A Multi-constraint Hybrid Network for Ultrasound Image Synthesis

Simulating ultrasound (US) through US image synthesis is crucial for preoperative training and learning ultrasound techniques for doctors. Due to the significant modality differences between magnetic resonance (MR) images and US images, the structure of the synthesized US images tends to be blurry and of lower quality. In this paper, we propose a hybrid network combining CNN and Transformer that enhances the synthesis performance by integrating their respective abilities in local and global feature extraction. We also introduce multiple constraints to improve the quality of the synthesized images. Firstly, we introduce a multi-scale perceptor (MSP) to enhance the structural content information of the synthesized images, thereby improving their authenticity. Additionally, we address the blurriness issue of synthesized images by incorporating a gradient difference loss in the loss function, which enhances the structural edge information and further improves the quality of synthesized images. Experimental results demonstrate that our proposed method achieves a PSNR of 14.7251 and an SSIM of 0.8556, outperforming other advanced methods currently available.