Deep Learning Architectures Comparison for Kidney Stone Classification in CT Images

Abstract

Kidney stones are a common condition that may cause severe pain and complications if not diagnosed early. Computed tomography (CT) imaging is widely used for identifying renal stones due to its high sensitivity. This study focuses on binary image-level classification (stone vs. normal) in CT images using deep learning architectures. Three convolutional neural networks—ResNet50, XResNet50, and DenseNet201—were evaluated under a unified preprocessing and training pipeline. Two publicly available CT datasets were used, each divided into 70% training and 30% testing sets. Models were trained using Google Colab with data augmentation to reduce overfitting. Performance was assessed using accuracy, precision, recall, and F1-score. Across the two datasets, XResNet50 achieved the highest accuracy (97% and 92%, respectively). While the reported results indicate strong performance under the defined experimental setup, further validation using patient-wise splitting and additional datasets is recommended to confirm generalisation. These findings provide a comparative reference for selecting suitable architectures for CT-based kidney stone classification.