MoRE: A Mixture-of-Experts-Based Task-Adaptive End-to-End Network for Multimodal MRI Reconstruction

Although accelerated MRI reconstruction has advanced rapidly through end-to-end learning, deploying a single unified network that generalizes across diverse anatomies and contrasts under constrained computational resources remains challenging. In this paper, we introduce MoRE, a sparsely activated mixture-of-experts (MoE) module integrated into an end-to-end variational network. MoRE couples a shared encoder with sample-wise, unsupervised routing to activate a minimal subset of expert decoders while strictly preserving physics-based data consistency. Evaluated on the fastMRI multi-coil brain and knee datasets under 8x undersampling, MoRE achieves highly stable SSIM and PSNR performance across multi-contrast datasets. Furthermore, t-SNE visualization of the routing embeddings reveals interpretable, modality-aware expert specialization. The sparse conditional computation mechanism ensures that the architectural overhead remains modest. These results demonstrate that MoE-style capacity scaling can significantly enhance general-purpose MRI reconstruction without requiring proportional increases in computational power.

Branch Learning in MRI: More Data, More Models, More Training

We investigated two complementary strategies for multicontrast cardiac MR reconstruction: physics-consistent data-space augmentation (DualSpaceCMR) and parameter-efficient capacity scaling via VQPrompt and Moero. DualSpaceCMR couples image-level transforms with kspace noise and motion simulations while preserving forwardmodel consistency. VQPrompt adds a lightweight bottleneck prompt; Moero embeds a sparse mixture of experts within a deep unrolled network with histogram-based routing. In the multivendor, multisite CMRxRecon25 benchmark, we evaluate fewshot and out-of-distribution generalization. On small datasets, k-space motion-plus-noise improves reconstruction; on the large benchmark it degrades performance, revealing sensitivity to augmentation ratio and schedule. VQPrompt produces modest and consistent gains with negligible memory overhead. Moero continues to improve after early plateaus and maintains baseline-like fewshot and out-of-distribution behavior despite mild overfitting, but sparse routing lowers PyTorch throughput and makes wall clock time the main bottleneck. These results motivate scale-aware augmentation and suggest prompt-based capacity scaling as a practical path, while efficiency improvements are crucial for sparse expert models.