New self-supervised MRI method uses multiple contrasts, no full scans needed

Deep learning has accelerated MRI by reconstructing high-quality images from under-sampled data, but most approaches rely on supervised learning with fully sampled k-space for training—expensive and often impractical. A new paper from Oxford researchers Brenden Kadota, Charles Millard, and Mark Chiew tackles this with a self-supervised multi-contrast framework. Their method jointly trains on multiple under-sampled contrasts, learning to map between partitioned sets of k-space data (inspired by the SSDU approach). Crucially, they also learn an optimal partitioning probability distribution for each contrast, sampled to generate masks that improve reconstruction fidelity.

Tested on two publicly available multi-contrast MRI datasets, the proposed approach outperforms current single-contrast self-supervised methods. The key contribution: removing the need for fully sampled k-space while using multi-contrast information—common in clinical scans (e.g., T1, T2, FLAIR)—to boost reconstruction quality. This could translate to shorter scan times and higher image fidelity in practice, all without requiring expensive fully sampled training data. The authors plan to release code and note that learned partitioning alone adds marginal computational overhead during inference.