Multi-modal MRI cascaded incremental reconstruction with coarse-to-fine spatial registration.
Authors
Affiliations (4)
Affiliations (4)
- Engineering Research Center of Digital Forensics, Ministry of Education, CICAEET, Nanjing University of Information Science and Technology, Nanjing, 210044, Jiangsu, China.
- Engineering Research Center of Digital Forensics, Ministry of Education, CICAEET, Nanjing University of Information Science and Technology, Nanjing, 210044, Jiangsu, China. Electronic address: [email protected].
- School of Computer and Information Science, Qinghai institute of technology, Xining, 810016, Qinghai, China.
- School of Computer Science, Nanjing University of Posts and Telecommunications, Nanjing, 210023, Jiangsu, China.
Abstract
Magnetic resonance imaging (MRI) typically utilizes multiple contrasts to assess different tissue features, but prolonged scanning increases the risk of motion artifacts. Compressive sensing MRI (CS-MRI) employs computational reconstruction algorithm to accelerate imaging. Full-sampled auxiliary MR images can effectively assist in the reconstruction of under-sampled target MR images. However, due to spatial offset and differences in imaging parameters, how to achieve cross-modal fusion is a key issue. In order to cope with this issue, we propose an end-to-end network integrating spatial registration and cascaded incremental reconstruction for multi-modal CS-MRI. Specifically, the proposed network comprises two stages: a coarse-to-fine spatial registration sub-network and a cascaded incremental reconstruction sub-network. The registration sub-network iteratively predicts deformation flow fields between under-sampled target images and full-sampled auxiliary images, gradually aligning them to mitigate spatial offsets. The cascaded incremental reconstruction sub-network adopts a new separated criss-cross window Transformer as the basic component and deploys them in dual-path to fuse inter-modal and intra-modal features from the registered auxiliary images and under-sampled target images. Through cascade learning, we can recover incremental details from fused features and continuously refine the target images. We validate our model using the IXI brain dataset, and the experimental results demonstrate that, compared to existing methods, our network exhibits superior performance.