de ses gens quel cose il feroient. Si ordonna
tout premierement le conte de Pennebruch
30 et se route à aler à Bregerach, et messire Richart de
[62] Stanfort et messire Estievene de Tonrbi et messire
Alixandre Ansiel et leurs gens à demorer en le ville
de Liebrone. Chil li acordèrent volentiers. Dont se
partirent li contes Derbi, li contes de Kenfort, messires
5 Gautiers de Mauni et li aultre, et chevaucièrent
devers Bourdiaus, et tant fisent qu’il y parvinrent.

§ 217. Au retour que li contes Derbi fist en le

cité de Bourdiaus, fu il liement recueillés et recheus
de toutes gens. Et vuidièrent li clergiés et li bourgois
10 de le ville à grant pourcession contre lui, et li
fisent tout honneur et reverense à leur pooir. Et li
abandonnèrent vivres et pourveances et toutes aultres
coses, à prendre ent à sen aise et volenté. Li
contes les remercia grandement de leur courtoisie
15 et de ce qu’il li offroient. Ensi se tint li contes Derbi
en le cité de Bourdiaus avoecques ses gens; si s’esbatoit
et jewoit entre les bourgois et les dames de
le ville.
Or lairons nous un petit à parler de lui, et parlerons
20 dou conte de [Lille], qui se tenoit en le Riolle,
et savoit bien tout le couvenant des Englès et le conquès
que li contes Derbi avoit fait, et point n’i avoit
peut pourveir de remède. Si entendi li dis contes de
[Lille] que li contes Derbi estoit retrais au sejour en
25 Bourdiaus et avoit espars ses gens et romput se chevaucie,
et n’estoit mies apparant que de le saison il
en fesist plus. Si s’avisa li dis contes qu’il feroit une
semonse et un mandement especial de gens d’armes,
et iroit mettre le siège devant Auberoce. Ensi qu’il
30 l’avisa, il le fist. Si escrisi devers le conte de Pieregort,
 celi de Quarmaing, celi de Commignes, celi de
[63] Brunikiel, celi de Villemur et devers tous les barons
de Gascongne, qui François se tenoient, qu’il fuissent
sus un jour qu’il leur assigna, devant Auberoce, car
il y voloit aler et mettre le siège. Li dessus dit conte,
5 visconte et baron de Gascongne obeirent à lui, car
il estoit comme rois ens ès marces de Gascongne.
Et assamblèrent leurs gens et leurs hommes, et furent
tout appareilliet au jour qui assignés y fu. Et vinrent
devant Auberoce telement que li chevalier dessus
10 nommé, qui le gardoient, ne s’en donnèrent de garde;
si se veirent assegiet de tous costés. Ensi que gens
de bon couvenant et de grant arroy, il ne furent de
riens esbahi, mès entendirent à leurs gardes et à
leurs deffenses. Li contes de [Lille] et li aultre baron,
15 qui là estoient venu moult poissamment, se logièrent
tout à l’environ, telement que nulz ne pooit entre[r],
ne issir en le garnison, qu’il ne fust aperceus. Et envoiièrent
querre quatre grans engiens à Thoulouse,
et les fisent là achariier et puis drecier devant le forterèce.
20 Et n’assalloient li François d’autre cose, fors
de ces engiens, qui nuit et jour jettoient pières de fais
ou chastiel. Che les esbahissoit plus c’autre cose, car
dedens six jours il desrompirent le plus grant partie
des combles des tours. Et ne s’osoient li chevalier
25 ne cil dou chastiel tenir, fors ens ès cambres votées,
par terre. Et estoit li intentions de chiaus de l’host
qu’il les occiroient là dedens, ou il se renderoient
simplement. Bien estoient venues les nouvelles à
Bourdiaus au conte Derbi et à monsigneur Gautier
30 de Mauni, que leur compagnon estoient assegiet dedens
Auberoce; mais point ne savoient qu’il fuissent
si apressé, ne si constraint qu’il estoient.

