MRecover: AI Model Recovers Motion-Corrupted MRI Using Synthetic Contrast
A group of researchers has developed MRecover, a conditional generative model that transforms standard T1-weighted MRI scans into T2-weighted turbo spin echo (TSE) images. This model uses autoregressive slice conditioning to maintain volumetric consistency. It was trained on a dataset of 577 7T MRI images and showed excellent fidelity, with an SSIM of 0.84 and FSIM of 0.94, even adapting well to 3T data. In the ADNI3 dataset, which had motion artifacts, MRecover improved the analyzable subjects by 31.8% after quality control (593 compared to 450). Additionally, the synthesized images provided larger effect sizes for differences in hippocampal subfield atrophy, which is crucial for accurate segmentation.
Key facts
- MRecover is a conditional generative model for recovering motion-corrupted MR images.
- It synthesizes T2w TSE images from T1w images using autoregressive slice conditioning.
- Trained on 7T MRI data from 577 subjects.
- In-domain testing (n=148) achieved SSIM=0.84 and FSIM=0.94.
- Generalized to 3T data: subfield volumes correlated r=0.87-0.97 (n=416).
- In ADNI3 dataset, analyzable subjects increased from 450 to 593 (31.8% more).
- Effect sizes for hippocampal subfield atrophy improved (ε²=0.121-0.100 vs 0.086-0.06).
- The model targets motion artifacts in hippocampal subfield segmentation.
Entities
Institutions
- ADNI3