Explainable MedAttn-ResNet50 for Automated Kidney Tumor Classification from CT Images With Grad-CAM Visualization

Abstract

Kidney tumor classification from CT imaging remains challenging due to CT-specific intensity variations, tumor heterogeneity, and limited clinically meaningful interpretability in existing deep learning models. To address these limitations, we propose an Explainable MedAttn-ResNet50 framework that integrates radiology-driven preprocessing, multi-scale feature extraction, and medical attention mechanisms within a unified architecture. The preprocessing pipeline applies radiology-inspired intensity normalization, contrast enhancement, and data augmentation to preserve diagnostically relevant tissue contrast. A Multi-Scale Feature Extraction Module (MSFEM) captures tumors of varying shapes and sizes, improving robustness to heterogeneous cases, while spatial, channel, and medical knowledge–gated attention guides the network to focus on anatomically relevant regions. Grad-CAM is applied at deeper layers to generate high-resolution visual explanations and was assessed qualitatively via representative overlays. The model was trained and evaluated on 10,000 CT images and achieved 99.40% accuracy, with 99.80% sensitivity, 99.00% specificity, and 0.993 AUC, outperforming several baseline architectures. These results demonstrate that the proposed method improves predictive performance while providing clinically aligned interpretability for kidney tumor classification.