- The paper introduces a self-supervised framework that learns optimal k-space partitioning across multiple contrasts, eliminating the reliance on fully-sampled data.
- It integrates dual-domain loss and a modified VarNet architecture to achieve high SSIM (up to 0.980) and low NMSE on benchmark datasets.
- The method generalizes across various under-sampling patterns and acceleration factors, demonstrating significant potential for clinical MRI applications.
Optimized Multi-Contrast Self-Supervised MRI Reconstruction via Learned k-space Partitioning
Introduction
The paper addresses the challenge of reconstructing high-fidelity MR images from under-sampled k-space acquisitions, focusing on clinical protocols that involve multi-contrast MRI. Traditional approaches accelerate MRI scans through under-sampling but require sophisticated reconstructions to recover high-quality images. While deep learning (DL) methods leveraging multi-contrast information have demonstrated enhanced reconstruction performance, these conventionally depend on supervised learning with fully-sampled k-space data. However, fully-sampled data is often impractical or unavailable in clinical settings, motivating self-supervised approaches. The study proposes an optimized self-supervised multi-contrast reconstruction framework in which k-space partitioning is learned end-to-end for each contrast, eliminating the reliance on reference fully-sampled data and advancing both practical and theoretical aspects of accelerated MRI.
Methodology
Multi-Contrast Self-Supervised Learning
The core innovation is the extension of self-supervised learning via data undersampling (SSDU) to joint multi-contrast reconstruction. The k-space from each contrast is independently partitioned into two disjoint sets, with the reconstruction network mapping between these sets. Channel-wise concatenation of the partitioned k-spaces across contrasts enables the network to exploit inter-contrast information. The loss is computed in both k-space and image space for each contrast, yielding multiple loss terms and facilitating dual-domain learning.
Learned Partitioning Strategy
Unlike heuristic, hand-chosen partitioning, the partitioning mask for each contrast is learned via a probabilistic model, parameterized by a sigmoid function. Sampling employs the reparameterization trick to enable backpropagation through stochastic sampling. The straight-through estimator provides meaningful gradients for learning partitioning probabilities. Partitioning distributions are dynamically optimized per contrast, improving fidelity beyond fixed or heuristically chosen partitioning.
Reconstruction Architecture
The methodology is compatible with any reconstruction network. The primary implementation utilizes a modified VarNet, an end-to-end variational network with recursive k-space data consistency and channel-concatenated U-Net refinement, unrolled for 12 cascades. The architecture accommodates multi-contrast input, scaling parameters and filters per contrast.
Datasets and Evaluation
Experiments are conducted on BraTS 2019 (simulated k-space for brain tumor imaging, four contrasts) and M4Raw (real low-field acquisitions, three contrasts). Multiple under-sampling patterns (1D/2D variable density, equal-spaced) and acceleration factors (R=4,6,8) are tested. Performance metrics used are SSIM and NMSE, computed after coil combination and masking to quantify image quality within signal regions.
Results
The proposed approach achieves measurable improvements in SSIM and NMSE over all contemporary self-supervised methods. Specifically, multi-contrast self-supervised reconstruction surpasses single-contrast self-supervised and even the supervised baselines (with access to fully-sampled k-space) at high acceleration factors. Incorporating learned partitioning further boosts fidelity, particularly in equal-spaced patterns where theoretical optimal partitioning is unknown.
Numerical Highlights
- On BraTS, for equal-spaced parallel imaging (R=8): Multi-contrast learned partitioning attains SSIM up to 0.980 and NMSE down to 0.0024, dominating other self-supervised variants.
- On M4Raw (R=6, 1D variable density): SSIM improves for all contrasts when using multi-contrast learned partitioning (SSIM ~0.887 vs. ~0.875 for single-contrast).
Partitioning Distribution Analysis
Learned partitioning distributions qualitatively mimic the initial undersampling distribution and manifest characteristic patterns (e.g., center-focused star shapes). Partitioning acceleration converges to approximately 1.3, demonstrating robust consistency across training setups.
Ablation and Generalizability
Ablation studies confirm the necessity of dual-domain loss and learned partitioning, each incrementally improving performance. The strategy generalizes to architectures beyond VarNet, such as IWNeXt, achieving comparable performance to supervised multi-contrast methods.
Contrast Scaling
The inclusion of additional contrasts yields monotonically increasing SSIM and decreasing NMSE, reinforcing the claim that aggregate k-space coverage and denoising via contrast transfer constitute key benefits.
Discussion
The empirical findings demonstrate that multi-contrast self-supervised MRI reconstruction, especially when augmented with learned k-space partitioning, achieves higher fidelity images than traditional supervised single-contrast methods. The performance advantage intensifies at higher acceleration factors, reflecting the ability of the network to exploit aggregate k-space coverage from multiple contrasts. The learned partitioning is primarily non-segregated across contrasts; explicit enforcement of segregated partitioning may further enhance performance—a direction for future research.
The dual-domain loss formulation and flexible reconstruction architectures enable broad applicability, including to real-world, unregistered multi-contrast data. Limitations include potential performance inflation in retrospectively under-sampled datasets and lack of explicit spatial alignment modules for unregistered data, both warranting future investigation.
Implications and Future Directions
Practically, this method enables high-quality, rapid MRI acquisitions in settings where fully-sampled k-space is unattainable, and facilitates protocol reduction in clinical workflows. Theoretically, the approach solidifies self-supervised learning as a competitive paradigm for inverse problems in medical imaging. Future work should investigate prospectively under-sampled multi-contrast protocols, segregated partitioning, and more sophisticated attention-based architectures. Extensions to multi-modal and denoising tasks are anticipated, as are integration with spatial alignment networks for improved robustness in real-world clinical contexts.
Conclusion
The paper demonstrates that multi-contrast self-supervised learning with optimized, learned partitioning yields superior MRI reconstruction performance, especially under aggressive acceleration and with limited training data. The approach is highly generalizable and establishes a new standard for practical, reference-free deep learning MRI reconstruction (2606.19182).