[64] § 218. Quant messires Franke de Halle et messires

Alain de Finefroide et messires Jehans de Lindehalle
 veirent le oppression que li François leur faisoient,
et si ne leur apparoit confors ne ayde de nul
5 costé, si se commencièrent à esbahir, et se consillièrent
entre yaus comment il se poroient maintenir.
«Il ne poet estre, disent il, que se li contes Derbi
savoit le dangier où cil François nous tiènent, qu’il
ne nous secourust, à quel meschief que ce fust. Si
10 seroit bon que nous li feissiens savoir, mès que nous
peuissions trouver messag[e].» Adonc demandèrent il
entre leurs varlès se il en y avoit nul qui volsist
gaegnier et porter ceste lettre qu’il avoient escript, à
Bourdiaus, et baillier au conte Derbi. Lors s’avança
15 uns varlès, et dist qu’il li porteroit bien et volentiers;
et ne le feroit mies tant, pour le convoitise de gaegnier,
que pour yaus delivrer de ce peril. Li chevalier
furent moult liet dou varlet qui s’offroit de faire le
message.
20 Quant ce vint au soir par nuit, li varlés prist la
lettre que li chevalier li baillièrent, qui estoit seellée
de leurs trois seaulz; et li encousirent en ses draps,
et puis le fisent avaler ens ès fossés. Quant il fu au
fons, il monta amont et se mist à voie parmi l’ost,
25 car aultrement ne pooit il passer. Et fu encontrés
dou premier get, et ala oultre, car il sçavoit bien
parler gascon, et nomma un signeur de l’host, et dist
qu’il estoit à lui. On le laissa passer par tant; et cuida
bien estre escapés, mès non fu, car il fu repris au
30 dehors des tentes, d’aultres varlès, qui l’amenèrent
devant le chevalier dou ghet. Là ne peut il trouver
nulle excusance qui riens li vaulsist. Si fu tastés et
[65] exquis, et la lettre trouvée sur lui. Si fu menés en
prison et gardés jusques au matin, que li signeur de
l’host furent tout levet; si furent tantost enfourmé
de le prise dou varlet. Adonc se traisent il tout ensamble
5 en le tente dou conte de [Lille]. Là fu la lettre
leute que li chevalier d’Auberoce envoioient au conte
 Derbi. Si eurent tout grant joie, quant il sceurent de
vérité que li chevalier d’Auberoce et li compagnon
englès qui dedens se tenoient estoient si astraint et
10 qu’il ne se pooient plus tenir: si ques, pour yaus
plus agrever, il prisent le varlet, et li pendirent les
lettres au col, et le misent tout en un mont et en le
fonde d’un engien, et puis le renvoiièrent dedens
Auberoce. Li varlès chei tous mors devant les chevaliers
15 qui là estoient, et qui furent esbahi et desconforté
quant il le veirent: «Ha! disent il, nostre
messagier n’a pas fait son message. Or ne savons
nous mès que viser, ne quel conseil avoir qui nous
vaille.»
20 A ces cops estoient montés à cheval li contes de
Pieregorth et messires Rogiers de Pieregorth ses oncles,
messires Charles de Poitiers, li viscontes de
Quarmaing et li sires de Duras, et passèrent devant
les murs de le forterèce au plus priès qu’il peurent.
25 Si escriièrent à chiaus dedens, et leur disent en gabois:
«Signeur, signeur englès, demandés à vostre
messagier où il trouva le conte Derbi si apparilliet,
quant à nuit se parti de vostre forterèce, et jà est
retournés de son voiage.» Adonc respondi messires
30 Frankes de Halle, qui ne s’en peut astenir, et dist:
«Par foy, signeur, se ceens nous sommes enclos,
nous en isterons bien, quant Diex vorra et li contes
[66] Derbi. Et pleuist à Dieu qu’il seuist en quel parti
nous sommes! Se il le savoit, il n’i aroit si avisé des
vostres qui ne ressongnast à tenir les camps. Et
se vous li volés segnefiier, li uns des nostres se mettera
5 en vostre prison pour rançonner, ensi que on
rançonne un gentil homme.» Dont respondirent li
François: «Nennil, nennil! Les pareçons ne se porteront
mies ensi. Li contes Derbi le sara tout à
temps, quant par nos engiens nous arons abatu rés
10 à rés terre che chastiel, et que vous, pour vos vies
 sauver, vous vos serés rendu simplement.»
—«Certainnement, ce respondi messires Franke de Halle, ce
ne sera jà que ensi nous nos doions rendre, pour
estre tout mort ceens.» Adonc passèrent li chevalier
15 françois oultre, et revinrent à leur logeis. Et li troi
chevalier englès demorèrent dedens Auberoce tout
esbahi, au voir dire; car ces pières d’engien leur
buskoient si grans horions, que ce sambloit effoudres
qui descendist dou ciel, quant elles frapoient contre
20 les murs dou chastiel.

