Medical Image Understanding and

Analysis 21st Annual Conference MIUA

2017 Edinburgh UK July 11 13 2017
Proceedings 1st Edition María Valdés
Hernández
Visit to download the full and correct content document:
https://textbookfull.com/product/medical-image-understanding-and-analysis-21st-ann
ual-conference-miua-2017-edinburgh-uk-july-11-13-2017-proceedings-1st-edition-mar
ia-valdes-hernandez/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Collaborative Computing Networking Applications and

Worksharing 13th International Conference
CollaborateCom 2017 Edinburgh UK December 11 13 2017
Proceedings Imed Romdhani
https://textbookfull.com/product/collaborative-computing-
networking-applications-and-worksharing-13th-international-
conference-collaboratecom-2017-edinburgh-uk-
december-11-13-2017-proceedings-imed-romdhani/

Image Analysis and Recognition 14th International

Conference ICIAR 2017 Montreal QC Canada July 5 7 2017
Proceedings 1st Edition Fakhri Karray

https://textbookfull.com/product/image-analysis-and-
recognition-14th-international-conference-iciar-2017-montreal-qc-
canada-july-5-7-2017-proceedings-1st-edition-fakhri-karray/

Image Analysis and Processing ICIAP 2017 19th

International Conference Catania Italy September 11 15
2017 Proceedings Part II 1st Edition Sebastiano
Battiato
https://textbookfull.com/product/image-analysis-and-processing-
iciap-2017-19th-international-conference-catania-italy-
september-11-15-2017-proceedings-part-ii-1st-edition-sebastiano-
battiato/

Image Analysis and Processing ICIAP 2017 19th

International Conference Catania Italy September 11 15
2017 Proceedings Part I 1st Edition Sebastiano Battiato

https://textbookfull.com/product/image-analysis-and-processing-
iciap-2017-19th-international-conference-catania-italy-
september-11-15-2017-proceedings-part-i-1st-edition-sebastiano-
Intelligent Human Computer Interaction 9th
International Conference IHCI 2017 Evry France December
11 13 2017 Proceedings 1st Edition Patrick Horain

https://textbookfull.com/product/intelligent-human-computer-
interaction-9th-international-conference-ihci-2017-evry-france-
december-11-13-2017-proceedings-1st-edition-patrick-horain/

Data Analytics 31st British International Conference on

Databases BICOD 2017 London UK July 10 12 2017
Proceedings 1st Edition Andrea Calì

https://textbookfull.com/product/data-analytics-31st-british-
international-conference-on-databases-bicod-2017-london-uk-
july-10-12-2017-proceedings-1st-edition-andrea-cali/

Research in Computational Molecular Biology 21st Annual

International Conference RECOMB 2017 Hong Kong China
May 3 7 2017 Proceedings 1st Edition S. Cenk Sahinalp
(Eds.)
https://textbookfull.com/product/research-in-computational-
molecular-biology-21st-annual-international-conference-
recomb-2017-hong-kong-china-may-3-7-2017-proceedings-1st-edition-
s-cenk-sahinalp-eds/

Logic, Language, Information, and Computation: 24th

International Workshop, WoLLIC 2017, London, UK, July
18-21, 2017, Proceedings 1st Edition Juliette Kennedy

https://textbookfull.com/product/logic-language-information-and-
computation-24th-international-workshop-wollic-2017-london-uk-
july-18-21-2017-proceedings-1st-edition-juliette-kennedy/

Reasoning Web Semantic Interoperability on the Web 13th

International Summer School 2017 London UK July 7 11
2017 Tutorial Lectures 1st Edition Giovambattista Ianni

https://textbookfull.com/product/reasoning-web-semantic-
interoperability-on-the-web-13th-international-summer-
school-2017-london-uk-july-7-11-2017-tutorial-lectures-1st-
María Valdés Hernández
Víctor González-Castro (Eds.)

Communications in Computer and Information Science 723

Medical Image
Understanding
and Analysis
21st Annual Conference, MIUA 2017
Edinburgh, UK, July 11–13, 2017
Proceedings

123
Communications
in Computer and Information Science 723
Commenced Publication in 2007
Founding and Former Series Editors:
Alfredo Cuzzocrea, Dominik Ślęzak, and Xiaokang Yang

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Phoebe Chen
La Trobe University, Melbourne, Australia
Xiaoyong Du
Renmin University of China, Beijing, China
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Orhun Kara
TÜBİTAK BİLGEM and Middle East Technical University, Ankara, Turkey
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Ting Liu
Harbin Institute of Technology (HIT), Harbin, China
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
More information about this series at http://www.springer.com/series/7899
María Valdés Hernández
Víctor González-Castro (Eds.)

Medical Image
Understanding and Analysis
21st Annual Conference, MIUA 2017
Edinburgh, UK, July 11–13, 2017
Proceedings

123
Editors
María Valdés Hernández Víctor González-Castro
Department of Neuroimaging Sciences Universidad de León
University of Edinburgh León
Edinburgh Spain
UK

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-3-319-60963-8 ISBN 978-3-319-60964-5 (eBook)
DOI 10.1007/978-3-319-60964-5

Library of Congress Control Number: 2017944214

© Springer International Publishing AG 2017

The chapters ‘Model-Based Correction of Segmentation Errors in Digitised Histological Images’ and
‘Unsupervised Superpixel-Based Segmentation of Histopathological Images with Consensus Clustering’ are
licensed under the terms of the Creative Commons Attribution 4.0 International License (http://
creativecommons.org/licenses/by/4.0/). For further details see license information in the chapters.
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. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

This volume comprises the proceedings of the 21st edition of the Medical Image
Understanding and Analysis (MIUA) Conference, an annual forum organized in the
United Kingdom for communicating research progress within the community interested
in biomedical image analysis. Its goals are the dissemination and discussion of research
in medical image analysis to encourage the growth and to raise the profile of this
multi-disciplinary field that has an ever-increasing real-world applicability. The con-
ference constitutes an excellent opportunity to network, generate new ideas, establish
new collaborations, learn about and discuss different topics, listen to speakers of
international reputation, and present and show medical image analysis tools.
This year’s edition was organized by The Row Fogo Centre for Research in Ageing
and Dementia in conjunction with Edinburgh Imaging (http://www.ed.ac.uk/clinical-
sciences/edinburgh-imaging) at The University of Edinburgh, in partnership with the
Scottish Imaging Network A Platform for Scientific Excellence (SINAPSE) (http://
www.sinapse.ac.uk/), the Journal of Imaging (http://www.mdpi.com/journal/jimaging),
and the EPSRC Medical Image Analysis Network (MedIAN) (http://www.ibme.ox.ac.
uk/MedIAN) and MathWorks (https://uk.mathworks.com/); it was supported by the
British Machine Vision Association (BMVA), Toshiba Medical Visualisation Systems
Europe, General Electric, Bayer, Siemens, Edinburgh Imaging, Optos, Holoxica Ltd.,
AnalyzeDirect and NVIDIA.
The number and level of submissions of this year’s edition were unprecedented.
In all, 105 technical papers and 22 abstracts showing clinical applications of
image-processing techniques (the latter not considered for inclusion in this volume)
were revised by an expert team of 93 reviewers from British(38), Spanish(25),
French(6), Italian(11), Swedish(5), Greek(2), New Zealand(1), American(3),
German(1), Dutch(1), Bangladeshi(1) and Swiss(1) institutions. Each of the 127 sub-
missions were reviewed by two to four members of the Program Committee. Based on
their ranking and recommendations, 18 of 105 papers were accepted, 65 of 105 papers
were provisionally accepted pending minor revisions and suggestions, and 22 of 105
papers were rejected for publication in this volume. From the papers rejected, 15 were
considered to need a major revision, and their authors were invited to address the
reviewers’ comments and submit their revised work to the Journal of Imaging (con-
ference partner), to which the reviews will be provided for facilitating the new review
process. After a second round of revision, we are including in this volume 82 full
papers. We hope you agree with us that they all show high quality and represent a step
forward in the medical image analysis field.
We thank all members of the MIUA 2017 Organizing, Program, and Steering
Committees and, particularly, all those who supported MIUA 2017 by submitting
papers and attending the meeting. We thank Professor Joanna Wardlaw, Head of the
Academic Hub and Centre that organized the conference, for her welcome words.
VI Preface

We also thank our speakers Professors Sir Michael Brady, Daniel Rueckert, Ingela
Nyström, and Jinah Park, and Dr. Constantino Carlos Reyes Aldasoro and Kon-
stantinos Kamnitsas for sharing their success, knowledge, and experiences. We hope
you enjoy the proceedings of MIUA 2017.

