Multi-Contrast MRI Synthesis Using Generative Adversarial Networks
Abstract
Purpose
Multi-contrast MRI—including T1-weighted contrast-enhanced (T1-CE), T2-weighted (T2), and T2-FLAIR—is essential for brain radiotherapy planning but increases scan time, patient discomfort, and exposure to gadolinium-based contrast agents (GBCA). We propose a dual-scale 2.5D generative adversarial network (GAN) to synthesize these contrasts from a single T1-weighted scan.
Methods
A conditional GAN with dual-scale discriminators was developed to capture local texture and global anatomical structure. A 2.5D strategy incorporating neighbouring slices was used to preserve 3D anatomical consistency while maintaining the computational efficiency of 2D models. Training employed a hybrid loss combining 2D adversarial, feature-matching, and perceptual losses with a 3D volumetric consistency loss computed over reconstructed volumes. The model was trained and evaluated on the BraTS 2021 dataset (1,000 training, 50 validation, 150 testing cases). Image quality was assessed using SSIM and PSNR, and clinical utility was evaluated via automatic tumour segmentation using Dice similarity and Hausdorff distance.
Results
The 2.5D GAN consistently outperformed 2D synthesis. For T1-CE, SSIM and PSNR were 0.95 ± 0.02 (p<0.0005) and 22.1 ± 2.5 dB (p<0.006), compared with 0.94 ± 0.03 and 21.9 ± 2.5 dB. For T2, SSIM and PSNR were 0.95 ± 0.03 (p<0.004) and 21.6 ± 2.3 dB (p<0.005), versus 0.94 ± 0.02 and 21.4 ± 2.1 dB. For T2-FLAIR, SSIM and PSNR reached 0.94 ± 0.02 (p<0.0005) and 21.3 ± 1.7 dB (p<0.0005), compared with 0.93 ± 0.02 and 20.8 ± 1.7 dB. Tumour segmentation improved using 2.5D synthetic images, with Dice scores increasing from 0.49 ± 0.22 to 0.53 ± 0.23 for tumour core and from 0.76 ± 0.10 to 0.81 ± 0.09 for whole tumour.
Conclusion
The dual-scale 2.5D GAN enables efficient generation of anatomically consistent multi-contrast MRI from a single T1 scan, supporting reduced scan time and GBCA exposure in brain radiotherapy workflows.