Abstract
PURPOSE: To develop a computer-aided automatic-detection (CAD) deep-learning algorithm to identify a urinary stone in low-dose non-enhanced CT images. MATERIALS AND METHODS: This retrospective study was performed at a single institution. Over a period of 14 months, the low-dose CT images of 486 patients with suspicious urinary stone disease were collected. The labeling of urinary stones, or not, in low-dose CT images was performed by an expert uroradiologist as a reference standard. We used labeled CT scans (axial 1,144, coronal 1,279, sagittal 765). We developed a CAD deep-learning algorithm using the YOLO v7 model. The data ratio for training, validation, and testing was set at 6:3:1. The performance of our proposed CAD deep-learning algorithm at identifying a urinary stone was analyzed using several parameters, such as the mean average performance (mAP), precision, recall, F1-score, and accuracy. RESULTS: The mAP of our proposed algorithm was 95%. The accuracy of the CAD deep-learning algorithm for urinary stone detection was 93% and 92%, in the training and test sets, respectively. CONCLUSION: The proposed CAD algorithm developed using a deep-learning model has high performance at urinary stone detection in low-dose CT images.