CT-Based Deep Learning Enables Postoperative Recurrence Prediction for Intrahepatic Cholangiocarcinoma

Abstract Preoperatively accurate evaluation of risk for postoperative recurrence contributes to maximizing the therapeutic success of surgical resection for intrahepatic cholangiocarcinoma (iCCA). This study aimed to investigate the potential of deep learning (DL) algorithms for predicting postoperative early recurrence through the use of preoperative images. We collected the dataset, including preoperative plain computed tomography (CT) images, from 41 patients undergoing curative surgery for iCCA at multiple institutions. We built a CT patch-based predictive model using a residual convolutional neural network and used 5-fold cross-validation. The prediction accuracy of the model was analyzed. We defined early recurrence as recurrence within a year after surgical resection. Of the 41 patients, early recurrence was observed in 20 (48.8%). A total of 71,081 patches were extracted from the entire segmented tumor area of each patient. The average accuracy of the ResNet model for predicting early recurrence was 98.2% for the training dataset. In the validation dataset, the average sensitivity, specificity, and accuracy were 97.8%, 94.0%, and 96.5%, respectively. Furthermore, the area under the receiver operating characteristic curve was 0.994. Our CT-based DL model exhibited high predictive performance in projecting postoperative early recurrence, proposing a novel insight into iCCA management.