Explainable Deep Learning for Kidney CT Scan Classification

This study presents a clear and accurate framework for automatically classifying kidney computed tomography (CT) scans into four important categories: normal, cyst, stone, and tumour. While current methods show high diagnostic accuracy, their lack of interpretability and confidence calibration limits their use in clinical settings. To address these issues, the proposed system uses an improved image analysis architecture along with three supportive interpretability methods: Gradient-weighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-Agnostic Explanations (LIME), and Score-CAM. These methods help visualize decision patterns at different levels. A realtime reliability diagram is included to regularly evaluate and ensure probability calibration for each classification. This helps provide reliable confidence estimates. The framework was built and tested using a publicly available dataset of 1 2, 4 4 6 kidney CT images that cover different anatomical orientations and contrast phases. It achieved a classification accuracy of 94.2%, a macroaveraged F1-score of 0.925, and an Expected Calibration Error (ECE) of 2.0%. The system shows strong ability to differentiate and maintain robust confidence. These results emphasize the importance of combining explainability and calibration to develop transparent, trustworthy, and clinically valuable diagnostic systems.