An automated approach for the kidney segmentation and detection of kidney stones on computed tomography using YOLO algorithms

Abstract

Background: For effective diagnosis and treatment planning, accurate segmentation of the kidneys and detection of kidney stones are crucial. Traditional procedures are time-consuming and subject to observer variation. This study proposes an automated method employing YOLO algorithms for renal segmentation and kidney stone detection on CT scans to address these issues. Methods: The dataset used in this study was sourced from the GitHub. The dataset contains a total of 1799 images, with 790 images labeled as 'containing kidney stones' and 1009 images labeled as 'not containing kidney stones'. U-Net architecture was utilized to precisely identify the region of interest, while YOLOv5 and YOLOv7 architecture was utilized to detect the stones. In addition, a performance comparison between the two YOLO models and other contemporary relevant models has been conducted. Results: We obtained a kidney segmentation IOU of 91.4% and kidney stone detection accuracies of 99.5% for YOLOv7 and 98.7% for YOLOv5. YOLOv5 and YOLOv7 outperform the best existing models, including CNN, KNN, SVM, Kronecker CNN, Xresnet50, VGG16, etc. YOLOv7 possesses superior accuracy than YOLOv5. The only issue we encountered with the YOLOv7 model was that it demanded more training time than the YOLOv5 model. Conclusion: The results demonstrate that the proposed AI-based method has the potential to improve clinical procedures, allowing radiologists and urologists to make well-informed decisions for patients with renal pathologies. As medical imaging technology progresses, the incorporation of deep learning techniques such as YOLO holds promise for additional advances in automated diagnosis and treatment planning.