A two-stage deep learning framework for kidney stone detection and clinical severity grading in CT imaging

Kidney stones affect millions globally, with non-contrast computed tomography (CT) serving as the gold standard for detection. Current manual analysis is time-consuming and subject to inter-observer variability, necessitating automated diagnostic solutions. This study presents an integrated, automated framework combining YOLOv8 object detection with ResNet-18 regression for comprehensive kidney stone detection and severity assessment from abdominal CT images. A novel two-stage inference pipeline was developed using the Kaggle kidney stone dataset (2,603 annotated CT images) split into training (80%), validation (10%), and testing (10%) sets. The first stage employs YOLOv8 for precise stone localization, achieving precision of 0.737, recall of 0.688, and mean Average Precision at Intersection over Union threshold 0.5 ( mAP@0.5 ) of 0.707. The second stage utilizes ResNet-18 regression for severity quantification, attaining mean absolute error (MAE) of 0.1080 for continuous severity prediction, subsequently mapped to clinical categories (Mild, Moderate, Severe) using predefined thresholds. The integrated pipeline demonstrates robust performance across diverse anatomical contexts with minimal false positives, offering interpretable outputs suitable for real-time clinical integration. This automated approach addresses critical gaps in standardized stone assessment and supports diagnostic decision-making in radiology and urology workflows.