Rethinking Dual-Domain Undersampled MRI reconstruction: domain-specific design from the perspective of the receptive field

TL;DR

The paper tackles the challenge of undersampled MRI reconstruction by rethinking dual-domain design through the lens of receptive field. It introduces two domain-specific modules, K-GLIM for global k-space initiation and I-PLDE for parallel local image-domain detail enhancement, embedded in a DuDoRNet+-style backbone. Across the IXI dataset, DuDoRNet+ consistently outperforms image-domain, dual-domain, and reference-guided baselines in PSNR and SSIM under 4x–8x accelerations, validating the value of domain-aware design. The findings highlight that tailoring network components to the distinct challenges of k-space interpolation and image restoration yields practical gains, with potential extensions to multi-coil reconstruction and varied sampling patterns.

Abstract

Undersampled MRI reconstruction is crucial for accelerating clinical scanning. Dual-domain reconstruction network is performant among SoTA deep learning methods. In this paper, we rethink dual-domain model design from the perspective of the receptive field, which is needed for image recovery and K-space interpolation problems. Further, we introduce domain-specific modules for dual-domain reconstruction, namely k-space global initialization and image-domain parallel local detail enhancement. We evaluate our modules by translating a SoTA method DuDoRNet under different conventions of MRI reconstruction including image-domain, dual-domain, and reference-guided reconstruction on the public IXI dataset. Our model DuDoRNet+ achieves significant improvements over competing deep learning methods.

Figures (2)

• Figure 1: Probabilityofafeasibleinterpolation$P$w.r.t.thesizeofreceptivefieldandaccelerationrate$a$when$R_{acs}$=0.125.
• Figure 2: FrameworkofarecurrentblockinDuDoRNet+.Itisconstructedbyadomain-specificShallowFeatureExtraction(X-SFE),GlobalFeatureRefinement(GFR)and4-stagedomain-specificbuildingblocks(X-$S_i$).Buildingblocks'colorfollowthenotationinthecolortable.ConvolutionisfollowedbyReLUunlessnoted.