May 2017 Maria Valdes Hernandez

Víctor González-Castro
 Organization

Program Chairs
Maria Valdes Hernandez University of Edinburgh, UK
Víctor González-Castro Universidad de León, Spain

MIUA Steering Committee

Yan Chen Loughborough University, UK
Bill Crum King’s College London, UK
Alastair Gale Loughborough University, UK
Víctor González-Castro Universidad de León, Spain
Tryphon Lambrou University of Lincoln, UK
Stephen McKenna University of Dundee, UK
Nasir Rajpoot University of Warwick, UK
Constantino Carlos City, University of London, UK
Reyes-Aldasoro
Greg Slabaugh City, University of London, UK
Maria Valdes Hernandez University of Edinburgh, UK
Xianghua Xie Swansea University, UK
Xujiong Ye University of Lincoln, UK
Reyer Zwiggelaar Aberystwyth University, UK

Local Organizing Committee

Devasuda Anblagan University of Edinburgh, UK
Lucia Ballerini University of Edinburgh, UK
Kristin Flegal University of Glasgow/SINAPSE, UK
Anne Grant University of Edinburgh, UK
Taku Komura University of Edinburgh, UK
Tom MacGillivray University of Edinburgh, UK
Ian Marshall University of Edinburgh, UK
Cyril Pernet University of Edinburgh, UK
Amos Storkey University of Edinburgh, UK
Emanuele Trucco University of Dundee, UK
Edwin van Beek Clinical Research Imaging Centre,
University of Edinburgh, UK

Technical Program Committee

Jose Luis Alba-Castro University of Vigo, Spain
Enrique Alegre Universidad de León, Spain
VIII Organization

Luciano Alparone University of Florence, Italy

Devasuda Anblagan University of Edinburgh, UK
Paul Armitage University of Sheffield, UK
Juan Arribas University of Valladolid, Spain
Lucia Ballerini University of Edinburgh, UK
Isabelle Bloch Telecom ParisTech, France
Leonardo Bocchi University of Florence, Italy
Gunilla Borgefors Centre for Image Analysis, Uppsala University, Sweden
Larbi Boubchir University of Paris 8, France
Gloria Bueno Universidad de Castilla-La Mancha, Spain
Maria J. Carreira Universidade de Santiago de Compostela, Spain
Alessandro Cavinato École Polytechnique Fédérale de Lausanne, Switzerland
Yan Chen Loughborough University, UK
Samuel Danso University of Edinburgh, UK
Johan Debayle École Nationale Supérieure des Mines de Saint-Étienne,
France
David Alexander Dickie University of Edinburgh, UK
Javier Escudero University of Edinburgh, UK
María del Milagro Universidad de Castilla-La Mancha, Spain
Fernández Carrobles
Laura Fernández-Robles Universidad de León, Spain
Samuele Fiorini Università degli studi di Genova, Italy
Francesco Fontanella Università di Cassino e del Lazio meridionale, Italy
Alejandro Frangi University of Sheffield, UK
Alastair Gale Loughborough University, UK
María García University of Valladolid, Spain
Andrea Garzelli University of Siena, Italy
Yann Gavet École Nationale Supérieure des Mines de Saint-Etienne,
France
Andrea Giachetti University of Verona, Italy
Calum Gray University of Edinburgh, UK
Manuel Graña Basque Country University (UPV/EHU), Spain
Enrico Grisan University of Padova, Italy
Juan M. Górriz University of Granada, Spain
Taku Komura University of Edinburgh, UK
Violet Kovacheva Institute of Cancer Research, UK
Tryphon Lambrou University of Lincoln, UK
Joakim Lindblad Mathematical Institute of the Serbian
Academy of Sciences and Arts, Serbia
Ruggiero Lovreglio University of Auckland, New Zealand
Evelyne Lutton INRA, France
Tom MacGillivray University of Edinburgh, UK
Sasan Mahmoodi University of Southampton, UK
Filip Malmberg Uppsala University, Sweden
Ian Marshall University of Edinburgh, UK
Stephen McKenna University of Dundee, UK
 Organization IX

Bjoern Menze Technical University of Munich, Germany

Mariofanna Milanova University of Arkansas at Little Rock, USA
Carmel Moran University of Edinburgh, UK
Philip Morrow Ulster University, UK
Susana Munoz Maniega University of Edinburgh, UK
Henning Müller HES-SO, Switzerland
Bill Nailon University of Edinburgh/NHS Lothian, UK
Mark Nixon University of Southampton, UK
Jorge Novo Buján University of A Coruña, Spain
Marcos Ortega University of A Coruña, Spain
Gonzalo Pajares University Complutense of Madrid, Spain
Pietro Pala University of Florence, Italy
Georgios Papanastasiou University of Edinburgh, UK
Roberto Paredes Universidad Politécnica de Valencia, Spain
Alejandro Pazos Sierra Universidade da Coruña, Spain
Enrico Pellegrini University of Edinburgh, UK
Cyril Pernet University of Edinburgh, UK
Jean-Charles Pinoli École Nationale Supérieure des Mines, France
Ian Poole TMVS, UK
Kashif Rajpoot University of Birmingham, UK
Nasir Rajpoot University of Warwick, UK
Javier Ramírez University of Granada, Spain
Constantino Carlos City, University of London, UK
Reyes-Aldasoro
Scott Semple University of Edinburgh, UK
Korsuk Sirinukunwattana Harvard Medical School, USA
Greg Slabaugh City, University London, UK
Natasa Sladoje Centre for Image Analysis, Uppsala University, Sweden
Robin Strand Centre for Image Analysis, Uppsala University, Sweden
Pablo G. Tahoces Universidade de Santiago de Compostela, Spain
Emanuele Trucco University of Dundee, UK
Maria Valdes Hernandez University of Edinburgh, UK
Cesar Veiga Instituto de Investigación Sanitaria Galicia Sur - IISGS,
Spain
Xianghua Xie Swansea University, UK
Xujiong Ye University of Lincoln, UK
Xenophon Zabulis Foundation for Research and Technology, Greece
Matteo Zanotto Istituto Italiano di Tecnologia, Italy
Reyer Zwiggelaar Aberystwyth University, UK
 Contents

Retinal Imaging

End-to-End Learning of a Conditional Random Field for Intra-retinal

Layer Segmentation in Optical Coherence Tomography . . . . . . . . . . . . . . . . 3
Arunava Chakravarty and Jayanthi Sivaswamy

Superpixel-Based Line Operator for Retinal Blood Vessel Segmentation . . . . 15

Tong Na, Yitian Zhao, Yifan Zhao, and Yue Liu

Automatic Detection and Identification of Retinal Vessel Junctions

in Colour Fundus Photography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Harry Pratt, Bryan M. Williams, Jae Ku, Frans Coenen,
and Yalin Zheng

Fast Optic Disc Segmentation in Retinal Images Using Polar Transform . . . . 38

Muhammad Nauman Zahoor and Muhammad Moazam Fraz

A Novel Technique for Splat Generation and Patch Level Prediction

in Diabetic Retinopathy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
I. Syed Muhammedh Ajwahir, Kumar Rajamani, and S. Ibrahim Sadhar

Ultrasound Imaging

Deep Residual Networks for Quantification of Muscle Fiber Orientation

and Curvature from Ultrasound Images . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Ryan Cunningham, Peter Harding, and Ian Loram

Modelling, Speckle Simulation and Quality Evaluation of Synthetic

Ultrasound Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Prerna Singh, Ramakrishnan Mukundan, and Rex de Ryke

Multi-level Trainable Segmentation for Measuring Gestational

and Yolk Sacs from Ultrasound Images . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Dheyaa Ahmed Ibrahim, Hisham Al-Assam, Sabah Jassim,
and Hongbo Du

Weakly Supervised Learning of Placental Ultrasound Images with Residual

Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Huan Qi, Sally Collins, and Alison Noble
XII Contents

Edge Aware Geometric Filter for Ultrasound Image Enhancement. . . . . . . . . 109

Deepak Mishra, Santanu Chaudhury, Mukul Sarkar,
and Arvinder Singh Soin

Cardiovascular Imaging

Tissues Classification of the Cardiovascular System

Using Texture Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Claudia Mazo, Enrique Alegre, Maria Trujillo,
and Víctor González-Castro

Multidimensional Assessments of Abdominal Aortic Aneurysms by

Magnetic Resonance Against Ultrasound Diameter Measurements . . . . . . . . . 133
G. Papanastasiou, V. González-Castro, C.D. Gray, R.O. Forsythe,
Y. Sourgia-Koutraki, N. Mitchard, D.E. Newby, and S.I.K. Semple

Comparison of Automatic Vessel Segmentation Techniques for Whole

Body Magnetic Resonance Angiography with Limited Ground Truth Data . . . 144
Andrew McNeil, Giulio Degano, Ian Poole, Graeme Houston,
and Emanuele Trucco

Evaluating Classifiers for Atherosclerotic Plaque Component

Segmentation in MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Arna van Engelen, Marleen de Bruijne, Torben Schneider,
Anouk C. van Dijk, M. Eline Kooi, Jeroen Hendrikse, Aart Nederveen,
Wiro J. Niessen, and Rene M. Botnar

Cardiac Mesh Reconstruction from Sparse, Heterogeneous Contours . . . . . . . 169

Benjamin Villard, Valentina Carapella, Rina Ariga, Vicente Grau,
and Ernesto Zacur

