Diabetic retinopathy detection using SWIN transformer

Abstract

Diabetic retinopathy (DR) is a diabetes related eye disorder that damages the retina. DR is among the most specific complications of diabetes. A vital challenge for automated detection systems in medical image diagnosis is to minimize the false negative rate for patients’ timely treatment. This paper presents a novel strategy employing the shifted window (SWIN) Transformer for efficiently modeling local and global visual information to address this challenge. We have proposed our work to maximize the true positive ratio and minimize the false negative ratio for the automated process of diagnosing the level of DR, so that patients with positive signs of DR can be predicted most accurately and can save vision. The results suggest that SWIN Transformer architecture, along with the contrast-limited adaptive histogram equalization (CLAHE) technique, provides a robust option for developing a reliable DR detection system. The results indicate that the proposed approach achieves 96% weighted recall across all the levels of DR detection and 97.45% validation accuracy for the eyePACS DR detection dataset, as well as 99% weighted recall across all the levels of DR detection, along with 99.26% validation accuracy for APTOS 2019 Blindness Detection dataset. Thus, this study aimed to develop a DR detection system focused on minimizing false negatives using the SWIN transformer.