CT-MR Synthesis of Medical Images based on Transformer Fusion
Zhicheng Wei, Weijie Huang, Shuhao Liu, Xingyu Li, Bingxuan Cheng, Menghua Zhang · 2025
In medical imaging, acquiring MR scans is costly and time-consuming, prompting interest in synthesizing MR images from CT using deep learning. To address issues such as detail loss and structural inconsistency in existing methods, we propose a generative adversarial framework incorporating a High-Frequency Feature Mask Fusion Transformer (HF Transformer) for CT-to-MR translation. A high-pass filter is applied to the CT input to extract edge-enhanced features, which, along with the original CT image, are processed through parallel encoders. The extracted features are fused via the HF Transformer and decoded to generate the MR image. Experimental results show that our method outperforms baseline models in MAE, PSNR, and SSIM, and better preserves bone marrow signals and key structures like the lumbar vertebral plate.