P Sivaprakash¹ , S Thenmozhi², Prathima Mabel J³, B Shuriya⁴

¹Department of Computer Science and Engineering (AI &ML), Rathinam Technical Campus, Coimbatore; ²Department of Artificial Intelligence & Data Science, Kalaignar Karunanidhi Institute of Technology, Coimbatore; ³Department of Information Science and Engineering, Dayananda Sagar College of Engineering, Karnataka; ⁴Department of Computer Science and Engineering, United Institute of Technology, Coimbatore, India.

For correspondence:-  P Sivaprakash   Email: sivaprakashcse0402@gmail.com

Received: 21 September 2025        Accepted: 12 December 2025        Published: 28 December 2025

Citation: Sivaprakash P, Thenmozhi S, J PM, Shuriya B. Breaking boundaries in diabetic retinopathy classification: A novel domain generalization approach with vision transformers for enhanced calibration and safety in medical imaging. Trop J Pharm Res 2025; 24(12):1479-1491 doi: https://dx.doi.org/10.4314/tjpr.v24i12.3

© 2025 The authors.
This is an Open Access article that uses a funding model which does not charge readers or their institutions for access and distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0) and the Budapest Open Access Initiative (http://www.budapestopenaccessinitiative.org/read), which permit unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited..

Abstract

Purpose: To develop a new domain generalization (DG) framework for diabetic retinopathy (DR) classification using Vision Transformers (VTs). Methods: A deep-learning-based prediction-softening mechanism has been used to allow self-distillation, whereby the knowledge extracted is transferred to intermediate layers. Intermediate representations are further improved with an adaptive convex combination of one-hot labels and internal predictions of the model, which helps regularize the network and counteract overfitting. Various publicly available DR datasets, such as APTOS and EYEPACS, in multi-source and single-source DG settings containing three VT backbones (DeiT, T2T-VT, and CvT) were used. Results: The proposed framework outperforms current DG methods, with better top-1 accuracy, strong calibration and predictive accuracy in unfamiliar domains. Conclusion: The Softening Predictions for Self-Distillation (SPSD)-VT framework is a robust, reliable, and generalizable framework that classifies DR and underscores the important role that domain generalization plays in medical imaging and sets a standard for future studies.

Keywords: Domain generalization, Classification, Vision transformers, Diabetic retinopathy, Deep learning

Impact Factor
Thompson Reuters (ISI): 0.6 (2023)
H-5 index (Google Scholar): 49 (2023)