§ 219. Toutes le[s] parolles et les devises et le

couvenant dou messagier, comment il avoit esté pris
devant Auberoce, et l’estat de la lettre, et le neccessité
de chiaus dou chastiel furent sceues et raportées à
25 Bourdiaus au conte Derbi et à monsigneur Gautier
de Mauni, par une leur espie qu’il avoient envoiiet
en l’ost, et qui leur dist: «Mi signeur, à ce que j’ai
pout entendre, se vo chevalier ne sont conforté dedens
trois jours, il seront ou mort ou pris. Et volentiers
30 se renderoient, se on les voloit prendre à merci,
mès il me samble que nennil.» De ces nouvelles ne
[67] furent mies li contes Derbi et messires Gautiers de
Mauni bien joiant, et disent entre yaus: «Ce seroit
lasqueté et villonnie, se nous [laissons[301]] perdre trois
si bons chevaliers que cil sont, qui si franchement
5 se sont tenu dedens Auberoce. Nous irons ceste part
et nous esmouverons tout premièrement, et manderons
au conte de Pennebruch, qui se tient à Bregerach,
qu’il soit dalés nous à cèle heure, et ossi à monsigneur
Richart de Stanfort et à monsigneur Estievene
10 de Tombi qui se tient à Liebrone.»
[Adoncques ly contes Derby se hasta durement, et
envoia ses messages et ses lettres devers le conte de
Pennebruk. Et se parti de Bourdiaux à ce qu’il avoit
de gens, et chevaucha tout couvertement devers Auberoche;
 15 bien avoit qui le menoit et qui congnissoit
le pais. Si vint ly contes Derby à Liebrone[302]] et là
sejourna un jour, attendans le conte de Pennebruch,
et point ne vint. Quant il vei qu’il ne venroit point,
si fu tous courouciés et se mist au chemin, pour le
20 grant desir qu’il avoit de conforter ses chevaliers qui
en Auberoce se tenoient, car bien [sçavoit[303]] qu’il en
avoient grant mestier.
Si issirent de Liebrone li contes [Derbi[304]] li contes
de Renfort, messires Gautiers de Mauni, messires
25 Richars de Stanfort, messires Hues de Hastinghes,
messires Estievenes de Tombi, li sires de Ferrières
et li aultre compagnon. Et chevaucièrent toute
[68] nuit, et vinrent à l’endemain à deux petites liewes
d’Auberoce. Si se boutèrent en un bois, et descendirent
de leurs chevaus; et les alloiièrent as arbres
et as foellies, et les laissièrent pasturer en l’erbe,
5 toutdis attendans le conte de Pennebruch. Et furent
là toute la matinée, et jusques à nonne. Si s’esmervilloient
trop durement de ce qu’il n’ooient nulle
nouvelle dou dit conte. Quant ce vint sus l’eure de
remontière, et il veirent que point ne venoit li
10 contes, si disent entre yaus: «Que ferons nous?
Irons nous assallir nos ennemis, ou nous retourons?»
Là furent en grant imagination quel cose il
en feroient, car ilz ne se veoient mies gens pour
combatre une tèle hoost qu’il y avoit devant Auberoce,
15 car il n’estoient non plus de trois cens lances
et de six cens arciers. Et li François pooient estre
entre dix mille et onze mille hommes. A envis ossi
le la[i]ssoient, car bien savoient, se il se partoient sans
le siège lever, il perd[r]oient le chastiel d’Auberoce et
20 les chevaliers leurs compagnons qui dedens estoient.
Finablement, tout consideret, et peset le bien contre
le mal, il s’acordèrent à ce que, ou nom de Dieu et
de saint Jorge, il iroient combatre leurs ennemis.
 Or avisèrent il comment; et l’avis là où le plus il
25 s’arrestèrent, il leur vint de monsigneur Gautier de
Mauni, qui dist ensi: «Signeur, nous monterons
tout à cheval, et costierons à le couverte ce bois où
nous sommes à present, tant que nous serons sus
l’autre cornée, au lés delà qui joint moult priès de
30 leur host. Et quant nous serons priès, nous feri[r]ons
chevaus des esporons et escrierons nos cris hautement;
nous y enterons droit sus l’eure dou souper:
[69] vous les verés si souspris et si esbahis de nous, qu’il
se desconfiront d’eulz meismes.» Adonc respondirent
li chevalier qui furent appellet à ce conseil: «Nous
le ferons ensi que vous l’ordonnés.» Si reprist cescuns
5 son cheval, et les recenglèrent estroitement; et
fisent restraindre leurs armeures, et ordonnèrent
tous leurs pages, leurs varlès et leurs malètes à là
demorer. Et puis chevaucièrent tout souef au loing
dou bois, tant qu’il vinrent sus l’autre cornée où li
10 hos françoise estoit logie assés priès, en un grant val,
sus une petite rivière. Lors qu’il furent là venu, il
desvolepèrent leurs banières et leurs pennons, et ferirent
chevaus des esperons, et s’en vinrent tout de
front sus le large planter et ferir en l’ost de ces signeurs
15 de Gascongne, qui furent bien souspris, et
leurs gens ossi; car de celle embusche ne se donnoient
il nulle garde, et se devoient tantos seoir au
souper. Et li pluiseur y estoient jà assis comme gent
asseguret, car il ne cuidaissent jamais que li contes
20 Derbi deuist là venir ensi à tèle heure.