Classification of Cross-sections for Vascular Skeleton Extraction

Using Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Kristína Lidayová, Anindya Gupta, Hans Frimmel, Ida-Maria Sintorn,
Ewert Bengtsson, and Örjan Smedby

Segmenting Atrial Fibrosis from Late Gadolinium-Enhanced Cardiac

MRI by Deep-Learned Features with Stacked Sparse Auto-Encoders . . . . . . . 195
Guang Yang, Xiahai Zhuang, Habib Khan, Shouvik Haldar,
Eva Nyktari, Xujiong Ye, Greg Slabaugh, Tom Wong,
Raad Mohiaddin, Jennifer Keegan, and David Firmin

Improved CTA Coronary Segmentation with a Volume-Specific

Intensity Threshold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Muhammad Moazzam Jawaid, Ronak Rajani, Panos Liatsis,
Constantino Carlos Reyes-Aldasoro, and Greg Slabaugh
 Contents XIII

Segmentation of Abdominal Aortic Aneurysm (AAA) Based

on Topology Prior Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Safa Salahat, Ahmed Soliman, Tim McGloughlin, Naoufel Werghi,
and Ayman El-Baz

Automated LGE Myocardial Scar Segmentation Using MaskSLIC

Supervoxels - Replicating the Clinical Method . . . . . . . . . . . . . . . . . . . . . . 229
Iulia A. Popescu, Alessandra Borlotti, Erica Dall’Armellina,
and Vicente Grau

Oncology Imaging

Multi-task Fully Convolutional Network for Brain Tumour Segmentation. . . . 239

Haocheng Shen, Ruixuan Wang, Jianguo Zhang, and Stephen McKenna

FF-CNN: An Efficient Deep Neural Network for Mitosis Detection

in Breast Cancer Histological Images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Boqian Wu, Tasleem Kausar, Qiao Xiao, Mingjiang Wang,
Wenfeng Wang, Binwen Fan, and Dandan Sun

Classification of Cervical-Cancer Using Pap-Smear Images:

A Convolutional Neural Network Approach . . . . . . . . . . . . . . . . . . . . . . . . 261
Bilal Taha, Jorge Dias, and Naoufel Werghi

New Level Set Model in Follow Up Radiotherapy Image Analysis . . . . . . . . 273

Roushanak Rahmat, William Henry Nailon, Allan Price,
David Harris-Birtill, and Stephen McLaughlin

Topological Analysis of the Vasculature of Angiopoietin-Expressing

Tumours Through Scale-Space Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Constantino Carlos Reyes-Aldasoro, Meit Bjorndahl, Chryso Kanthou,
and Gillian M. Tozer

Quantitative Electron Density CT Imaging for Radiotherapy Planning . . . . . . 297

Jonathan H. Mason, Alessandro Perelli, William H. Nailon,
and Mike E. Davies

3D Texton Based Prostate Cancer Detection Using Multiparametric

Magnetic Resonance Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Liping Wang and Reyer Zwiggelaar

Tumor Segmentation in Whole Slide Images Using Persistent Homology

and Deep Convolutional Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
Talha Qaiser, Yee-Wah Tsang, David Epstein, and Nasir Rajpoot

Multispectral Biopsy Image Based Colorectal Tumor Grader . . . . . . . . . . . . 330

Suchithra Kunhoth and Somaya Al Maadeed
XIV Contents

Semi-automatic Bone Marrow Evaluation in PETCT

for Multiple Myeloma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
Patrick Leydon, Martin O’Connell, Derek Greene, and Kathleen Curran

Mammography Image Analysis

A Texton-Based Approach for the Classification of Benign

and Malignant Masses in Mammograms. . . . . . . . . . . . . . . . . . . . . . . . . . . 355
Zobia Suhail, Azam Hamidinekoo, Erika R.E. Denton,
and Reyer Zwiggelaar

Breast Density Classification Using Multiresolution Local Quinary

Patterns in Mammograms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Andrik Rampun, Philip Morrow, Bryan Scotney, and John Winder

Rich Interaction and Feedback Supported Mammographic Training:

A Trial of an Augmented Reality Approach . . . . . . . . . . . . . . . . . . . . . . . . 377
Qiang Tang, Yan Chen, and Alastair G. Gale

A Robust Algorithm for Automated HER2 Scoring in Breast Cancer

Histology Slides Using Characteristic Curves . . . . . . . . . . . . . . . . . . . . . . . 386
Ramakrishnan Mukundan

Investigating the Effect of Various Augmentations on the Input Data Fed

to a Convolutional Neural Network for the Task of Mammographic
Mass Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
Azam Hamidinekoo, Zobia Suhail, Talha Qaiser, and Reyer Zwiggelaar

Brain Imaging

Learning Longitudinal MRI Patterns by SICE and Deep Learning:

Assessing the Alzheimer’s Disease Progression. . . . . . . . . . . . . . . . . . . . . . 413
Andrés Ortiz, Jorge Munilla, Francisco J. Martínez-Murcia,
Juan M. Górriz, Javier Ramírez, and for the Alzheimer’s Disease
Neuroimaging Initiative

Improved Reference Tracts for Unsupervised Brain White

Matter Tractography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Susana Muñoz Maniega, Mark E. Bastin, Ian J. Deary,
Joanna M. Wardlaw, and Jonathan D. Clayden

Review of Fast Density-Peaks Clustering and Its Application to Pediatric

White Matter Tracts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
Shichao Cheng, Yuzhuo Duan, Xin Fan, Dongyu Zhang, and Hua Cheng
 Contents XV

A Deep Learning Pipeline to Delineate Proliferative Areas of Intracranial

Tumors in Digital Slides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
Zaneta Swiderska-Chadaj, Tomasz Markiewicz, Bartlomiej Grala,
Malgorzata Lorent, and Arkadiusz Gertych

Tree-Based Ensemble Learning Techniques in the Analysis

of Parkinsonian Syndromes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
J.M. Górriz, J. Ramírez, M. Moreno-Caballero, F.J. Martinez-Murcia,
A. Ortiz, I.A. Illán, F. Segovia, D. Salas-González, and M. Gomez-Rio

Evaluating Alzheimer’s Disease Diagnosis Using Texture Analysis . . . . . . . . 470

Francisco Jesús Martinez-Murcia, Juan Manuel Górriz, Javier Ramírez,
Fermin Segovia, Diego Salas-Gonzalez, Diego Castillo-Barnes,
Ignacio A. Illán, Andres Ortiz, and for the Alzheimer’s Disease
Neuroimaging Initiative

Evaluation of Four Supervised Learning Schemes in White Matter

Hyperintensities Segmentation in Absence or Mild Presence
of Vascular Pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
Muhammad Febrian Rachmadi, Maria del C. Valdés-Hernández,
Maria Leonora Fatimah Agan, Taku Komura, and The Alzheimer’s
Disease Neuroimaging Initiative

Context-Aware Convolutional Neural Networks for Stroke Sign

Detection in Non-contrast CT Scans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
Aneta Lisowska, Alison O’Neil, Vismantas Dilys, Matthew Daykin,
Erin Beveridge, Keith Muir, Stephen Mclaughlin, and Ian Poole

Automatic Brain Tumor Detection and Segmentation Using U-Net

Based Fully Convolutional Networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
Hao Dong, Guang Yang, Fangde Liu, Yuanhan Mo, and Yike Guo

Modeling Diffusion Directions of Corpus Callosum. . . . . . . . . . . . . . . . . . . 518

Safa Elsheikh, Andrew Fish, Roma Chakrabarti, Diwei Zhou,
and Mara Cercignani

Feature Extraction and Classification to Diagnose Hypoxic-Ischemic

Encephalopathy Patients by Using Susceptibility-Weighted MRI Images . . . . 527
Sisi Wu, Sasan Mahmoodi, Angela Darekar, Brigitte Vollmer,
Emma Lewis, and Maria Liljeroth

Evaluation of an Automatic ASPECT Scoring System for Acute Stroke

in Non-Contrast CT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
Matt Daykin, Erin Beveridge, Vismantas Dilys, Aneta Lisowska,
Keith Muir, Mathini Sellathurai, and Ian Poole
XVI Contents

Image Enhancement and Alignment

Pre-processing Techniques for Colour Digital Pathology Image Analysis . . . . 551

Wael Saafin and Gerald Schaefer

Motion Compensation Using Range Imaging in C-Arm Cone-Beam CT. . . . . 561

Bastian Bier, Mathias Unberath, Tobias Geimer, Jennifer Maier,
Garry Gold, Marc Levenston, Rebecca Fahrig, and Andreas Maier

Medical Image Colorization for Better Visualization and Segmentation . . . . . 571

Muhammad Usman Ghani Khan, Yoshihiko Gotoh, and Nudrat Nida

Fetoscopic Panorama Reconstruction: Moving from Ex-vivo to In-vivo . . . . . 581

Floris Gaisser, Suzanne H.P. Peeters, Boris Lenseigne,
Pieter P. Jonker, and Dick Oepkes

Finite Element Based Interactive Elastic Image Registration . . . . . . . . . . . . . 594

