Deep learning for the prediction of residual tumor after radiotherapy and treatment decision-making in patients with nasopharyngeal carcinoma based on magnetic resonance imaging

Abstract

Background: Concurrent chemoradiotherapy (CCRT) and induction chemotherapy (IC) plus CCRT (IC + CCRT) are the main treatments for patients with advanced nasopharyngeal carcinoma (NPC). We aimed to develop deep learning (DL) models using magnetic resonance (MR) imaging to predict the risk of residual tumor after each of the 2 treatments and to provide a reference for patients to select the best treatment option.

Methods: A retrospective study was conducted on 424 patients with locoregionally advanced NPC who underwent CCRT or IC + CCRT between June 2012 and June 2019 in the Renmin Hospital of Wuhan University. According to the evaluation of MR images taken 3 to 6 months after radiotherapy, patients were divided into 2 categories: residual tumor and non-residual tumor. Transferred U-net and Deeplabv3 neural networks were trained, and the better-performance segmentation model was used to segment the tumor area on axial T1-weighted enhanced MR images. Then, 4 pretrained neural networks for prediction of residual tumors were trained with CCRT and IC + CCRT datasets, and the performances of the models trained using each image and each patient as a unit were evaluated. Patients in the test cohort of CCRT and IC + CCRT datasets were successively classified by the trained CCRT and IC + CCRT models. Model recommendations were formed according to the classification and compared with the treatment decisions of physicians.

Results: The Dice coefficient of Deeplabv3 (0.752) was higher than that of U-net (0.689). The average area under the curve (aAUC) of the 4 networks was 0.728 for the CCRT and 0.828 for the IC + CCRT models trained using a single image as a unit, whereas the aAUC for models trained using each patient as a unit was 0.928 for the CCRT and 0.915 for the IC + CCRT models, respectively. The accuracy of the model recommendation and the decision of physicians was 84.06% and 60.00%, respectively.

Conclusions: The proposed method can effectively predict the residual tumor status of patients after CCRT and IC + CCRT. Recommendations based on the model prediction results can protect some patients from receiving additional IC and improve the survival rate of patients with NPC.

Keywords: Deep learning; concurrent chemoradiotherapy (CCRT); induction chemotherapy; nasopharyngeal carcinoma (NPC); residual tumor.