Multi-contrast MRI Reconstruction with Deformable Attention and Invertible Network

Tiancheng Zhang, Xinyi Zhang

Abstract: As an effective technique to accelerate magnetic resonance imaging (MRI), multi-contrast reconstruction (MCR) has been widely studied in recent years. The core problems of MCR are: 1) how to extract structurally consistent information from complex reference images, and 2) how to effectively fuse the multi-contrast features. Several methods have been proposed to solve the above two problems, however, they still show limited performance. In this paper, we propose a novel and effective method by exploiting a deformable attention module and an information-lossless invertible neural module for effective feature extraction and feature fusion. To be specific, the deformable attention module first calculates the cross-attention between the multi-contrast images and then applies a content-aware deformable convolution to adaptively sample useful information from the reference images. Using the extracted features, the proposed invertible neural module enables effective feature fusion to generate the expected high-quality target contract images. Extensive experiments on two benchmark multi-contrast MRI datasets demonstrate that our method outperforms state-of-the-art algorithms both visually and quantitatively.

External IDs:dblp:conf/icmcs/ZhangZ24