An Inpainting-based Method for MRI Synthesis

Motivation: The existing deep learning-based MRI sequence synthesis methods are prone to obscure or simulate pathology. Goal(s): To develop an inpainting-based image synthesis network (IBSNet) for MRI sequence high-fidelity synthesis. Approach: We designed a dual-branch structure to focus on global and local information extraction. Moreover, a joint loss is proposed to constrain the network from signal intensity, structural similarity, and edge preservation. An attention module is used to refine the intermediate feature maps from both the channel and spatial dimensions. Results: The results show that our method outperforms other methods based on the encoder-decoder network, generative adversarial network (GAN), and diffusion model. Impact: Our proposed inpainting-based image synthesis network can generate the target sequence from existing sequences, which can reduce the scanning time of MRIs and improve the patient experience.