GCL-RCA-Net: an RCA-RemUNet-enhanced graph-contrastive framework for hybrid data-driven and dual-branch physics-guided parallel MRI reconstruction.

OBJECTIVE: Parallel magnetic resonance imaging (pMRI) with Cartesian equispaced undersampling accelerates acquisition but introduces structured aliasing that degrades reconstruction quality, especially at high acceleration factors. Although recent self-supervised and contrastive reconstruction methods have shown promise, most operate slice-wise and do not capture cross-slice anatomical continuity or align learned representations with physics-guided reconstruction. APPROACH: To address these limitations, we propose GCL-RCA-Net, an RCA-RemUNet-enhanced graph-contrastive framework for pMRI reconstruction under Cartesian equispaced undersampling. The method combines measurement-consistent data-driven initialization, a physics-guided dual-branch module, anatomy-aware graph contrastive priors, and multi-stage cross-domain refinement. The initial stage stabilizes reconstruction by directly restoring acquired k-space samples and learning missing data. A dual-branch module then exploits complementary physics through calibration-based generalized autocalibrating partially parallel acquisitions (GRAPPA) interpolation in k-space and sensitivity-encoding (SENSE)-based conjugate gradient refinement in the image domain, followed by adaptive branch selection and fusion. To model anatomical context, a self-supervised graph contrastive module encodes cross-slice continuity and non-local structural relationships, providing descriptors that guide the cascaded refinement network. Finally, a multi-stage lattice-style cross-domain fusion module jointly refines the data-driven estimate and fused physics output through coupled image- and k-space pathways, followed by final data consistency enforcement. RESULTS: On fastMRI brain data with acceleration factors from 2× to 8×, and in cross-anatomy evaluation on fastMRI knee data, GCL-RCA-Net consistently outperforms recent baselines under equispaced Cartesian undersampling. Relative to the strongest competing method, it achieves PSNR gains of 2.66-4.35 dB on brain and 0.31-3.03 dB on knee reconstruction, with corresponding improvements in structural similarity. SIGNIFICANCE: The proposed framework shows that combining measurement-consistent data-driven estimation, physics-guided dual-branch refinement, and anatomy-aware graph contrastive priors yields robust and generalizable pMRI reconstruction under severe undersampling, with strong potential for faster, higher-quality, and clinically reliable MRI.