利用人工智慧判讀腹部Ｘ光(KUB)之腎臟與輸尿管結石

陳約任¹、馮思中¹、吳俊德¹、莊正鏗¹、虞凱傑¹、林柏宏¹、邵翊紘¹、黃亮鋼¹、范佐搖²、郭昶甫²、陳嶽鵬^2*、甘弘成^1*

¹ 林口長庚紀念醫院泌尿科;² 林口長庚醫院醫療人工智能核心實驗室;^*共同通訊作者

The Application of Deep Learning Models for Kidney and Upper Ureteral Stone Detection on Abdominal Plain X-ray Images

Jonathan YJ Chen ¹, See-Tong Pang ¹, Chun-Te Wu ¹, Cheng-Keng Chuang ¹, Kai-Jie Yu ¹, Po-Hung Lin ¹, I-Hung Shao ¹, Liang-Kang Huang ¹, Tzuo-Yau Fan ², Chang-Fu Kuo ², Yueh-Peng Chen ²*, Hung-Cheng Kan ¹*

¹ Divisions of Urology, Department of Surgery, Chang-Gung Medical Foundation Linkou Chang-Gung Memorial Hospital; ² Center for Artificial Intelligence in Medicine, Chang Gung Memorial Hospital, Linkou, Taiwan; *Co-corresponding author

Purpose:

        The plain X-ray examination of the abdomen (KUB) serves as a widely employed and convenient diagnostic modality for the detection of urolithiasis, offering advantages over computed tomography (CT) in terms of cost and radiation exposure. Nonetheless, achieving accurate diagnosis of urinary stones on KUB remains a significant challenge. In recent years, computer-aided diagnosis (CAD) has emerged as a promising approach in the realm of urinary stone detection, predominantly applied with CT imaging. However, investigations pertaining to CAD implementation on plain radiographs for urolithiasis detection remain limited. Our study endeavors to address this gap by developing a CAD system utilizing deep learning architecture for the precise detection of kidney and upper ureter stones on KUB images, accompanied by an assessment of its accuracy.

Materials and Methods:

        KUB images were sourced from the Chang Gung Research Database and manually annotated by three urologists to identify kidney or upper ureter stones. A total of 1900 images depicting urolithiasis from 989 patients were amassed and annotated for model development. Among these, 1523 images constituted the training set, while 377 images comprised the validation set. Additionally, the model's performance was validated using an additional 726 images, consisting of 652 images depicting stones and 74 images devoid of any stone, all confirmed via CT imaging. RetinaNet served as the deep learning model backbone for stone detection, with mean average precision (mAP) employed for performance evaluation. Metrics such as F1-score, precision, sensitivity, false discovery rate (FDR), and false negative rate (FNR) were utilized to assess the accuracy of the model.

Results:

The optimal performance metrics achieved on the validation dataset were as follows: F1-score of 82%, precision of 96%, sensitivity of 71%, FDR of 4%, and FNR of 29%. On the test dataset, the model exhibited the following metrics for kidney and upper ureter stone detection: F1-score of 81%, precision of 95%, sensitivity of 70%, FDR of 5%, and FNR of 30%. The mAP attained on the testing dataset was 0.65.

Conclusions:

        The application of a deep learning-based CAD system for urolithiasis detection on plain X-ray imaging demonstrates feasibility and potential utility in clinical practice. Our model exhibits promising accuracy in identifying kidney and upper ureter stones on KUB images. The findings of this study suggest potential avenues for extension, particularly in the development of detection models targeting lower ureter stones and bladder stones, entities typically challenging to diagnose via plain radiography.