Edge information preserving generative adversarial network combining Transformer and multi-scale fusion for MRI modality transfer

Abstract Magnetic Resonance Imaging (MRI) is a multi-sequence imaging technology that assists clinicians in diagnosis. However, due to factors such as long scanning time, inability of patients to cooperate, and high cost of obtaining the full sequence, clinically only Part of the comparison sequence. However, often missing sequences can help diagnose clinical diseases. Therefore, generating missing MRI sequences is very important for clinical diagnosis. However, most existing methods for MRI generation lack the learning of global context information, thereby losing part of the task-related feature information, or only consider the pixel-level difference between the generated image and the real image, ignoring the details and details of the image. Texture information affects the quality of image generation. In response to the above problems, a edge information preserving generative adversarial network combining Transformer and multi-scale fusion (ETF-GAN) for MRI modality transfer was proposed.. This model combined with Transformer improves the accuracy of To capture global context information, the Iterative Multi-scale Concatenate Fusion (IMCF) module is proposed in the decoder part to enhance the decoder’s ability to utilize features of different scales, while also introducing Sobel Edge Extraction (SEE) module to retain the edge information of the image and improve the quality of image generation. Experiments show that the proposed method is superior to other advanced MRI modality transfer methods in quantitative comparison and visual comparison.