Enhanced Kidney Diagnosis by CT Scan Stone Annotation

The increasing prevalence of kidney stone disease necessitates efficient and accurate diagnostic methods to alleviate the burden on healthcare systems and professionals. Traditional manual methods of CT scan analysis are time-intensive and prone to human error, often delaying critical diagnoses. This study introduces an automated detection framework utilizing the YOLO NAS model, specifically optimized for real-time kidney stone annotation in CT scans. The dataset includes over 10,000 CT images sourced from Kaggle and Roboflow, enriched with additional scans manually annotated some image using the VGG Image Annotator tool to ensure comprehensive coverage of kidney stone types, sizes, and densities. The YOLO NAS model was selected due to its superior performance in object detection, leveraging Neural Architecture Search for optimization and trained using the SuperGradients library. The proposed model achieves a mean average precision (mAP) of 93% at a 0.50 Intersection over Union (IoU) threshold, demonstrating its high accuracy and efficiency. This automated solution reduces radiologists' workload, enhances diagnostic precision, and enables timely interventions, ultimately improving patient care. Future enhancements include expanding the dataset, integrating multi-modal data, and optimizing deployment for real-time clinical applications. By automating kidney stone detection, this system offers a robust approach to improving medical diagnostics.