An explainable deep-learning model reveals clinical clues in diabetic retinopathy through SHAP

Abstract

Diabetic retinopathy (DR) is one of the leading causes of blindness globally. Several studies indicate that 90% of cases are preventable through early detection and appropriate treatment. Due to the increasing number of diabetic patients, the number of images that ophthalmologists have to manually analyze is becoming unaffordable. In this study, we propose a robust method for the automatic grading of the DR, while emphasizing the importance of providing visual explanations. The proposed method leans on a modified layer architecture of the ResNet-50 network. It also includes additional techniques such as data augmentation, regularization, early stopping criteria, transfer learning, and fine-tuning. In addition, in order to assist in the interpretation of the results of the deep-learning model, we introduce a visual Explainable Artificial Intelligence approach using SHapley Additive exPlanations (SHAP). We evaluated the effectiveness of our method using five publicly available databases of retinal images: APTOS-2019, EyePACS, DDR, IDRiD, and SUSTech-SYSU, achieving accuracy rates of 94.64%, 86.36%, 84.23%, 82.79%, and 85.65%, respectively. Notably, SHAP analysis revealed insights into our results, suggesting that retinal vasculature changes are potential DR risk indicators. We also found that peripheral retinal observations proved crucial in predicting DR progression, with initial lesions often found there. Moreover, this work overcomes the challenges of a highly imbalanced dataset, commonly encountered in clinical environments. To the best of our knowledge, our results show for the first time the usefulness of SHAP visual explanations in DR grading, thus contributing to an early adoption of automated solutions in real clinical environments.