MRecover AI recovers motion-corrupted MRI scans, boosting usable data by 31.8%

Motion artifacts are a persistent problem in high-resolution MRI, especially for T2-weighted turbo spin echo (TSE) sequences needed for hippocampal subfield segmentation. A team from the University of Pittsburgh and collaborators introduces MRecover, a conditional generative model that synthesizes TSE images from standard T1-weighted scans. The model uses autoregressive slice conditioning to maintain volumetric consistency across slices. Training on 577 ultra-high-field 7T MRI scans allowed the model to learn the mapping between contrast types. In-domain testing on 148 subjects yielded strong fidelity (SSIM 0.84, FSIM 0.94).

Critically, MRecover generalized to lower-field 3T clinical data. When applied to the motion-affected ADNI3 dataset, the model recovered 31.8% more analyzable scans after quality control (593 vs 450 subjects). Subfield volumes from synthesized images closely matched as-acquired ones (correlation r=0.87-0.97). This increase in sample size led to larger effect sizes for detecting hippocampal atrophy differences between diagnostic groups (whole hippocampus epsilon-squared 0.121-0.100 vs 0.086-0.062). The approach reduces data loss from motion artifacts without requiring rescanning, potentially accelerating clinical research.