Date of Award

3-10-2025

Document Type

Doctoral Thesis

Degree Name

Doctor of Philosophy

First Advisor

Dr. Ruairi O’Reilly

Second Advisor

Dr Mubashir Husain Rehmani

Abstract

Artificial intelligence (AI) plays a pivotal role in computer-aided diagnosis. In image-based diagnosis, AI systems rely on biomedical images containing sensitive disease-oriented information. AI systems utilize machine learning models to automate this diagnosis. Machine learning models such as deep neural networks require large training datasets to enhance the diagnostic precision of these models. Models can underperform when they are trained on imbalanced datasets. Imbalance in datasets refers to the skewed numbers of images per image class. There is a need to enhance the training of these models via data augmentation. In the context of biomedical imagery, where acquiring large, balanced datasets is often infeasible due to privacy concerns, cost, and rarity of specific disease conditions, Generative Adversarial Networks (GANs) have emerged as powerful tools for data augmentation. GANs learn image features’ distribution of the original data to produce synthetic images representative of that data to address data imbalances. However, despite their potential, GANs experience problems in training such as mode collapse (limiting a GAN’s ability to generate diversified images), non-convergence (imbalanced state deviating GAN training from Nash equilibrium), and instability (unstable training due to a vanishing gradient). These problems in GANs present a distinct challenge for biomedical image synthesis as they can significantly degrade the diversity and quality of generated images. To date, the existing literature on GANs has overlooked these training problems individually in generating synthetic biomedical images. This PhD thesis investigates these training challenges by identifying their root causes and quantifying them systematically for biomedical imagery. Evaluation metrics associated with assessing the output requiring highly scientific rigor to enable a better representation of the loss, diversity, and quality of synthetic images are explored. Pre-processing techniques and architectural modifications in GANs to address the training challenges are proposed. A novel early stopping criterion is introduced to minimize computational costs during training ensuring the generation of diverse and high-quality synthetic images. Furthermore, this thesis investigates the utility of diverse and high-quality synthetic images generated from proposed GANs in augmenting imbalanced datasets for image classification, enhancing deep learning model performance on augmented datasets.

Access Level

info:eu-repo/semantics/openAccess

Download

Share

COinS