Interpretable Generative Anomaly Detection for Diabetic Retinopathy using Fundus Images

View at publisher Summarise with VIRAT
Authors: Vipin Kumar Bansal, Manisha Malhotra
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Diabetic Retinopathy (DR) is one of the significant complications of diabetes and a leading cause of vision impairment. In many cases, vision loss due to DR is preventable if the disease is diagnosed at an early stage. Reliable screening and timely intervention are therefore critical. Besides manual analysis of retinal screening images, several automated approaches based on artificial intelligence and deep learning have evolved due to their promising performance in DR detection. However, most existing AI-based solutions are supervised and provide only final predictions, without explaining the underlying reasoning. This lack of interpretability is a key limitation and significantly reduces clinical acceptance, primarily because of a lack of trust in such black-box systems. Generative Anomaly Detection (AD), based on the principles of Generative AI, aims to either generate entirely new content or reconstruct existing data. It has emerged as a strong alternative to supervised approaches, particularly for image-based anomaly detection tasks. This paradigm is desirable in scenarios where obtaining annotated data is difficult or costly. However, despite its advantages, generative AD also has notable limitations. Although generated or reconstructed images can be visually inspected, ensuring explainability remains challenging because generative models often operate as black-box systems.This work proposes an explainable generative anomaly detection framework for diabetic retinopathy (DR) screening using fundus images, trained exclusively on healthy data. This framework introduces a multi-level explainability mechanism that combines both qualitative and quantitative i.e. pixel-level reconstruction-based anomaly localisation along with feature-level latent-space interpretability respectively. Qualitatively, it produces clinically meaningful visual explanations aligned with ophthalmological reasoning. Quantitatively, the stastical methods achieve strong anomaly detection and lesion-level localisation performance on public datasets, outperforming post-hoc explainability approaches. The main contributions include an explainable generative DR screening model, integrated multi-level explanations, and comprehensive qualitative and quantitative validation demonstrating improved trustworthiness and clinical relevance.

Fields of Study & Subjects

Retinal Imaging and Analysis Explainable Artificial Intelligence (XAI) Anomaly Detection Techniques and Applications

Publication details

Journal / source
International Conference on Image Processing and Capsule Networks (ICIPCN)
Pages
1207–1213
Publication date
27 January 2026
Language
EN