DiaKANet: An Explainable Attention-Enhanced Deep Feature Learning for Diabetic Retinopathy Classification from Retinal Fundus Images

Abstract

Diabetic retinopathy (DR) is a progressive complication of diabetes that significantly damages the blood vessels within the retina. However, it is difficult to detect in the early-stage owing to its asymptomatic nature; patients may experience intermittent or blurred vision as the condition develops. Retinal fundus images offer valuable information about the pathological changes in the retinal region, facilitating computer-aided detection and severity classification. Despite recent advances, existing techniques often struggle to deal with variations in retinal image quality and to extract subtle pathological patterns, thus limiting their reliability in multi-class classification. To tackle these issues, this study develops an Interpretable KAN-Based Framework for Multi-Class Diabetic Retinopathy Classification (IKAN-MDRC), aiming to develop an automated, robust, and explainable multi-class classification model by integrating advanced techniques using fundus images. To achieve this, the model uses a RegNet feature extraction strategy along with an efficient channel attention mechanism to effectively capture hierarchical retinal features and perform lightweight channel-wise attention to highlight informative retinal feature channels and eliminate less informative features. The extracted features are classified through the EfficientKAN classifier by effectively learning nonlinear relationships from the retinal features. In addition, the Harris Hawks optimization technique is used to select the most suitable hyperparameters to increase the classification performance and provide faster convergence. The model transparency is improved through incorporating the GradCAM++ explainability model by highlighting the significant retinal region that contributes more to the model decision. The experimental analysis is conducted using the Diabetic Retinopathy Detection dataset across various performance metrics. These findings demonstrate that the proposed IKAN-MDRC framework attains superior classification results with an accuracy of 97.61% compared to existing models and validates its potential in accurate categorization of multi-class DR with improved model interpretability to support clinical decision-making.