Gadolinium-free contrast-enhanced MRI (GFCE-MRI) synthesis via deep learning for radiotherapy of nasopharyngeal carcinoma

Abstract

Nasopharyngeal carcinoma (NPC) is a highly infiltrative and radiosensitive malignancy. Radiotherapy is currently the mainstay therapeutic remedy. In radiotherapy of NPC patients, the gadolinium-based contrast enhanced MRI (CE-MRI) plays a critical role in NPC delineation. However, the gadolinium-based contrast agents (GBCAs) associated safety issues have attracted serious attention of clinicians in recent years. To reduce or eliminate the use of GBCAs, deep learning has been proposed to synthesize the gadolinium-free contrast enhanced MRI (GFCE-MRI), aiming at providing an alternative to gadolinium-based CE-MRI. Nevertheless, recent studies mostly focus on novel deep learning algorithms development or feasibility investigations for disease diagnosis in different anatomies, such as brain, liver, and breast. Currently, these is no study has been reported for NPC radiotherapy. In this study, we for the first time developed deep learning algorithm to synthesize GFCE-MRI from contrast-free T1-weighted (T1w) and T2-weighted (T2w) MRI for radiotherapy of NPC patients. Specifically, we achieved three research objectives in this study: (i) to develop a novel multimodality-guided synergistic neural network (MMgSN-Net) that tailored for GFCE-MRI synthesis of NPC patients; (ii) to investigate and improve the GFCE-MRI model generalizability using multi-institutional MRI data; and (iii) to investigate the clinical efficacy of GFCE-MRI in primary NPC tumor delineation. Our experiments showed that the proposed MMgSN-Net is able to generate highly realistic GFCE-MRI images and the quantitative results outperformed three comparing state-of-the-art methods. We also found that the heterogeneity of multi-institutional MRI heavily affects generalizability of the well-trained single-institutional model. After training the model with multi-institutional data and shorting the multi-institutional data distribution variations, the model generalizability has been significantly improved. The clinical evaluation results also suggest that our synthetic GFCE-MRI is highly promising for clinical use, with Dice Similarity Coefficient (DSC) of 0.762 and Hausdorff Distance (HD) of 1.932mm, respectively. The dosimetric difference of planning target volumes between real patients and synthetic patients was less than 1%, which is acceptable for radiotherapy as reported by two board-certified oncologists.