§ 220. Evous les Englès venant frapant en celle

host, pourveus et avisés de ce qu’il devoient faire, en
escriant: «[D]erbi[305], [D]erbi au conte!» et «Mauni,
Mauni au signeur!» Puis commencièrent à coper et à
25 decoper tentes, trés et pavillons, et reverser l’un sus
l’autre, et abatre et occire et mehagnier gens, et mettre
 en grant meschief; ne [les François[306]] ne savoient auquel
entendre, tant estoient il quoitiet et fort hastet. Et
[70] quant il se traioient sus les camps pour yaus recueillier
et assambler, il trouvoient arciers tous appareilliés
qui les traioient et bersoient et occioient sans
merci et sans pité. Là avint soudainnement sus ces
5 chevaliers de Gascongne uns grans meschiés, car il
n’eurent nul loisir d’yaus armer ne traire sus les
camps. Mais fu li contes de [Lille] pris en son pavillon
et moult durement navrés, et li contes de Pieregorch
ossi dedens le sien, et messires Rogiers, ses oncles,
10 et occis li sires de Duras et messires Aymars
de Poitiers, et pris li contes de Valentinois ses frères.
Briefment, on ne vit onques tant de bonnes gens,
chevaliers et escuiers qui là estoient, perdu à si peu
de fait, car cescuns fuioit que mieulz mieulz. Bien
15 est verité que li contes de Commignes et li viscontes
de Quarmaing et cil de Villemur et cils de Brunikiel
et li sires de la Barde et li sires de Taride, qui estoient
logiet d’autre part le chastiel, se recueillièrent
et misent leurs banières hors, et se traisent vassaument
20 sus les camps. Mais li Englès, qui avoient jà
desconfis le plus grant partie de l’ost, s’en vinrent
en escriant leurs cris celle part, et se boutèrent e[n]s
de plains eslais, ensi que gens tous reconfortés et
qui veoient bien, se fortune ne leur estoit trop contraire,
25 que li journée estoit pour yaus. Là eut fait
mainte belle apertise d’armes, mainte prise et mainte
rescousse. Quant messires Franke de Halle et messires
Alains de Finefroide [et messires Jehans de Lindehalle[307]],
qui estoient ens ou chastiel [de Auberoche[308]],
[71] entendirent le noise et le hue, et recogneurent les
banières et les pennons de leurs gens, si s’armèrent
et fisent armer tous chiaus qui avoecques euls
estoient. Et puis montèrent à cheval, et issirent de le
5 forterèce d’Auberoce, et s’en vinrent sus les camps
et se boutèrent ou plus fort de le bataille: ce rafresci
 et resvigura grandement les Englès.