Yechiel Lamash, Anath Fischer, and Jonathan Lessick

Significance of Magnetic Resonance Image Details in Sparse

Representation Based Super Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 605
Prabhjot Kaur, Srimanta Mandal, and Anil K. Sao

Restoration of Intensity Uniformity of Bi-contrast MRI Data with Bayesian

Co-occurrence Coring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
Stathis Hadjidemetriou, Marios Nikos Psychogios, Paul Lingor,
Kajetan von Eckardstein, and Ismini Papageorgiou

Can Planning Images Reduce Scatter in Follow-Up Cone-Beam CT?. . . . . . . 629

Jonathan H. Mason, Alessandro Perelli, William H. Nailon,
and Mike E. Davies

Super Resolution Convolutional Neural Networks for Increasing Spatial

Resolution of 1 H Magnetic Resonance Spectroscopic Imaging . . . . . . . . . . . 641
Sevim Cengiz, Maria del C. Valdes-Hernandez, and Esin Ozturk-Isik

Radial Basis Function Interpolation for Rapid Interactive Segmentation

of 3-D Medical Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Negar Mirshahzadeh, Tanja Kurzendorfer, Peter Fischer, Thomas Pohl,
Alexander Brost, Stefan Steidl, and Andreas Maier

Modeling and Ssegmentation of Preclinical, Body

and Histological Imaging

Deep Quantitative Liver Segmentation and Vessel Exclusion

to Assist in Liver Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663
Benjamin Irving, Chloe Hutton, Andrea Dennis, Sid Vikal,
Marija Mavar, Matt Kelly, and Sir J. Michael Brady
 Contents XVII

Initial Results of Multilevel Principal Components Analysis

of Facial Shape. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
D.J.J. Farnell, J. Galloway, A. Zhurov, S. Richmond, P. Perttiniemi,
and V. Katic

Estimating Rodent Brain Volume by a Deformable Contour Model . . . . . . . . 686

Julio Camacho-Cañamón, María J. Carreira, Pedro Antonio Gutiérrez,
and Ramón Iglesias-Rey

MIMONet: Gland Segmentation Using Multi-Input-Multi-Output

Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698
Shan E. Ahmed Raza, Linda Cheung, David Epstein, Stella Pelengaris,
Michael Khan, and Nasir M. Rajpoot

Automated Polyp Segmentation in Colonoscopy Frames

Using Fully Convolutional Neural Network and Textons . . . . . . . . . . . . . . . 707
Lei Zhang, Sunil Dolwani, and Xujiong Ye

Model-Based Correction of Segmentation Errors in Digitised

Histological Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 718
David A. Randell, Antony Galton, Shereen Fouad, Hisham Mehanna,
and Gabriel Landini

A 2D Morphable Model of Craniofacial Profile and Its Application

to Craniosynostosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
Hang Dai, Nick Pears, and Christian Duncan

A Comparison of Texture Features Versus Deep Learning for Image

Classification in Interstitial Lung Disease . . . . . . . . . . . . . . . . . . . . . . . . . . 743
Alison O’Neil, Matthew Shepherd, Erin Beveridge, and Keith Goatman

A Novel High-Throughput Multispectral Cell Segmentation Algorithm . . . . . 754

Jenia Golbstein, Yaniv Tocker, Revital Sharivkin, Gabi Tarcic,
and Michael Vidne

Unsupervised Superpixel-Based Segmentation of Histopathological

Images with Consensus Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 767
Shereen Fouad, David Randell, Antony Galton, Hisham Mehanna,
and Gabriel Landini

A Non-integer Step Index PCNN Model and Its Applications . . . . . . . . . . . . 780

Zhen Yang, Yanan Guo, Xiaonan Gong, and Yide Ma

Segmentation of Overlapping Macrophages Using Anglegram Analysis . . . . . 792

José Alonso Solís-Lemus, Brian Stramer, Greg Slabaugh,
and Constantino Carlos Reyes-Aldasoro
XVIII Contents

New Disagreement Metrics Incorporating Spatial Detail – Applications

to Lung Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
Alberto M. Biancardi and Jim M. Wild

Unsupervised Segmentation of Cervical Cell Nuclei via Adaptive

Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
Srishti Gautam, Krati Gupta, Arnav Bhavsar, and Anil K. Sao

Feature Detection and Classification

Simultaneous Cell Detection and Classification with an Asymmetric

Deep Autoencoder in Bone Marrow Histology Images. . . . . . . . . . . . . . . . . 829
Tzu-Hsi Song, Victor Sanchez, Hesham EIDaly, and Nasir Rajpoot

Glomerulus Classification with Convolutional Neural Networks . . . . . . . . . . 839

Anibal Pedraza, Jaime Gallego, Samuel Lopez, Lucia Gonzalez,
Arvydas Laurinavicius, and Gloria Bueno

Paediatric Frontal Chest Radiograph Screening with Fine-Tuned

Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 850
Jonathan Gerrand, Quentin Williams, Dalton Lunga, Adam Pantanowitz,
Shabir Madhi, and Nasreen Mahomed

Automatic Hotspots Detection for Intracellular Calcium Analysis

in Fluorescence Microscopic Videos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 862
David Traore, Katja Rietdorf, Nasser Al-Jawad, and Hisham Al-Assam

Cervical Nuclei Classification: Feature Engineering Versus Deep

Belief Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 874
Christoph Rasche, Ciprian Ţigăneşteanu, Mihai Neghină,
and Alina Sultana

A New Method of Surgical Tracking System Based on Fiducial Marker . . . . 886

Shuaiyifan Ma and Zijian Zhao

Automated Detection of Barrett’s Esophagus Using Endoscopic Images:

A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 897
Noha Ghatwary, Amr Ahmed, and Xujiong Ye

Estimating Bacterial Load in FCFM Imaging . . . . . . . . . . . . . . . . . . . . . . . 909

Sohan Seth, Ahsan R. Akram, Kevin Dhaliwal,
and Christopher K.I. Williams

Random Forest-Based Feature Importance for HEp-2 Cell

Image Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 922
Vibha Gupta and Arnav Bhavsar
 Contents XIX

Automatic Quantification of Epidermis Curvature in H&E Stained

Microscopic Skin Image of Mice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
Saif Hussein, Sabah Jassim, and Hisham Al-Assam

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 947

Retinal Imaging
 End-to-End Learning of a Conditional Random
Field for Intra-retinal Layer Segmentation
in Optical Coherence Tomography

Arunava Chakravarty(B) and Jayanthi Sivaswamy

Center for Visual Information Technology, IIIT Hyderabad, Hyderabad, India

arunava.chakravarty@research.iiit.ac.in, jsivaswamy@iiit.ac.in

Abstract. Intra-retinal layer segmentation of Optical Coherence

Tomography images is critical in the assessment of ocular diseases. Exist-
ing Energy minimization based methods employ handcrafted cost terms
to define their energy and are not robust to the presence of abnormalities.
We propose a novel, Linearly Parameterized, Conditional Random Field
(LP-CRF) model whose energy is learnt from a set of training images
in an end-to-end manner. The proposed LP-CRF comprises two convo-
lution filter banks to capture the appearance of each tissue region and
boundary, the relative weights of the shape priors and an additional term
based on the appearance similarity of the adjacent boundary points. All
the energy terms are jointly learnt using the Structured Support Vec-
tor Machine. The proposed method segments all retinal boundaries in
a single step. Our method was evaluated on 107 Normal and 220 AMD
B-scan images and found to outperform three publicly available OCT seg-
mentation software. The average unsigned boundary localization error is
1.52±0.29 pixels for segmentation of 8 boundaries on the Normal dataset
and 1.9±0.65 pixels for 3 boundaries on the combined AMD and Normal
dataset establishing the robustness of the proposed method.

Keywords: CRF · SSVM · OCT · Segmentation

1 Introduction
Optical Coherence Tomography (OCT) images of retina provide a cross-sectional
view of the intra-retinal tissue which is composed of 7 adjacent layers [2], sepa-
rated by 8 boundaries as depicted in Fig. 1a. The boundaries ordered from the
top to bottom are the: (i) Inner Limiting Membrane (ILM) separating the Nerve
Fiber Layer (NFL) from vitreous, (ii) NFL/GCL boundary separating NFL from
the Ganglion Cell and Inner Plexicon layer (GCL-IPL), (iii) IPL/INL separat-
ing GCL-IPL from the Inner Nuclear Layer (INL), (iv) INL/OPL separating
INL from the Outer Plexiform Layer (OPL), (v) OPL/ONL separating OPL
from the Outer Nuclear and Inner Segment (ONL-IS) layer, (vi) IS/OS sepa-
rating ONL-IS from the Outer Segment (OS) layer (vii) RP Ein separating OS
from the Retinal Pigment Epithelium (RPE) layer and finally the (viii) RP Eout
boundary separating RPE from the choroid.

c Springer International Publishing AG 2017
M. Valdés Hernández and V. González-Castro (Eds.): MIUA 2017, CCIS 723, pp. 3–14, 2017.
DOI: 10.1007/978-3-319-60964-5 1
4 A. Chakravarty and J. Sivaswamy

a. b.

