Date of Defense
12-5-2025 11:00 AM
Location
F1–1164
Document Type
Thesis Defense
Degree Name
Master of Science in Electrical Engineering (MSEE)
College
COE
Department
Electrical and Communication Engineering
First Advisor
Prof. Nabil Albastaki
Keywords
Ocular diseases, data augmentation, generative adversarial networks (GANs), deep learning, Image classification, region of interest (ROI).
Abstract
Automatic detection of ocular diseases helps medical professionals efficiently identify eye disorders, reduce diagnostic errors, and accelerate diagnoses to prevent blindness. Deep learning has been successfully utilized in various fields, including medical image classification. However, in spite of these advancements, challenges remain in ocular disease classification.
/="/">The objective of this work is to address these challenges using data processing, data augmentation in combination with Region of Interest (ROI) techniques. Medical datasets often suffer from scarcity, imbalance, and low-quality images, leading to inaccurate classification. To mitigate these issues, we utilize the ODIR dataset, which contains 7,000 labelled training images for both left and right eyes, and propose a data augmentation method using a Generative Adversarial Networks (GANs) algorithm. This method aims to augment and balance the fundus images across eight different categories of ocular diseases, including normal fundus images. The generated images were then used to train transfer learning models, with 3,000 generated images per category. The data was split into 80% (2,400) for training, 10% (300) for testing, and 10% (300) for validation. Using only generated images, the testing accuracy of a transfer learning model Inception V3 – after training for 25 epochs, improved significantly to about 95% for eight eye disease categories. We further trained the filtered and separated ocular disease images using an unsupervised model – StyleGAN – for 160 to 260 iterations. A portion of the 2,400 generated images was then used to train transfer learning models to avoid overfitting and enhance accuracy. Additionally, we applied ROI and post-processing techniques to enhance the images. Using only real images for the binary glaucoma test resulted in an improvement of approximately 10% compared to unmodified images which achieved only 72% testing accuracy.
/="/">We have demonstrated that augmenting medical datasets with GANs, combined with the use of appropriate ROI techniques tailored to the characteristics of two categories of eye diseases, significantly improves classification accuracy.
/="/">This comprehensive study includes numerous experiments using ocular disease images, ranging from imbalanced, unmodified, non-augmented data to balanced, augmented, and edited data. The study employs multi-approach methods with a standard vi benchmark, using only real, unadjusted images as testing data, and hence highlighting the challenges and uniqueness of this work.
Included in
A STUDY OF THE IMPACT OF BALANCING, GEOMETRIC TRANSFORMATION, GENERATIVE NETWORKS AUGMENTATION, AND ROI TECHNIQUES IN EYE DISEASES CLASSIFICATION
F1–1164
Automatic detection of ocular diseases helps medical professionals efficiently identify eye disorders, reduce diagnostic errors, and accelerate diagnoses to prevent blindness. Deep learning has been successfully utilized in various fields, including medical image classification. However, in spite of these advancements, challenges remain in ocular disease classification.
/="/">The objective of this work is to address these challenges using data processing, data augmentation in combination with Region of Interest (ROI) techniques. Medical datasets often suffer from scarcity, imbalance, and low-quality images, leading to inaccurate classification. To mitigate these issues, we utilize the ODIR dataset, which contains 7,000 labelled training images for both left and right eyes, and propose a data augmentation method using a Generative Adversarial Networks (GANs) algorithm. This method aims to augment and balance the fundus images across eight different categories of ocular diseases, including normal fundus images. The generated images were then used to train transfer learning models, with 3,000 generated images per category. The data was split into 80% (2,400) for training, 10% (300) for testing, and 10% (300) for validation. Using only generated images, the testing accuracy of a transfer learning model Inception V3 – after training for 25 epochs, improved significantly to about 95% for eight eye disease categories. We further trained the filtered and separated ocular disease images using an unsupervised model – StyleGAN – for 160 to 260 iterations. A portion of the 2,400 generated images was then used to train transfer learning models to avoid overfitting and enhance accuracy. Additionally, we applied ROI and post-processing techniques to enhance the images. Using only real images for the binary glaucoma test resulted in an improvement of approximately 10% compared to unmodified images which achieved only 72% testing accuracy.
/="/">We have demonstrated that augmenting medical datasets with GANs, combined with the use of appropriate ROI techniques tailored to the characteristics of two categories of eye diseases, significantly improves classification accuracy.
/="/">This comprehensive study includes numerous experiments using ocular disease images, ranging from imbalanced, unmodified, non-augmented data to balanced, augmented, and edited data. The study employs multi-approach methods with a standard vi benchmark, using only real, unadjusted images as testing data, and hence highlighting the challenges and uniqueness of this work.