MEDICAL IMAGE ANALYSIS OF VERTEBRAE AND LESIONS USING CONVOLUTIONAL NEURAL NETWORKS

MPhil Thesis Defence


Title: "MEDICAL IMAGE ANALYSIS OF VERTEBRAE AND LESIONS USING 
CONVOLUTIONAL NEURAL NETWORKS"

By

Miss Han ZHANG


Abstract

Deep learning has been very successful in different tasks with natural 
images in recent years. Furthermore, analyzing vertebrae and lesions in 
medical images seems challenging because of the high accuracy requirements 
and the variability of inter- and intra-class disparities in lesions from 
different organs. Therefore, the effectiveness of convolutional neural 
networks (CNN) needs to be further optimized for medical image analysis 
tasks. The methods and techniques proposed in this thesis can be 
classified into two main categories: methods for scoliosis assessment and 
methods for multi-organ lesion detection. Both proposed methods can 
improve the precision of diagnosis in Computer-Aided Diagnosis (CADx) 
systems.

Computer tomography (CT) scans’ capabilities in detecting lesions have 
been increasing remarkably in the past decades. Thus, researchers are no 
longer satisfied with the conventional CADx systems, which can detect a 
single kind of diseases alone. Therefore, we propose a multi-organ lesion 
detection (MOLD) approach to better address real-life chest- related 
clinical needs. MOLD is a challenging task, especially within a large, 
high resolution image volume, due to various types of background 
information interference and large difference in lesion sizes. 
Furthermore, the appearance similarity between lesions and other normal 
tissues demands more discriminative features. In order to overcome these 
challenges, we introduce depth-aware (DA) and skipped-layer hierarchical 
training (SHT) mechanisms with the novel Dense 3D context enhanced (Dense 
3DCE) lesion detection model. The novel Dense 3DCE framework considers the 
shallow, medium, and deep- level features together comprehensively. In 
addition, equipped with our SHT scheme, the backpropagation process can 
now be supervised under precise control, while the DA scheme can 
effectively incorporate the depth domain knowledge into the scheme. 
Extensive experiments have been carried out on a publicly available, 
widely-used DeepLesion dataset, and the results prove the effectiveness of 
our DA-SHT Dense 3DCE network in the MOLD task.

Spinal diseases are common and difficult to cure, which causes much 
suffering. Accurate diagnosis and assessment of these diseases can 
considerably improve cure rates and the quality of life for patients. The 
spinal disease assessment relies primarily on accurate vertebra landmark 
detection, such as scoliosis assessment. However, existing approaches do 
not adequately exploit the relationships between vertebrae and analyze the 
global spine structure, meaning scarcity annotations are underutilized. In 
addition, the practical design of ground-truth is also deficient in model 
learning due to the sub- optimal coordinate system. Therefore, we propose 
a unified end-to-end vertebra land- mark detection network called 
Dcor-VLDet, contributing to the scoliosis assessment task. This network 
takes the positional information from within and between vertebrae into 
account. At the same time, through fusing the advantages of both Cartesian 
and polar coordinate systems, the symmetric mean absolute percentage error 
(SMAPE) value can be reduced significantly in scoliosis assessment. The 
experimental results demonstrate that our proposed method is superior in 
measuring Cobb angle and detecting landmarks on low-contrast X-ray 
images.


Date:  			Tuesday, 23 August 2022

Time:			4:00pm - 6:00pm

Zoom Meeting:
https://hkust.zoom.us/j/96539549972?pwd=V3VRb2prUmNJYnM3Ti9wL1c0WkZQZz09

Committee Members:	Prof. Albert Chung (Supervisor)
 			Prof. Pedro Sander (Supervisor)
 			Dr. Dan Xu (Chairperson)
 			Dr. Hao Chen


**** ALL are Welcome ****