§ 221. Que vous feroi je lonch parlement? Chil

de le partie le conte de [Lille] furent là tout desconfi,
10 et priès que tout mort et tout pris. Jà ne s’en fust
nulz escapés, se la nuis ne fust si tost venue. Là y
ot pris, que contes que viscontes, jusques à neuf, et
des barons et des chevaliers tel fuison qu’il n’i avoit
homme d’armes des Englès qui n’en euist deux ou
15 trois dou mains, des quelz il eurent depuis grant
pourfit. Ceste bataille fu desous Auberoce, le nuit
Saint Laurens en aoust, l’an de grasce Nostre Signeur
mil trois cens quarante quatre.

§ 222. Apriès le desconfiture qui fu là si grande

20 et si grosse pour les Gascons et si adamagable, car
il estoient là venu en grant arroi et en bonne ordenance,
mais petite songne les fist perdre ensi qu’il
apparu, li Englès, qui estoient mestre et signeur
dou camp, entendirent à leurs prisonniers, et, comme
25 gens qui leur ont toutdis estet courtois, leur fisent
très bonne compagnie, et en recrurent assés sus leurs
fois à revenir dedens un certain jour à Bourdiaus ou
à Bregerach; [et] il se retraisent dedens Auberoce. Et
là donna à souper li contes Derbi le plus grant partie
30 des contes, des viscontes, qui prisonnier estoient, et
[72] ossi les chevaliers de se compagnie. Si devés croire
et savoir qu’il furent celle nuit en grant reviel, et rendirent
grans grasces à Nostre Signeur de le belle journée
qu’il avoient eu, quant une puignie de gens qu’il
5 estoient, environ mil combatans parmi les arciers,
uns c’autres, en avoient desconfi dix mil et plus, et
rescous le ville et le chastiel d’Auberoce, et les chevaliers
leurs compagnons qui dedens estoient moult
astraint, et qui dedens deux jours euissent esté pris
10 et en le volenté de leurs ennemis.
 Quant ce vint au matin, un peu apriès soleil levant,
li contes de Pennebruch vint à bien trois cens lances
et quatre cens archiers, qui jà avoit esté enfourmés
sus son chemin de l’avenue de le bataille. Si estoit
15 durement courouciés de ce qu’il n’i avoit esté; et en
parla par mautalent au conte Derbi, et dist: «Certes,
cousins, il me samble que vous ne m’avés fait maintenant
point d’onneur ne de courtoisie, quant vous
avés combatus vos ennemis sans moy, qui m’aviés
20 mandé si acertes. Et bien poiés savoir que je ne me
fuisse jamais souffers que je ne fuisse venus.» Donc
respondi li contes Derbi, et dist tout en riant: «Par
ma foy, cousins, nous desirions bien vostre venue.
Et nous souffresi[o]ns toutdis, en vous sourattendant
25 dou matin jusques as vespres. Et quant nous veimes
que vous ne veniés point, nous en estions tout
esmervilliet. Si n’osames plus attendre que nostre
anemi ne seuissent nostre venue. Car, se il le seuissent,
il euissent eu l’avantage sur nous. Et, Dieu
30 merci, nous l’avons eu sur yaus; si les nous aiderés
à garder et à conduire jusques à Bourdiaus.» Adonc
se prisent par les mains, et entrèrent en une cambre,
[73] et issirent de ce pourpos. Tantost fu heure de disner;
si se misent à table; si mengièrent et burent,
tout aise et à grant loisir, des pourveances des François
qu’il avoient amené devant le chastiel de Auberoce,
5 dont il estoient bien raempli. Tout ce jour et le
nuit ensievant, se tinrent il en Auberoce, et se reposèrent
et rafreschirent grandement. Et l’endemain
au matin, il furent tout armé et tout monté. Si se
partirent de Auberoce, et y laissièrent à chapitainne
10 et à gardiien un chevalier gascon, qui toutdis avoit
estet de leur partie, qui s’appelloit messires Alixandres
de Chaumont. Et chevaucièrent devers Bourdiaus,
et emmenèrent le plus grant partie de leurs
