What question did this study set out to answer?

The aim is to enhance the classification of diabetic retinopathy from retinal fundus images using advanced deep learning techniques.

April 25, 2026Open Access

REGANXNet: GAN-Augmented Hybrid Architecture With LIME- and SHAP-Based XAI Heat Maps for the Classification of Diabetic Retinopathy

Key Points

The aim is to enhance the classification of diabetic retinopathy from retinal fundus images using advanced deep learning techniques.
Utilized a GAN-augmented hybrid architecture combining EfficientNet-B4, InceptionResNetV2, and ResNet101.
Augmented a dataset of 19,500 images through class-conditional GAN synthesis and rigorous preprocessing techniques.
Implemented Explainable AI with LIME and SHAP to generate heatmaps for lesion identification.
Achieved 96.4% accuracy, 96.9% precision, and 96.1% recall with a mean IoU score of 90.8%.
Outperformed various architectures including Swin Transformer-Tiny (91.0%) and ResNet101 (89.8%).
Confusion matrices showed minimal errors between adjacent diabetic retinopathy stages.

Abstract

Accurate classification of diabetic retinopathy (DR) from retinal fundus images is essential for timely intervention, yet challenges like class imbalance, hidden logic, and limited generalization slow down the adoption of deep learning in clinical practice. To address these challenges, we propose REGANXNeta Retinal GAN-augmented eXplainable Network that integrates EfficientNet-B4, InceptionResNetV2, and ResNet101 to utilize their complementary strengths. Class-conditional GAN synthesis augments a 19, 500image dataset from APTOS 2019, EyePACS, DDR, and IDRiD, with rigorous preprocessing including CLAHE and 512 512 normalization to ensure balanced representation and mitigate overfitting. Explainable AI via LIME and SHAP generates heatmaps that highlight specific regions of the lesion such as microaneurysms and hemorrhages, achieving mean IoU scores of 0. 58 (LIME) and 0. 66 (SHAP) against expert masks. Comprehensive evaluation yields 96. 4% accuracy, 96. 9% precision, 96. 1% recall, 96. 5 \% ~F 1 score, and 90. 8% IoU, outperforming Swin Transformer-Tiny (91. 0%), ResNet101 (89. 8%), MobileNetV3Large (88. 9%), EfficientNet-B4 (89. 3%), DenseNet201 (88. 1%), and ConvNeXt-Tiny (87. 2%). Ablation studies confirm synergy from feature fusion, while confusion matrices reveal minimal errors between adjacent DR stages. REGANXNet provides reliable and easy-to-understand DR screening that works on basic computers but needs future improvements to provide faster use on small devices.

REGANXNet: GAN-Augmented Hybrid Architecture With LIME- and SHAP-Based XAI Heat Maps for the Classification of Diabetic Retinopathy

Key Points

Abstract

Cite This Study