An enhanced interpretable deep learning approach for diabetic retinopathy detection

Abstract

Diabetic Retinopathy (DR) is a consequence of type1 or type-2 diabetes. It is critical to identify complications early since they may result in visual issues such as retinal detachment, vitreous hemorrhage, and glaucoma. The interpretability of automated classifiers for medical diagnoses such as diabetic retinopathy is critical. The primary issue is the difficulties inherent in inferring reasonable conclusions from them. In recent years, numerous efforts have been made to transform deep learning classifiers from statistical black box machines with high confidence to self-explanatory models. The concern of effective data preprocessing before classification remains unsolved. Although the application of machine Learning schemes has proven to be effective when trained in a supervised way, it still has limitations with data redundancy, feature selection, and human expert interference. Hence, a combinatorial deep learning approach is proposed to interpret diabetic retinopathy (DR) detection. The proposed method combines the Shapley Additive Explainability (SHAP) and Local 127:1679 Model-Agnostic Explanations (LIME) to analyze the deep learning output effectively. Results from our experiment show that our proposed approach outperformed the existing schemes in detecting DR. � 2022 IEEE.