prisonniers.
 15 § 223. Tant chevaucièrent li dessus dit Englès et
leurs routes qu’il vinrent en le cité de Bourdiaus où
il furent recheu à grant joie. Et ne savoient li Bourdelois
comment bien festiier le conte Derbi et monsigneur
Gautier de Mauni, car li renommée aloit que
20 par leur emprise avoient esté devant Auberoce li
Gascon desconfi, et pris li contes de [Lille] et plus
de deux cens chevaliers. Si leur faisoient grant joie
et haute honneur. Ensi se passèrent il cel yvier qu’il
n’i eut nulles besongnes ens ès marces par de delà
25 qui à recorder face. Si ooit souvent li rois d’Engleterre
bonnes nouvelles dou conte Derbi, son cousin,
qui se tenoit à Bourdiaus sus Gironde et là environ.
Si en estoit tous liés et à bonne cause, car li contes
Derbi faisoit tant qu’il estoit amés de tous ses amis
30 et ressongniés de tous ses ennemis.
Quant ce vint apriès Paskes, que on compta l’an
[74] mil trois cens quarante cinq, environ le moiiené de
may, li contes Derbi, qui s’estoit tenus et yvrenés
tout le temps à Bourdiaus ou là priès, fist une coeilloite
et un amas de gens d’armes et d’arciers, et dist
5 qu’il voloit faire une chevaucie devers le Riolle que
li François tenoient, et le assegeroit, car elle estoit
bien prendable. Quant toutes ses besongnes furent
ordenées et ses gens venus, il se partirent de Bourdiaus
en bon arroy et en grant couvenant, et vinrent
10 ce premier jour en le ville de Bregerach. Là trouvèrent
il le conte de Pennebruch, qui avoit fait ossi
sen assamblée d’autre part.
Si furent cil signeur et leurs gens dedens Bregerach
trois jours; au quatrime, il s’en partirent.
15 Quant il furent sus les camps, il esmèrent leurs gens
et considerèrent leur pooir, et se trouvèrent mil
combatans et deux mil archiers. Si chevaucièrent
 tout ensi, et fisent tant qu’il vinrent devant un chastiel
que on claime Sainte Basille. Quant il furent
20 là venu, il le assegièrent de tous lés, et fisent grant
apparant de l’assallir. Chil de Sainte Basille veirent
les Englès et leur force, comment il tenoient les
camps, et que nulz ne lor aloit au devant. Mès encores
estoient prisonnier de le bataille de Auberoce
25 tout li plus grant de Gascongne, dont il deuissent
estre aidié et conforté: si ques, tout consideret, il
se misent en l’obeissance dou conte Derbi qui representoit
adonc là le personne dou roy d’Engleterre,
et li jurèrent feaulté et hommage, et le recogneurent
30 à signeur.
Par ensi il passa oultre bellement, et prist le chemin
de Aiguillon; mais, ançois qu’il y parvenist, il
[75] trouva en son chemin un chastiel que on appelle le
Roce Millon, qui estoit bien pourveus de bons saudoiiers
et d’arteillerie. Non obstant ce, li contes commanda
que li chastiaus fust assallis. Donc s’avancièrent
5 Englès et arcier, et le commencièrent à assallir
fortement et durement, et chil de dedens à yaus
deffendre vassaument. Et jettoient pières et baus et
grans barriaus de fier, et pos plains de cauch: de
quoi il blechièrent pluiseurs assallans qui montoient
10 contremont et s’abandonnoient folement, pour leurs
corps avancier.

§ 224. Quant li contes Derbi vei que ses gens se

travilloient et tuoient sans riens faire, si les fist retraire
et revenir as logeis. A l’endemain, il fist par les
15 villains dou pays achariier et aporter grant fuison de
busce et de velourdes et d’estrain, et tout jetter et
tourner ens ès fossés, et mettre ossi grant plenté de
terre. Quant une partie des fossés furent tout empli,
que on pooit bien aler seurement jusques au piet
20 dou mur, il fist arouter bien trois cens arciers, et par