Author

Supervisor

Dr Muhammad Iqbal

Programme

BSc (Hons) in Computing in IT

Subject

Computer Science

Abstract

Retinal diseases such as Choroidal Neovascularization, Diabetic Macular Edema and Drusen can affect vision if they are not detected and treated early. Optical Coherence Tomography images provide detailed cross sectional views of the retina, making them useful for identifying retinal abnormalities. This project investigates the use of deep learning for classifying retinal images from the OCT2017 dataset into four categories: CNV, DME, DRUSEN, and NORMAL. The project follows the CRISP-DM methodology, covering data understanding, data preparation, modelling, evaluation and deployment. During data preparation, duplicate images were removed, image dimensions were standardised and class imbalance was addressed using class weights during model training.

Two deep learning approaches were developed and compared a Custom Convolutional Neural Network trained from scratch and a ResNet50 model using transfer learning. Hyperparameter tuning was performed for the Custom CNN using Keras Tuner Hyperband and both models were trained and evaluated using till 30 epochs. The Custom CNN was selected as the final model because it achieved the best overall balance between validation performance and test performance, ResNet50 also performed well, but it was slightly lower than the Custom CNN. Finally, the selected model was deployed using Gradio in Google Colab, allowing users to upload an OCT image and receive a predicted class, confidence score and Grad-CAM heatmap to visualise the model’s focus area.

Date of Award

2026

Full Publication Date

2026

Access Rights

open access

Document Type

Undergraduate Project

Resource Type

bachelor thesis

Share

COinS