Explainable Kidney Stone Detection and Localization in CT with U-Net, EfficientNet, and Multi-Method XAI

Abstract

Kidney stone disease affects approximately 10% of the global population, underscoring the clinical importance of accurate CT-based stone classification and precise localization. This paper presents a leakage-controlled deep learning framework that delivers both pixel-level localization and transparent image-level prediction through two complementary model pipelines that are trained and evaluated independently. In the first pipeline, a U-Net is trained for grayscale-to-color reconstruction and stone localization; a binary stone mask is then obtained via post-processing to visually delineate suspected regions. In the second pipeline, an EfficientNet-B0 model is trained for binary classification (Normal vs. Stone) using patientdisjoint train/validation/test splits to mitigate data leakage and better reflect realistic deployment, and is benchmarked against standard Convolutional Neural Network (CNN) baselines. We further incorporate explainable AI to validate model reasoning: Gradient-weighted Class Activation Mapping (Grad-CAM) highlights discriminative regions, Local Interpretable Modelagnostic Explanations (LIME) provides superpixel-based local evidence, and SHapley Additive exPlanations (SHAP) assigns pixel-wise contribution scores. Across experiments, grayscale representations substantially outperformed color inputs, and augmentation improved generalization. Using two state-of-theart datasets, the proposed classifier achieved 99.76% and 99.65% test accuracy, respectively. For segmentation/localization, the U-Net component achieved Dice scores of 0.91-0.93 on the first dataset and 0.99 on the second dataset. Qualitative explainability results consistently focused on stone-relevant kidney regions, supporting interpretability and clinical plausibility of predictions. These findings indicate that combining robust training practices with multi-method explainability can yield highly accurate and clinically interpretable kidney stone analysis from CT images.