Image analysis of diabetes pathology: classifying with precision via an upgraded resnet50 framework

Diabetes is a growing global health issue, and effective diagnostic tools are needed to support early detection. This study proposes an enhanced deep learning framework, SE-ResNet50, which integrates a squeeze-and-excitation (SE) block into the conventional ResNet50 architecture to improve the classification of diabetic kidney pathology from glomerular images. The SE block adaptively recalibrates feature responses, enabling the model to emphasize diagnostically relevant structures better. The proposed framework was trained and validated on a kidney tissue dataset from Kanazawa Medical University, achieving superior performance with an accuracy of 97.02%, precision of 0.96, and an AUC of 0.9856. SE-ResNet50 exhibited superior robustness and generalizability compared to established CNN architectures such as EfficientNet B0, Inception V3, and ConvNeXt. Further visualization via Grad-CAM revealed that the model effectively localizes critical regions within glomerular images. These results highlight the potential of SE-ResNet50 as a reliable and interpretable tool for advancing diabetes-related CKD diagnosis in clinical settings. • A SE-ResNet50 framework is proposed for accurate classification of diabetic kidney pathology from glomerular images. • The model incorporates a squeeze-and-excitation block to enhance feature channel recalibration and diagnostic sensitivity. • SE-ResNet50 achieves superior accuracy and AUC compared to ResNet50, EfficientNet B0, and Inception V3. • Grad-CAM visualization confirms that the model focuses on clinically relevant glomerular structures, enhancing interpretability.