Therapeutic Response Prediction to Neoadjuvant Chemotherapy for Rectal Cancer Using the Deep Learning Approach

Abstract Background The prediction of response to chemotherapy can lead to the optimization of neoadjuvant chemotherapy (NAC). This study aimed to develop a predicting model of therapeutic response to NAC for rectal cancer (RC). Methods Three courses of S-1 and oxaliplatin (SOX) NAC were administered before total mesorectal excision. We collected the dataset for the prechemotherapy arterial phase of enhanced computed tomography (CT) images from 57 patients undergoing rectal surgery after NAC for RC. In all cases, the therapeutic response to NAC had been pathologically confirmed. We established three prediction labels; poor response, marked response, and complete response (CR). We built a predictive model using a residual convolutional neural network (ResNet50) and used 3-fold cross-validation. The prediction accuracy of the model was analyzed. Results Of the 57 patients, pathological CR was observed in 9 (15.8%). A total of 4,607 squares were extracted from the segmented tumor area of each patient. The average accuracy of the ResNet model for predicting pathological CR was 99.9% for the training dataset. In the test dataset, the average accuracy was 94.9%. Likewise, in the prediction of marked and poor responses, the models demonstrated high accuracy (93.6% and 93.3%, respectively) and high AUC (0.966 and 0.976, respectively). Conclusions Our deep learning model, using prechemotherapy CT images of RC, exhibited high predictive performance in projecting therapeutic response to SOX NAC. This study presents a novel insight into the optimization of NAC for RC.