Fig. 1. Retinal layer boundaries in OCT B-scans of (a) Healthy retina, listed from
top to bottom: ILM(Red), NFL/GCL(Green), IPL/INL(blue), INL/OPL(yellow),
OPL/ONL(cyan), IS/OS(magenta), RP Ein (pink) and RP Eout (purple); (b) Retina
with AMD: ILM(Red), RP Ein (Green) and RP Eout (Blue). (Color figure online)

Accurate segmentation of these layers is necessary to quantify the morpholog-

ical changes in the retinal tissue and detect ocular diseases such as Age Related
Macular Degeneration (AMD) [1]. In AMD, the drusen deposits in the RPE
layer lead to irregularities and undulations in the RP Ein boundary as depicted
in Fig. 1b. Currently, commercial OCT systems are equipped to segment only
two or three layers and often fail for images of poor quality or/and with lesions.
Early work on OCT layer segmentation focussed on the segmentation of few
(2–4) prominent layers. These methods used peak, valley and signed gradient
analysis on the intensity profiles for each A-scan (column) in the OCT slice fol-
lowed by regularization across adjacent A-scans [5]. Their performance suffered
due to the lack of strong boundaries and overlapping intensities between the
adjacent layers, presence of speckle noise and vessel shadows.
To overcome these challenges, deformable model based methods were pro-
posed that incorporated shape priors in addition to the boundary and regional
appearance of the layers. Constraints on the thickness and smoothness of lay-
ers were imposed in [9] while in [11] the parallelism between the adjacent layer
boundaries was enforced. A circular arc based shape prior was employed in [14].
However, these methods are sensitive to the initial contour initialization leading
to high convergence time or entrapment of the curve in the local minima.
Graph based methods have been explored in [1,2] where the layers in each
OCT slice is segmented sequentially using the Dijkstra’s shortest path algo-
rithm. In [4], an energy composed of multiple cost terms was defined to capture
the appearance and shape priors for each layer. The energy minimization was
formulated as a Minimum Cost Closed Set (MCCS) problem on a geometric
graph. Each cost term in the energy had to be handcrafted manually and then
combined using empirically determined relative weights. These methods were
initially designed for the segmentation of healthy images and require complex
disease-specific modifications in [1,12] to handle abnormalities.
Thus, the goal of obtaining an optimal energy function for the joint multi-
layer segmentation of OCT images remains an open problem. The minima of
such an energy must correspond to the desired boundaries across a large set of
images and should be applicable for both healthy and abnormal cases without
 Learning CRF for OCT Segmentation 5

the need for disease specific modifications. We address this problem with a novel
linearly parameterized Conditional Random Field (LP-CRF) formulation whose
parameters are learnt in a data-driven, end-to-end manner with a Structured
Support Vector Machine (SSVM) [3]. The convolution filters used to capture the
appearance of each layer region and boundary as well as the relative weights of
the appearance and shape prior based cost terms are implicitly modelled within
the LP-CRF model. As a result our method doesnot require any explicit feature
extraction or tunable weights. The contributions of this paper are:
– We eliminate the need to handcraft the energy by learning it from a set of
training images in an end-to-end manner. The proposed method outperforms
the existing methods [4] with similar but handcrafted energy cost terms.
– Our CRF formulation seamlessly incorporates both hard and soft constraints
on shape priors in a single pairwise edge between each neighbor which is
difficult to implement in the MCCS approach [4,13], requiring additional
directed edges in the graph construction.
– We jointly segment all the layers in a single optimization step in contrast to
the existing methods that need to handle each layer differently.
– Our method is able to learn a single energy function for both Normal as well
as AMD cases for a three layer boundary segmentation problem.
– The proposed method has been shown to be robust to data acquired from
multiple centres with different scanners and at different resolutions.

2 Method
An overview of the proposed method is presented in Fig. 2a. The input image
is standardized by extracting the region of interest (ROI), flattening the retinal
tissue and removing speckle noise (see Sect. 2.1). The task of joint extraction of
the multiple layer boundaries is posed as a CRF based Energy Minimization (see
Sect. 2.2). The total CRF energy is learnt in an end-to-end manner by employing
SSVM (see Sect. 2.3). This involves parameterization of the energy as a linear
function LP-CRF (see Sect. 2.4). During testing, a CRF is constructed from
the standardized image using the learnt parameters. Thereafter, the optimal
labelling of the CRF is evaluated and brought back to the original image space
by reversing the image flattening and ROI extraction operations.

Fig. 2. a. Overview of the proposed method, training is represented with dotted lines.
b. Graphical model of the proposed CRF formulation.
6 A. Chakravarty and J. Sivaswamy

2.1 Image Preprocessing

The curvature of the retinal surface is flattened by using the method reported in
[2]. Each image column is shifted by an offset obtained by fitting a quadratic poly-
nomial on a rough estimate of the RP Eout boundary. Next, the ROI is extracted
by cropping out the dark (intensity values close to 0) background regions at the
top and bottom of the image. The ROI is estimated as the rectangular region
encompassing the largest connected component obtained by thresholding the
input image at 0.3 after scaling the intensity to [0,1]. To reduce holes due to the
dark layers within the ROI, a large Gaussian filter with σ empirically set to 9
was employed prior to thresholding. Thereafter, the speckle reducing anisotropic
diffusion [15] is applied. The ROI is resized to 190 × 600 to handle the variations
in image resolution and an intensity standardization scheme based on [10] is
applied to handle the inter and intra-scanner intensity variations.

2.2 Problem Formulation as a CRF

The joint multi-layer segmentation problem seeks to extract L layer boundaries

in an OCT image I of size X × Y . Each boundary is labeled as 1 ≤ l ≤ L
ordered from the top to bottom and represented by N points with coordinates
(xl,n , yn ), obtained by uniformly sampling the image columns at yn positions.
The height along the X-axis where the lth boundary passes through the column
yn is denoted by xl,n . Thus, each xl,n is a discrete random variable that can take
a value from the label set Ω = {1 ≤ i ≤ X, i ∈ Z + }. The set of random variables
X = {xln |1 ≤ l ≤ L, 1 ≤ n ≤ N } defines a random field and x ∈ Ω L×N denotes
a feasible labelling of X obtained by assigning a label from Ω to each xl,n .
A unary boundary cost εlbnd (xl,n ) is defined for each xl,n to capture the like-
lihood of the lth boundary to pass through (xl,n , yn ) given the appearance of the
OCT image around that location. Moreover, based on the Markovian assump-
tion, the label of each xl,n is also considered to be dependent on its immediate
neighbors xl,n+1 on the same boundary and xl+1,n on the adjacent (l + 1)th
boundary which are captured by the pairwise Intra-layer cost εl,n intra (xl,n , xl,n+1 )
and the pairwise Inter-layer cost εl,n (x
inter l,n , xl+1,n ) terms respectively. An undi-
rected graphical representation of the CRF for a 4-neighborhood is depicted in
Fig. 2b. The smoothness and similarity in the appearance of the adjacent points
on a boundary is captured by εl,n l,n
intra (xl,n , xl,n+1 ) while εinter (xl,n , xl+1,n ) cap-
tures the tissue appearance and layer thickness priors between the l and (l + 1)th
layer. Thus, all the three cost terms are dependent on the observed image I. To
simplify the notation, the input arguments for the cost terms have been omit-
ted in the rest of the paper and simply represented by εlbnd , εl,n l,n
intra and εinter
respectively. Thus, the CRF energy for I is defined as
 Learning CRF for OCT Segmentation 7


L 
N  
L N −1 
L−1 N
E(x, I) = εlbnd + εl,n
intra + εl,n
inter
l=1 n=1 l=1 n=1 l=1 n=1
= Ebnd (x, I) + Eintra (x, I) + Einter (x, I), (1)

where Ebnd (x, I), Eintra (x, I) and Einter (x, I) are the sum of all the unary, intra-
layer and the inter-layer cost terms in the entire CRF respectively. The labelling
x that maximizes E(x, I) for an image I corresponds to the desired segmenta-
tion. During implementation, the max CRF inference in Eq. 1 is converted into
a minimization problem by taking the negative of all the unary and pairwise
cost terms, and solved using the Sequential-Tree Reweighted Message Passing
algorithm [6].
Our objective is to parameterize E(x, I) by a set of parameters θ which can
be learnt from a set of training images. Next, we look at the problem of learning
θ in Sect. 2.3 followed by an appropriate definition of Eθ (x, I) in Sect. 2.4.

