Towards Generalizable Diabetic Retinopathy Grading: A Multi-Dataset Cross-Validation Study Using Hybrid Deep Learning Across APTOS, IDRiD, Messidor, and EyePACS

DOI

10.2991/978-94-6239-821-4_10How to use a DOI?

Keywords

Diabetic Retinopathy; Multi-Dataset Validation; Domain Generalisation; EfficientNet; Swin Transformer; Cross-Validation; Explainable AI

Abstract

A persistent and largely unacknowledged weakness in deep learning-based diabetic retinopathy (DR) grading is the near-universal practice of training and evaluating models on a single benchmark dataset. This creates a dangerous illusion of clinical readiness: a model achieving 99% accuracy on EyePACS may fail substantially when presented with images from an Indian ophthalmology clinic captured under different lighting, on a different camera, with a different patient demographic. This paper directly addresses this generalisation gap by proposing and investigating a hybrid EfficientNet-B4 and Swin Transformer model evaluated under rigorous five-fold cross-validation across four major benchmark datasets: APTOS 2019, IDRiD, Messidor-2, and EyePACS. We systematically analyse the sources of performance variation across these datasets — including domain shift due to imaging device differences, label shift due to differing grading protocols, and prior shift due to class frequency variation — and introduce a domain-agnostic preprocessing pipeline with contrast-limited adaptive histogram equalisation (CLAHE) and standardised resizing to mitigate these effects. We also integrate Grad-CAM + + for cross-dataset explainability analysis. Our experiments demonstrate that models trained and validated on merged multi-dataset splits consistently outperform single-dataset counterparts on unseen clinical data, with average accuracy improvements of 4.2–7.8% on cross-dataset test splits. These findings establish multi-dataset validation as a necessary standard for clinically trustworthy DR grading research.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Dhumal Surekha Vitthalrao
AU  - Mayur Jakhete
PY  - 2026
DA  - 2026/10/31
TI  - Towards Generalizable Diabetic Retinopathy Grading: A Multi-Dataset Cross-Validation Study Using Hybrid Deep Learning Across APTOS, IDRiD, Messidor, and EyePACS
BT  - Proceedings of the International Conference on Information Systems Design And Intelligent Applications (ICISDA 2026)
PB  - Atlantis Press
SP  - 98
EP  - 112
SN  - 1951-6851
UR  - https://doi.org/10.2991/978-94-6239-821-4_10
DO  - 10.2991/978-94-6239-821-4_10
ID  - Vitthalrao2026
ER  -