Self-auditing residual drifting model accelerates knee MRI while preserving pathology

Accelerated MRI sacrifices image quality, especially for clinical pathology. A new method from researchers at Wake Forest and RPI introduces SA-RDM-DC (Self-Auditing Residual generative Drifting Model with Data Consistency). Instead of slow iterative diffusion, it trains a drift field from zero-filled reconstruction toward fully sampled residual correction. The model simultaneously predicts image and k-space residuals, enforces data consistency with acquired measurements, and uses frequency-aware supervision to recover fine detail. Critically, it produces dense error maps and slice-level risk scores in the same inference—essentially auditing its own failures.

Evaluated on multi-coil fastMRI knee data at acceleration factors 4, 8, and 12, SA-RDM-DC outperformed UNet-image-SENSE, DC-UNet, score diffusion, ELF-Diff, SENSE-VarNet, and MoDL. It achieved the highest SSIM across all acceleration factors while running subsecond per slice—avoiding the long runtime of diffusion models. Pathology-aware analysis using fastMRI+ annotations showed better lesion-region structural fidelity and reduced meniscus prediction instability. On SKM-TEA with protocol shift, its self-auditing scores transferred partially as a selective-review signal. This work suggests that combining generative drifting with self-auditing enables fast, reliable accelerated MRI suitable for clinical use.