2.3 The Structured Support Vector Machine Formulation

Let {I (k) , x(k) }K

k=1 denote a set of K training OCT image slices I
(k)
with corre-
sponding ground truth (GT) labelling x(k) . Given a feasible labelling x for I (k) ,
L N (k)
we define a loss function Δ(x(k) , x) = l=1 n=1 | xl,n − xl,n | as sum of the
unsigned distances between the corresponding boundary points across all layers.
The energy function Eθ (x, I) must map all possible labellings x for each
image I to energy values such that the correct labelling corresponds to the
highest energy, i.e., x(k) = argmaxx Eθ (x, I (k) ), ∀k. Moreover, Eθ must tend to
assign higher energy scores to labellings with a lower loss Δ.
SSVM [3] is an extension of the Support Vector Machines that can be applied
to solve the above problem under the assumption that Eθ is a linear function i.e.,
Eθ (x, I) = θ  .F (x, I) where F (x, I) is known as the joint feature function. It
imposes L2 regularization on θ to improve the generalization on unseen images
by solving the following optimization,

1 
M
λ
min || θ || +
2
ξk
θ,ξ≥0 2 M
k=1

s.t. θ  .(F (k) (x(k) ) − (F (k) (x̄(k) )) ≥ Δ(x(k) , x̄(k) ) − ξk

∀k, ∀x̄(k) ∈ Y (k) ,
(2)
where ξk are the slack variables and λ is the regularization parameter which was
fixed to 10−4 in our experiments. The constraints ensure that for each image, the
GT labelling x(k) has a higher score than all incorrect labellings x̄(k) by a margin
scaled by the loss Δ(x(k) , x̄(k) ). Here Y (k) denotes the combinatorially
 large set of
all incorrect labellings for Ik resulting in an extremely large ( k |Y (k) |) number
of constraints making it difficult to be solved directly. In this work, we employed
the Block Co-ordinate Frank Wolfe Algorithm [7] to solve Eq. 2 which replaces
the |Y (k) | constraints for each image Ik by the most violating constraint. We refer
the readers to [7] for more details. The method used to find the most violating
8 A. Chakravarty and J. Sivaswamy

constraint (known as the max oracle) is problem specific and involves solving the
following optimization problem: yk∗ = argmaxx (Δ(x(k) , x̄(k) ) + Eθ (x, I)). Since
in our case, Δ(x(k) , x̄(k) ) is separable at each xln , the max oracle is defined similar
to the CRF inference problem in Eq. 1. with an additional term | xln − x̄ln | added
to the unary terms for each xln .

2.4 Linear Parameterization of CRF Energy

We define the individual cost terms in Eq. 1 as the linear functions denoted
by, Ebnd (x, I) = w 
bnd .Fbnd (x, I), Eintra (x, I) = wintra .Fintra (x, I) and

Einter (x, I) = winter .Finter (x, I) respectively. Then, the net energy is given
by Eθ (x, I) = θ  .F (x, I) where F (x, I) = [Fbnd  
Fintra 
Finter ] and θ  =
  
[wbnd wintra winter ]. The details of the individual cost terms are discussed below.

Boundary Cost: For each layer boundary l, we aim to learn a p × p convolution

filter ul which would have a high response only at the pixels lying on the lth
boundary. Let Il,n represent a p × p image patch centered at (xl,n , yn ). Both ul
and Il,n are linearly indexed to p2 × 1 vectors so that their dot product gives
the filter response at xl,n . Thus, the boundary cost for each xl,n is defined as
εlbnd (xln ) = u
l .Il,n and the total Boundary cost over the entire CRF is


L 
N 
L 
N
Ebnd (x, I) = u
l .Il,n = u
l Il,n = w
bnd .Fbnd (x, I), (3)
l=1 n=1 l=1 n=1

N N
where w  
bnd = [u1 ...uL ] and Fbnd (x, I) = [ n=1 Il,n ... n=1 Il,n ] .
Pairwise Intra-layer Energy: The interaction between the adjacent points
on the lth boundary consists of a shape and an appearance term. The shape
−1 (xl,n+1 −xl,n )−μl,n
prior dl,n
intra (xl,n , xl,n+1 ) = exp{ 2 .( l,n
σintra
) }
intra 2
preserves the bound-
ary smoothness by penalizing large deviations of the signed height gradient
(xl,n+1 − xl,n ) from the Gaussian distributions whose mean μl,n intra and stan-
l,n
dard deviation σintra are separately learnt from the GT of the training images
for each layer l, and column yn .
Emphasising the shape prior can lead to gross segmentation errors in the
presence of abnormalities such as AMD which affect the boundary smoothness.
Hence, an additional appearance-based term is introduced to favour the sim-
ilarity in the appearance of the adjacent boundary points. S(xl,n , xl,n+1 ) =
255
1 − min k=1 min{hk (xl,n , yn ), hk (xl,n+1 , yn+1 )} is a histogram intersection
based similarity measure between the normalized histograms hk of p × p image
patches centered at (xl,n , yn ) and (xl,n+1 , yn+1 ) respectively.
The pairwise intra-layer energy is modelled as a linear combination of the
shape and appearance terms, εl,n l,n l
intra = αl .dintra (xl,n , xn+1 ) + βl .S(xl,n , xl,n+1 ),
where αl and βl are relative weight coefficients learnt automatically in an end-to
end manner during training. Therefore, the total Intra-layer pairwise energy is
 Learning CRF for OCT Segmentation 9

 
L N −1
Eintra (x, I) = (αl .dl,n
intra (xl,n , xl,n+1 ) + βl .S(xl,n , xl,n+1 ))
l=1 n=1

L 
N −1 
L 
N −1
= αl .( dl,n
intra (xl,n , xl,n+1 )) + βl .( S(xl,n , xl,n+1 ))
l=1 n=1 l=1 n=1

= wintra .Fintra (x, I). (4)

Here, Eintra (x, I) is linearized by taking w intra= [α1 α2 ...αL β1 β2 ...βL ] and
N −1
Fintra (x) = [d1 d2 ... dL S 1 S 2 ... S L ] , where di = n=1 di,nintra (xi,n , xi,n+1 ) and
i
 N −1
S = n=1 S(xi,n , xi,n+1 ) respectively.
Pairwise Inter-layer Energy: It captures the interaction between the adjacent
layer boundaries by employing a shape and an appearance based cost. The shape
prior dl,n
inter penalizes the deviation of the layer thickness (xl+1,n − xl,n ) from
apriori learnt Gaussian distributions with mean μl,n inter and standard deviation
l,n
σinter for each column yn . Moreover, hard constraints are also imposed on the
minimum Tmn l
and maximum Tmx l
layer thickness of each layer l by assigning ∞
to the infeasible labellings. Therefore,
 l,n
(xl+1 l
n −xn )−μinter 2
l,n exp{ −1
2 .( l,n ) }, if Tmnl
≤ (xl+1,n − xl,n ) ≤ Tmx
l
dinter = σinter (5)
∞, otherwise.

The parameters μl,n l,n l l

inter , σinter , Tmn and Tmx are learnt for each layer from the GT
l
of training images. Tmn > 0 ensures that the layer boundaries donot intersect.
Let Rl denote the tissue region between the l and (l + 1)th boundary. To
capture the appearance of each Rl , a q × q convolution filter vl is defined for
each of the  L − 1 regions such that vl has a high average response given by
1 xl+1,n 
|xl+1,n −xl,n | ( j=xl,n vl .Ij,n ) in each column yn within the layer Rl .
εl,n
inter is defined as a linear combination of the appearance and shape cost
xl+1,n 
terms, εl,n
inter = 1
|xl+1,n −xl,n | ( l,n
j=xl,n vl .Ij,n ) + γl .dinter (xl,n , xl+1,n ), where γl
provides a relative weight between the two terms. Both vl and γl are learnt in
an end-to-end manner. Therefore, the total Inter-layer pairwise Energy is
⎧ ⎛ ⎞ ⎫

L−1 N ⎨ 
xl+1,n ⎬
1 ⎝ ⎠ + γl .dl,n
Einter (x, I) = vl .Ij,n
⎩ | xl+1,n − xl,n | inter
⎭
l=1 n=1 j=xl,n
⎧ ⎫ N 

L−1 ⎨ N
1 
xl+1,n ⎬ L−1  l,n

= vl Ij,n + γl dinter
⎩ | xl+1,n − xl,n | ⎭
l=1 n=1 j=xl,n l=1 n=1

= w
inter .Finter (x, I) (6)

Here, Einter (x, I) is linearized by taking w inter = [v1 v

2 ...vL−1 γ1 γ2 ...γL−1 ] and
Finter (x) = [r1 r2 ... rL−1 t1 t2 ... tL−1 ] , where ti =
N i,n
  n=1 dinter (xi,n , xi+1,n )
x
and ri = n=1 |xi+1,n1−xi,n | j=xi,n Ij,n respectively.
N i+1,n
10 A. Chakravarty and J. Sivaswamy

