利用人工智慧對電腦斷層影像上良性和惡性的腎臟腫瘤進行分類

陳約任¹、馮思中²、虞凱傑²、林柏宏²、劉忠一²、邵翊紘²、莊正鏗²、張英勛²、

甘弘成²

¹林口長庚紀念醫院；²林口長庚紀念醫院泌尿科

Automated differentiation of benign and malignant renal tumors on computed tomography using artificial intelligence

Chen Jonathan YJ ¹, Pang ST ², Yu KJ ², Lin PH ², Liu CY ², Shao IH ², Chuang CK ², Chang YH ², Kan HC²

¹ Chang-Geng Medical Foundation Linkou Chang-Geng Memorial Hospital; ² Divisions of Urology, Department of Surgery, Chang-Geng Medical Foundation Linkou Chang-Geng Memorial Hospital

Purpose:

To develop and evaluate the feasibility of an objective method using artificial intelligence (AI) to differentiate different cell type renal tumors using convolutional neural networks (CNNs) on computed tomography images.

Materials and Methods:

We used CNN trained and validated to identify different cell type renal tumors in contrast-enhanced computed tomography images from 251 patients (2162 images). The ResNet-18 model was cross-trained to perform this classification. We developed two models, model one was used to identify benign and malignant renal tumors and model two was used to identified different cell type of malignant renal tumors. The network's performance was evaluated through analysis of the accuracy, sensitivity and specificity from five-fold cross-validation.

Results:

Model one achieved accuracy of 82% in benign and malignant renal tumor type classification (87% sensitivity and 92% specificity) compared with the final pathology results. Model two achieved accuracy of 84% in clear cell renal cell carcinoma and non-clear cell renal cell carcinoma classification (86% sensitivity and 97% specificity) compared with the final pathology results.

Conclusion:

Machine learning analysis of CT texture features can facilitate the accurate differentiation of different cell type renal tumors.