DEEP LEARNING-BASED MODEL FOR THE IDENTIFICATION OF RETINAL DETACHMENT

Abstract

Retinal detachment (RD) is an ocular emergency condition that requires urgent diagnosis, as delays in diagnosis and treatment can result to rapid and permanent loss of vision. Currently, there is no established confirmatory marker for early RD diagnosis, and clinical detection relies heavily on the judgment of ophthalmologists, often at a later stage when invasive surgical intervention becomes the only treatment option. To address this gap, a scoping review was conducted to systematically map existing literature on artificial intelligence (AI) applied to RD detection using medical images. Findings from the review indicate that while AI has shown promise, most studies included RD as part of multi-disease classification tasks rather than focusing exclusively on RD detection, and there remains a lack of standardized datasets and optimized model architectures specific to RD. Building on these gaps, this thesis proposes RDNet, a novel deep learning-based model based on the SqueezeNet architecture, trained on a publicly available dataset of 1,693 fundus images consisting of rhegmatogenous RD and non-RD cases. The final model achieved 97.55% sensitivity, 99.26% specificity, and 98.23% accuracy, outperforming existing models with an AUC of 0.995. These results demonstrate that AI-based models—when carefully designed and trained on clinically relevant data—can assist in early and reliable RD detection and have the potential to support ophthalmologists in timely decision-making, thereby reducing the need for late-stage invasive interventions and improving patient outcomes.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering