Prediction of Anatomical Presentation during Radiotherapy of Nasopharyngeal Carcinoma Using GAN-LSTM for Plan Adaption Decision

Abstract

During treatment, patients frequently undergo anatomical changes that might result in dose degradation. Adaptive radiation therapy (ART) is now available to overcome this problem. However, this method is time-consuming, and the lack of criteria to trigger replanning prevents its widespread use. To overcome these obstacles, anatomical presentation models are necessary. In this study, we developed a novel deep-learning method to make a series of predictions for the remaining treatment course. Total 230 nasopharyngeal carcinoma patients who received radiotherapy in 33 fractions were enrolled. The data included cone-beam computed tomography (CBCT) and planning CT images. CBCT image quality was improved to CT level using an in-house software. A generative adversarial network-long short-term memory network model was proposed, with the generation abilities of the former network and forecasting abilities of the latter. To predict the anatomical presentation for 3-6 weeks, we trained four models. The planning CT and CBCT acquired earlier were used as the input. Physicians segmented the gross target volume (GTVnx) and parotid glands on prediction and real CBCT (ground truth). Contours, dosimetry parameters, and plan adaptation decision were used to evaluate the models. The table shows the overall performance of the test set (18 cases). The anatomical changes were predicted over the treatment course with a dice similarity coefficient (DSC) of 0.96, 0.90 and 0.92 and mean distance to agreement (MDA, in mm) of 0.37, 0.70, and 0.60 for GTVnx, left parotid, and right parotid, respectively. Bland-Altman analysis revealed that dosimetry parameters did not show significant difference between prediction and ground truth. The prescription coverage (%) of GTVnx, V30 of the left parotid, and V30 of the right parotid had mean absolute biases of 0.09, 1.09, and 0.27, respectively. At week 6, there were two cases that required plan adaptation, and the model effectively triggered replanning one week in advance. We developed a framework that predicts the anatomical changes occurring in future fractions. Establishing such a framework provides a proactive approach to ART and saves clinical time by anticipating and preparing for treatment strategies in advance.