3 Results
Dataset: The proposed method was evaluated on the macular OCT B-scans
of both Normal and AMD subjects using two datasets, Normal dataset [2] and
AMD dataset [1] kindly provided by Chiu et al.
The Normal dataset consists of 107 B-scans obtained from 10 OCT volumes.
The B-scans in 5 volumes are of size 300 × 800 pixels with a pixel resolution
of (3.23, 6.7, 67) µm/pixel along the axial, lateral and azimuthal directions. The
AMD dataset is characterized by the presence of drusen and geographic atro-
phy. It consists of 220 B-scans sampled from 20 OCT volumes collected from 4
different clinics. All B-scans were of size 512 × 1000 pixels with pixel resolution
varying in the range of (3.06–3.24, 6.50–6.60, 65–69.8) μm/pixel.
In both datasets, eleven B-scans are linearly sampled from the OCT volumes
such that the sixth B-scan is centered at the fovea. The GT comprises of manual
markings by a senior grader. While GT for all 8 layer boundaries (see Fig. 1a.)
are available for the Normal dataset, GT of only 3 boundaries (see Fig. 1b.) are
available for the AMD dataset.
Evaluation Metrics and Benchmarking: The accuracy of the proposed
method was analyzed in terms of the boundary localization error (BLE) for each
layer boundary, the Dice coefficient (DC) and the retinal Layer Thickness Error
(LTE) for the segmented tissue regions. BLE is defined as the average (signed or
unsigned) distance in pixels between the extracted boundary and the GT along
each column (A-scan) in the image. DC measures the extent of overlap between
the extracted and GT tissue regions. LTE is defined as the average absolute dif-
ference in the thickness (in pixels) between the extracted and GT tissue region
across each column in the image. While DC provides a global measure of the
accuracy, LTE is more sensitive to localized inaccuracies at each column. Ideally,
BLE, LTE should be 0 and DC should be 1.
The proposed method was benchmarked against the results obtained from
three publicly available OCT segmentation software: (A) CASEREL1 based on
[2], (B) the Iowa Reference Algorithm (IRA) based on [4] and (C) OCTSEG2
based on [8]. Each software extracts different number of layers and the RP Ein
boundary which is critical for AMD detection is only available in IRA. Addi-
tionally, the IN L/OP L boundary is also unavailable in OCTSEG.
Experimental Setup: The experiments were carried out in two settings. In
Experiment-1, the proposed method was evaluated on the task of the joint seg-
mentation of all the 8 layer boundaries on the Normal dataset. In Experiment-2,
a single CRF model was learnt on the combined Normal and AMD dataset for
extraction of the three layers, ILM, RP Ein and RP Eout as the GT of only these
three boundaries were available for the AMD dataset. In both experiments, yn
was sampled 4 pixels apart for computational tractability and the intermediate
boundary points were obtained using b-spline interpolation. The size of both ul

1
http://pangyuteng.github.io/caserel/.
2
https://www5.cs.fau.de/research/software/octseg/.
 Learning CRF for OCT Segmentation 11

and vl filters were fixed to 11 × 11. A Matlab implementation of our method

took around 9 seconds to process each B-scan on a i7 processor with 8 GB RAM.
Experiment-1: The proposed method was evaluated using a five-fold cross-
validation on the Normal dataset. In each fold, the B-scans from two OCT
volumes were used for testing while the B-scans from the remaining 8 volumes
were used for training. The boundary and region based metrics for each layer
are provided in Tables 1 and 2, respectively. The proposed method consistently
outperforms other methods for all the 8 boundaries showing a 45% improve-
ment in the average unsigned BLE (1.52/0.29) across all layers compared to
IRA (2.80/0.91). This indicates the advantage of learning the energy in contrast
to the handcrafted cost terms used in IRA based on [4]. Moreover, while IRA
performs the segmentation in 2 steps, namely, the outer (1, 7, 8) layer bound-
aries followed by the inner layers (2–6), the proposed method extracts all the

Table 1. (mean/std.) of BLE (in pixels) for 8 layer boundaries on the Normal dataset

ILM NFL/GCL IPL/INL INL/OPL OPL/ONL IS/OS RP Ein RP Eout

U nsignedError
CASEREL 1.31/0.40 2.83/2.59 5.02/2.67 5.42/2.59 5.37/4.71 1.72/0.81 — 1.70/0.60
OCTSEG 1.87/3.19 6.56/2.92 3.89/3.92 — 3.48/2.99 1.11/1.49 — 1.11/1.69
IRA 3.10/0.99 3.39/1.89 2.40/1.15 4.06/1.30 3.33/1.21 1.08/0.48 2.19/0.84 2.84/1.06
Proposed 1.09/0.28 1.66/0.64 1.51/0.47 1.68/0.55 1.95/0.81 1.15/0.85 1.47/0.75 1.67/0.76
SignedError
CASEREL −1.02/0.59 1.36/3.14 −3.24/3.96 −3.00/4.26 −3.71/5.68 −0.98/1.04 — 0.13/1.03
OCTSEG −0.36/2.97 0.42/5.32 2.73/1.54 — 1.54/3.35 −0.22/1.42 — 0.01/1.78
IRA −2.99/1.04 −2.88/2.29 −2.10/1.36 −3.98/1.39 −2.54/1.98 0.72/0.70 0.90/2.02 −2.77/1.10
Proposed −0.00/0.73 0.17/1.23 0.00/0.97 0.10/1.30 −0.06/1.63 0.28/1.03 0.01/1.37 −0.21/1.52

Table 2. (mean/std.) of LTE and Dice for 7 tissue regions on the Normal dataset.
Ground Truth mean layer thickness is reported to provide a context for the LTE.

NFL GCL-IPL INL OPL ONL-IS OS RPE

GT layer thickness
8.23 22.63 11.29 8.21 25.45 8.13 9.33
Avg. layer thickness err.
CASEREL 3.31/2.84 5.52/2.51 2.33/0.62 5.03/1.52 5.23/4.30 — —
OCTSEG 6.84/2.74 5.71/1.83 — — 3.36/2.61 — —
IRA 2.39/1.00 2.44/0.94 2.54/0.91 2.67/1.29 3.85/1.25 2.12/0.88 3.87/1.44
Proposed 1.97/0.67 2.06/0.59 1.85/0.53 2.32/0.96 2.26/1.08 1.73/0.77 1.83/0.85
Dice coeff.
CASEREL 0.79/0.13 0.82/0.09 0.60/0.19 0.60/0.16 0.88/0.08 — —
OCTSEG 0.61/0.14 0.79/0.13 — — 0.91/0.06 — —
IRA 0.68/0.11 0.88/0.07 0.71/0.11 0.63/0.11 0.92/0.03 0.82/0.06 0.70/0.07
Proposed 0.84/0.06 0.93/0.02 0.87/0.03 0.80./0.07 0.94/0.02 0.86/0.07 0.85/0.06
Another random document with
no related content on Scribd:
 Lawn Benches Made from Old Bedsteads

Painted Green, These Rebuilt Bedsteads Served as Lawn Benches

Old bedsteads were converted into serviceable lawn, or porch,

benches, as shown in the photograph reproduced, by the addition of
a suitable seat, properly supported. The transformation was a simple
one. Only the foot and headpieces of the bedsteads were used. The
front legs and other pieces were made from other wood. The front
legs are of square stock, about 2¹⁄₂ by 2¹⁄₂ in. The crosspieces,
supporting the wide board seats, are mortised into the legs and
fastened with glue and screws. The seat is fastened from the under
side by cleats. The lumber was carefully planed and sandpapered so
that the benches presented a smooth finish when painted green, to
match other outdoor furniture.—F. E. Tuck, Nevada City, Calif.
 Repairing Wood-Wind Instruments
Wood-wind instruments sometimes “leak” at the joints or keys and
make playing of the instrument difficult. Many such instruments are
made in sections, with ends that telescope to form a tight fit. This fit
is maintained by the use of a cork band cemented around the tenon
end of the telescoping joint. The renewal of these cork joints, and the
addition of new pads on the keys, will make an old instrument nearly
as good as it was when new, so far as playing is concerned,
provided the work is correctly done and the wood of the sections
themselves has not cracked. Many musicians have spare time and
can do this work themselves. The outlay for materials for the job is
from 75 cents to $1.00. A small alcohol, or even a kerosene, lamp
and an old knife, or old file, are required.

The Cork is Fitted Carefully into Place, and Glued

All traces of the old cork on the joint can be removed with
sandpaper, leaving it as shown at the left. The cork comes in strips
of about the proper thickness, and wide and long enough to allow for
trimming. The ends of the strip should be beveled to make a ¹⁄₄-in.
lap joint.
A small quantity of the cement is heated over the lamp and six
drops poured on the joint; then with the end of the file, which should
be heated also, it is spread to give an even, thin coating. The
beveled ends of the strip are similarly treated. By working quickly
and carefully, the coating on the joint and strip are brought to a
plastic state by holding in the flame, and the strip is quickly laid in
place. Before the cement has time to harden, press the cork in,
forming a neat joint. Bind a rag around the cork, leaving it until the
cement is thoroughly set.
The corked joint will be too large to go into the joining section of
the instrument. File and sandpaper it to a twisting fit. Though the
cork should be truly cylindrical, it may be tapered a trifle smaller at
the forward end. A coating of tallow applied to the joint will make it
easy-fitting, but air-tight and moisture-proof.
The pads are disks of felt incased in thin sheepskin. After long
usage, they become too hard to make an air-tight fit. Repadding
should, therefore, be anticipated. Shellac will give good results in
putting on pads. It is heated until liquid and poured into the key
recess. The new pad is pressed into the liquid shellac, care being
taken to have it well centered. For different keys, it will be necessary
to use varying quantities of shellac to make the pad sit higher or
lower, as required.—Donald A. Hampson, Middletown, N. Y.

