We aimed to predict the prognosis of advanced nasopharyngeal carcinoma (stage Ⅲ-Ⅳa) using Pre- and Post-treatment MR images based on deep learning (DL). A total of 206 patients with primary nasopharyngeal carcinoma who were diagnosed and treated at the Renmin Hospital of Wuhan University between June 2012 and January 2018 were retrospectively selected. A rectangular region of interest (ROI), which included the tumor area, surrounding tissues and organs, was delineated on each Pre- and Post-treatment MR image. Two Inception-Resnet-V2 based transfer learning models, named Pre-model and Post-model, were trained with the Pre-treatment images and the Post-treatment images, respectively. In addition, an ensemble learning model based on the Pre-model and Post-models was established. The three established models were evaluated by receiver operating characteristic curve (ROC), confusion matrix, and Harrell's concordance indices (C-index). High-risk-related gradient-weighted class activation mapping (Grad-CAM) images were developed according to the DL models. The Pre-model, Post-model, and ensemble model displayed a C-index of 0.717 (95% CI: 0.639 to 0.795), 0.811 (95% CI: 0.745-0.877), 0.830 (95% CI: 0.767-0.893), and AUC of 0.741 (95% CI: 0.584-0.900), 0.806 (95% CI: 0.670-0.942), and 0.842 (95% CI: 0.718-0.967) for the test cohort, respectively. In comparison with the models, the performance of Post-model was better than the performance of Pre-model, which indicated the importance of Post-treatment images for prognosis prediction. All three DL models performed better than the TNM staging system (0.723, 95% CI: 0.567-0.879). The captured features presented on Grad-CAM images suggested that the areas around the tumor and lymph nodes were related to the prognosis of the tumor. The three established DL models based on Pre- and Post-treatment MR images have a better performance than TNM staging. Post-treatment MR images are of great significance for prognosis prediction and could contribute to clinical decision-making.

Importance: Accurate pre-treatment prediction of distant metastasis in patients with Nasopharyngeal Carcinoma (NPC) enables the implementation of appropriate treatment strategies for high-risk individuals. Purpose: To develop and assess a Convolutional Neural Network (CNN) model using pre-therapy Magnetic Resonance (MR) imaging to predict distant metastasis in NPC patients. Methods: We retrospectively reviewed data of 441 pathologically diagnosed NPC patients who underwent complete radiotherapy and chemotherapy at Renmin Hospital of Wuhan University (Hubei, China) between February 2012 and March 2018. Using Adobe Photoshop, an experienced radiologist segmented MR images with rectangular regions of interest. To develop an accurate model according to the primary tumour, Cervical Metastatic Lymph Node (CMLN), the largest area of invasion of the primary tumour, and image segmentation methods, we constructed intratumoural and intra-peritumoural datasets that were used for training and test of the transfer learning models. Each model’s precision was assessed according to its receiver operating characteristic curve and accuracy. Generated high-risk-related Grad-Cams demonstrated how the model captured the image features and further verified its reliability. Results: Among the four models, all intra-peritumoural datasets performed better than the corresponding intratumoural datasets, with the CMLN intra-peritumoural dataset exhibiting the best performance (average area under the curves (AUCs) = 0.88). There was no significant difference between average AUCs of the Max and NPC tumour datasets. AUCs of the eight datasets for the four models were higher than those of the Tumour-Node-Metastasis staging system (AUC=0.67). In most datasets, the xception model had higher AUCs than other models. The efficientnet-b0 and xception models efficiently extracted high-risk features. Conclusion: The CNN model predicted distant metastasis in NPC patients with high accuracy. Compared to the primary tumour, the CMLN better predicted distant metastasis. In addition to intratumoural data, peritumoural information can facilitate the prediction of distant metastasis. With a larger sample size, datasets of the largest areas of tumour invasion may achieve meaningful accuracy. Among the models, xception had the best overall performance.