DReXDetect: An Explainable Deep Learning Framework for Diabetic Retinopathy Classification Viable in Clinical Practice
Pallabi Das, Kasturi Barik, Rajashree NayakIntroduction:
Diabetic Retinopathy (DR) is a severe outcome of diabetes that may lead to vision impairment. Manual screening of DR is time-demanding, subjective, and less accessible in resource-limited areas. Existing deep learning (DL) models outperform in classifying between the No-DR and Severe-DR classes. However, several existing methods report reduced sensitivity in detecting mild and moderate DR stages due to biased training using publicly available imbalanced datasets.
Methods:
This study introduces DReXDetect, a hybrid computer-aided diagnostic (CAD) framework driven by Explainable Artificial Intelligence (XAI). The proposed model exclusively hybridizes established transfer learning (TRL) models, ResNet-50 and Inception-V3, for robust DR severity classification. The incorporation of XAI analysis, including Gradientweighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-Agnostic Explanations (LIME), and Shapley Additive Explanations (SHAP), offers visual and feature-level insights into model decision-making, potentially fostering trust in clinical environments.
Results:
Experimental results contribute towards the development of a clinically trustworthy XAI-assisted CAD system, with particular emphasis on grade-wise reliability and interpretability across different DR severity stages.
Discussion:
The performance of DReXDetect is evaluated against various cutting-edge TRL architectures, including standalone DenseNet-121, Inception-V3, and Xception. ResNet-50 is considered the baseline model because of its reliable residual learning. DenseNet-121 and Xception are excluded from the fusion process to prevent feature redundancy and to enhance discriminatory information; instead, ResNet-50 is hybridized with Inception-V3 to extract multi- scale, diverse, and discriminative deep features.
Conclusion:
This paper presents a systematic, interpretable, and hybrid feature-learning architecture within a unified, transparent evaluation framework for five-stage DR classification, thereby emphasizing robustness, transparency, and class-wise reliability.