¶A simple method of bracing a screen door is to stretch a stout wire

diagonally across the lower portion of it.
Rustic Trellis to Shade Door or Window
Rustic Trellises are Easily Constructed and When Covered with Vines Add
to the Attractiveness of the Home
 Proper preparation in the early spring will make it possible for the
householder to shade doors and windows from the hot summer’s
sun by means of inexpensive rustic trellises that add not a little to the
beauty of the home. A suggestion for a trellis at a doorway and one
for a window are shown in the illustration. They are made of straight
tree trunks and small limbs, having the bark on them. The curved
portions of the window trellis may be made easily by using twigs that
are somewhat green. Morning-glories, or other suitable climbing
plants, may be trained over the trellises.—J. G. Allshouse,
Avonmore, Pa.
 Making Scale Enlargements with a Rubber Band
For reducing or enlarging maps, and similar drawings of irregular
design, the device shown in the illustration will replace the ordinary
instruments, and enable the draftsman to turn out a given amount of
work in much less time than required when proportional dividers are
used. The materials needed are an eraser, a rubber band, two pins,
two thumb tacks, and a few drops of rubber cement. From the eraser
two pieces are cut, as shown in the sketch, about ¹⁄₄ by ¹⁄₂ by 1¹⁄₄ in.
Cut deep slits in each end of these pieces. Insert the end of the
rubber band, cut at the splice, in one of these slits and place a thumb
tack in the other. A pin is thrust through the eraser and trimmed
close, to prevent the thumb tack from tearing the eraser. Cement the
slits with rubber cement, and place the assembled device under a
book weight, until the cement has set.
 This Simple Device Is Useful in Enlarging or Reducing Drawings and Maps

Assuming that a contour map is to be enlarged, the rectangular

divisions of the original map, ordinarily section lines or the
boundaries of quarter sections, are drawn on the larger sheet as a
base for the reproduction. Place the device on the original map, as
indicated, the edge of the rubber band touching a “horizontal” section
line between two “vertical” ones, the rubber band under slight
tension. On the black surface of the band, dot white points, with
water color, along the section line at which the contour lines intersect
it. Also place a dot at each end of the band to indicate the position of
the two “vertical” section lines between which the band is set.
Transfer the device to the same relative position on the
enlargement, stretching the rubber band. Make dots at each end,
denoting the “vertical” section lines, for the corresponding lines on
the enlargement. The series of intermediate points along the band
will be in the same relative position on the enlargement as they were
on the original. They can be connected on the enlargement with as
accurate a result as obtained by the use of proportional dividers, and
more rapidly.
After the points are indicated upon the enlargement, the
reproducing device is removed and the surface of the rubber band
cleaned instantly by touching it with a moist cloth. The exposed part
of the rubber band is a variable, and the device can be made with
this dimension adapted to the work. It is capable of enlarging or
reducing at a ratio not greater than six to one, above which the
rubber band approaches its elastic limit.—H. L. Wiley, Seattle, Wash.
 Signal Telegraph with Green and Red Lights

By arranging a circuit with batteries, lights, and keys, as shown in

the diagram, a signal telegraph may be made that will afford much
pleasure to boys and may be used for practical purposes. The keys
A and B are wired into the circuit with a battery C and a red and a
green incandescent lamp. A simple set of signals may be devised
easily so that messages may be sent in the code.—James R.
Townsend, Itasca, Texas.
 A Circular Swing
By DAVIS FOSS GETCHELL

W hile on the farm I constructed a circular swing which proved very

attractive to my boys and their friends. By its side, and
suspended from the same tree branch, was an ordinary swing.
During the eight weeks of our stay the latter was seldom in use. The
circular swing was a far greater favorite with all the young people,
boys and girls alike.
Around a branch of a large elm and 18 or 20 ft. from the tree trunk
was looped a 10-ft. length of chain and to the hanging end of this
was made fast a 1-in. rope nearly 10 ft. longer than was needed to
reach the ground. Directly beneath the point where the chain went
around the limb, as determined by a plumb bob, was set a 6-in.
piece of cedar post 3¹⁄₂ ft. into the ground. This was sawed off
square 2¹⁄₂ ft. above the ground. Into the top of this post was set a
¹⁄₂-in. rod, to serve as a pivot for the swing. It was set in firmly about
6 in. and projected about 3 in. from the top of the post.
 The Circular Swing will be Found Very Safe and Pleasurable, but, as is the
Case of an Ordinary Swing, Anyone Careless Enough to Get in the Way of It
will Get Badly Bumped

A straight-grained piece of pine board, 15 ft. long, 8 in. wide, and 1

in. thick, was procured and a hole bored in one end large enough to
make it turn freely on the pin in the upper end of the post. Two holes
were bored in the other end of the board large enough to admit the
rope. The first hole was 6 in. from the end, and the second hole, 3 ft.
The hanging end of the rope was passed down through one of these
holes and back up through the other and then made fast to itself
about 3 ft. above the board after the board had been adjusted so that
it would swing throughout its length at the height of the post, or 2¹⁄₂
ft. from the ground. The swing was then complete except for a
swivel, which was put in the rope within easy reach of one standing
on the board, so that it could be oiled.
One good push would send the board with a boy on the end three
or four times about the 90-ft. circle. The little fellows would like to get
hold of the board in near the post and shove it around. Once started,
it could be kept going with very little effort.
In putting up such a swing, make sure to have the post set solidly
in the ground, as it has a tendency to work loose. Tie all the knots
tightly. Do not look upon the swivel as unnecessary. The first swing I
put up was without one, and the rope twisted off in a few days.
It is not necessary to climb a tree; just throw a stout cord over the
limb by means of a stone or nut tied to the end, then haul the rope
and chain up over the limb with the cord. Before the chain leaves the
ground loop the end of it and pass the cord through the loop. The
higher the limb from the ground the better the swing will work, but 25
ft. will be about right.
 Hand-Operated Motorboat Whistle

Bellows Operated by Hand for Blowing a Whistle on a Power Boat

Anyone with a power boat can construct a blower for the whistle
very cheaply. The whistle is attached to a suitable length of pipe,
threaded on each end. The blower is made of two white-pine boards,
1 in. thick, cut as shown at A; a thin piece of leather is cut like the
pattern B, to form the bellows part, and after it is shaped, the edges
of the boards are glued and the leather placed in position, where it is
fastened with tacks driven in about 1 in. apart. The bellows are
fastened to the under side of a seat with screws, and a tension
spring is attached to the bottom of the bellows and the floor of the
boat. A cord is fastened to the lower board of the bellows and run up
through to the cabin roof over suitable pulleys to a handle within
convenient reach of the operator.—Contributed by John I. Somers,
Pleasantville, N. J.
 Filling In Broken Places on Enamel
Ordinary putty will not do to fill in cracks or broken spots on an
enameled surface, such as a clockface. Fine sealing wax is much
better, as it hardens at once, takes color without absorbing the oil,
and does not shrink like putty. Use a wax of the proper color to
match the surface as closely as possible. Fit it in and smooth with a
warm, flexible piece of metal, such as a palette knife. Give it one or
two coats of thin color to exactly match the other surface, and
varnish. If the article has not a high polish, the gloss of the varnish
can be cut a little with pumice stone.
 A Twisting Thriller Merry-Go-Round
By R. E. EDWARDS

“Stepdime!”
right up; three twisting thrillers for a penny—a tenth of a
was the familiar invitation which attracted customers to
the delights of a homemade merry-go-round of novel design. The
patrons were not disappointed, but came back for more. The power
for the whirling thriller is produced by the heavy, twisted rope,
suspended from the limb of a tree, or other suitable support. The
rope is cranked up by means of the notched disk A, grasped at the
handle B, the car being lifted off. The thriller is stopped when the
brakeplate I rests on the weighted box L.
 The Supporting Ropes are Wound Up at the Disk A, the Car is Hooked into
Place, and the Passengers Take Their Seats for a Thrilling Ride, Until the
Brakeplate I Rests on the Box

Manila rope, ³⁄₄ in. or more in diameter, is used for the support,
and is rigged with a spreader, about 2 ft. long, at the top, as shown.
The disk is built up of wood, as detailed, and notches, C, provided
for the ropes. The rope is wound up and the car is suspended from it
by the hook, which should be strong, and deep enough so that it
cannot slip out, as indicated at H.
The car is made of a section of 2 by 4-in stuff, D, 10 ft. long, to
which braces, E, of 1 by 4-in. stuff are fastened with nails or screws.
The upper ends of the pieces E are blocked up with the centerpiece
F, nailed securely, and the wire link G is fastened through the joint.
The seats J are suspended at the ends of the 2 by 4-in. bar, with
their inner ends lower, as shown, to give a better seating when the
thriller is in action. The seats are supported by rope or strap-iron
brackets, K, set 15 in. apart. The box should be high enough so that
the seats do not strike the ground.