Causality-driven dual-domain network enhanced by Gaussian splatting for magnetic resonance image reconstruction.

Magnetic resonance imaging (MRI) is a crucial tool in modern clinical diagnostics due to its non-invasive nature and high-resolution imaging capabilities. However, existing MRI reconstruction methods face two major challenges: (i) reconstruction artifacts that degrade image quality, and (ii) a lack of causal modeling, as most methods focus solely on correlations while ignoring confounding factors such as varying sampling strategies, dataset heterogeneity, and temporal fluctuations. To address these issues, we propose CauGD2-Net - a causality-driven dual-domain network enhanced by Gaussian splatting for magnetic resonance image reconstruction. First, a temporal causal graph (TCG) is constructed to explicitly model causal relationships throughout the reconstruction process. To further identify and eliminate spurious correlations and confounding factors, we introduce intermediate variables to build a phased temporal causal graph (PTCG). To mitigate checkerboard artifacts, we propose a (CGS) mechanism that enhances spatial consistency during reconstruction. Furthermore, a causal dual-domain low-rank (CD2LR) module is developed to effectively integrate spatial and frequency domain features in a low-rank manner, reducing parameter complexity while preserving multi-dimensional feature representations. Extensive experiments under varying acceleration factors demonstrate the superiority of the proposed approach. Results on the publicly available IXI dataset and two in-house clinical datasets show that CauGD2-Net outperforms baseline methods by an average of 5.35/6.69 in PSNR and 0.058/0.064 in SSIM, respectively.