CW-GAN: Controllable-Weighting Generative Adversarial Networks for Cross-Domain Multi-Contrast MR Image Synthesis

Motivation: Clinical scan time constraints limit the number of modalities and weighting degrees of images acquired. Generating high-quality images with controlled weighting from limited data can significantly enhance diagnostic accuracy. Goal(s): To develop a generalizable CW-GAN for synthesizing MRI images with controllable contrast and weighting degrees, while improving cross-dataset generalization. Approach: CW-GAN learns target modality features through adversarial training, using special masks as priori information and data enhancement for improved generalization. Results: CW-GAN produces high-quality MR images with adjustable weighting, outperforming CycleGAN and pGAN in both image quality and controllability. Impact: CW-GAN provides a powerful tool for MRI image synthesis with controlled weighting, outperforming existing methods in generalization and controllability, offering valuable potential for clinical applications and advancing MRI deep learning-based tasks.