Deep Learning-Based Synthesis of Contrast-Enhanced MRI for Automated Delineation of Primary Gross Tumor Volume in Radiotherapy of Nasopharyngeal Carcinoma.

Abstract

Contrast-enhanced MRIs are necessary to delineate the primary gross tumor volume (GTVp) in radiotherapy of nasopharyngeal carcinoma (NPC). However, using contrast agents to scan contrast-enhanced MRIs is not applicable to some patients due to metal implants or their allergy, and it increases the treatment cost of patients. To address these problems, this work aims at synthesizing contrast-enhance MRIs from unenhanced MRIs by implementing generative adversarial network (GAN). In this work, 324 MRI datasets of patients with NPC were retrospectively collected between September 2016 and September 2017 from a single institute. MRI examinations were performed with un-enhanced T1-weighted (T1) and T2-weighted (T2) sequences, and contrast-enhanced T1-weighted (T1C) and fat-suppressed T1-weighted (T1FSC) sequences. We designed and developed a modified pix2pix network to synthesize T1C (sT1C) and T1FSC (sT1FSC) from real T1. The end of the generator in this network was assembled with multiple heads (the classification head and gradient head) to learn more representation information and features from real images, the discriminator in this network distinguished whether the synthesized image is real and fake and supervised that the generator outputs more realistic synthesized image. We verified the performance of the synthesized images for automated delineation of GTVp. In an independent testing set of 11 patients, the synthesized sT1C and sT1FSC were inputted into the segmentation deep learning network along with their corresponding T1 and T2 sequences to generate GTVp contours. Delineation performance of the synthesized images and real images for automated delineation were evaluated by dice similarity coefficient (DSC), and average surface distance (ASD), using human expert contours as the ground truth. In automated contouring of GTVp for NPC, the segmentation deep learning network using one or two synthesized MRIs showed equivalent performance when compared with the automated contours which generated from four real MRI sequences. Mean DSCs between automated contours by sT1C-replaced or sT1C and sT1FSC-replaced network and ground truth contours were 0.726 ± 0.143 and 0.711 ± 0.157, respectively, slightly inferior to that of contours generated from four real MRI sequences (0.740 ± 0.154, both P >0.05). In terms of mean ASD, there was also no significant difference between automated contours generated from synthesized images and real images (3.056 ± 4.216 mm and 3.537 ± 4.793 mm vs. 3.124 ± 4.637 mm; both P > 0.05). We proposed an MRI-synthesis method based on GAN and the synthesized contrast-enhanced MRIs performed equivalent as the real contrast-enhanced MRIs in the automated delineation of gross tumor volume for radiotherapy of NPC.