Interpretable Deep Learning Frameworks for Multi-Source Image Analysis of Diabetic Retinal Pathologies
Abstract: Integrating artificial intelligence (AI) and computer vision into medical imaging has led to transformative advances in diagnostic healthcare, particularly within ophthalmology. Among various eye-related diseases, Diabetic Retinopathy (DR) is one of the most prevalent and severe complications of diabetes mellitus, posing a leading cause of blindness globally. Early identification and precise classification of Diabetic Retinopathy (DR) are crucial for effective intervention and treatment planning. However, the increasing volume of retinal images that need to be analyzed and the scarcity of expert ophthalmologists necessitate the development of reliable automated screening systems. This thesis introduces a comprehensive deep learning framework designed to enhance the reliability and scalability of automated Diabetic Retinopathy (DR) diagnosis by addressing three fundamental challenges: image quality assessment (first), referable DR classification (second), and interpretability of both image quality and diagnostic outputs (third).
First, the accuracy of automated Diabetic Retinopathy (DR) screening systems is highly dependent on the quality of the input retinal images. Suboptimal images—caused by blur, poor illumination, or patient movement—can significantly degrade model performance, leading to misdiagnoses. We develop an automated image quality assessment (IQA) module based on convolutional neural networks (CNNs) to mitigate this. This module performs a critical pre-screening step to identify and filter out low-quality images, ensuring that only diagnostically useful images proceed to the classification stage. The model is trained using expert-labeled data, capturing a wide range of quality variations, and demonstrates high reliability in distinguishing between acceptable and non-acceptable image inputs. This step enhances the trustworthiness and clinical utility of downstream diagnostic predictions.
Second, the core component of this research focuses on developing a deep learning-based classification system for referable diabetic retinopathy. Traditional Diabetic Retinopathy (DR) diagnosis relies on manual assessment by specialists, which is both time-intensive and subject to inter-observer variability. Our proposed system leverages deep CNN architectures to automatically learn hierarchical, discriminative features from raw fundus images, enabling accurate classification across all Diabetic Retinopathy (DR) severity levels. The model is trained on publicly available and proprietary datasets with annotations for both referable and non-referable cases. The model achieves robust performance through careful tuning, data augmentation, and optimization, especially in identifying cases requiring referral to an ophthalmologist. This automated classification system offers a scalable solution for large-scale diabetic eye disease screening and can significantly reduce the burden on healthcare professionals.
Third, while high-performing deep learning models are increasingly adopted in medical contexts, their lack of transparency remains a barrier to widespread clinical deployment. Therefore, this work strongly emphasizes model interpretability for image quality assessment (IQA) and Diabetic Retinopathy (DR) classification tasks. To this end, we incorporate explainability techniques such as class activation maps (CAMs), Grad-CAM, and saliency maps to visualize which regions of the retinal image contributed most to the model’s decision. For image quality assessment, clinicians can understand why an image was rejected—e.g., due to peripheral blur or underexposure. For Diabetic Retinopathy (DR) classification, these visualization tools highlight pathological features like microaneurysms, haemorrhages, or exudates, providing a clinically meaningful rationale for each prediction. These interpretability components enhance clinician trust fac, facilitate model validation, and align the system with emerging regulatory requirements for transparent AI in healthcare.
In summary, this thesis proposes an end-to-end framework for automated Diabetic Retinopathy (DR) detection that integrates a deep learning-based image quality assessment module (first), an accurate and scalable referable Diabetic Retinopathy (DR) classification system (second), and interpretable visual feedback mechanisms (third). By ensuring high-quality inputs, delivering precise diagnostic outputs, and enabling transparent decision-making, the proposed approach contributes to developing robust, clinically applicable AI tools for diabetic eye disease management. This work represents a significant step toward realizing intelligent, trustworthy, and accessible Diabetic Retinopathy (DR) screening systems suitable for real-world healthcare environments.