MRecover: A Conditional Generative Model for Recovering Motion- Corrupted MR images Using AI Generated Contrast

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• High-resolution T2-weighted (T2w) turbo spin echo (TSE) MRI is crucial for hippocampal subfield segmentation.
• T2w TSE sequences are highly sensitive to motion artifacts, often resulting in significant data loss and reduced analytical power.
• Existing methods struggle to mitigate motion artifacts effectively, limiting the utility of valuable MRI datasets.

Purpose of the Study:

• To develop and validate a novel conditional generative model, MRecover, for synthesizing T2w TSE MRI from routine T1w images.
• To address data loss caused by motion artifacts in MRI scans, particularly for hippocampal subfield analysis.
• To improve the quality and quantity of analyzable data for neurodegenerative disease research.

Main Methods:

• A conditional generative model (MRecover) was developed using autoregressive slice conditioning to synthesize T2w images from T1w images.
• The model was trained on 7 Tesla (7T) MRI data (n=577) and validated on both in-domain (n=148) and out-of-domain 3T data (n=416).
• Performance was evaluated using metrics like Structural Similarity Index Measure (SSIM) and Feature Similarity Index Measure (FSIM), and by comparing hippocampal subfield volumes and diagnostic group differences.

Main Results:

• MRecover demonstrated high fidelity in synthesizing T2w images, achieving SSIM of 0.84 and FSIM of 0.94 on in-domain data.
• The model generalized effectively to 3T data, with synthesized and acquired subfield volumes showing strong correlation (r=0.87-0.97).
• Application to the ADNI3 dataset increased analyzable subjects by 31.8% (593 vs 450) and revealed larger effect sizes for hippocampal subfield atrophy in diagnostic groups.

Conclusions:

• The MRecover model successfully synthesizes high-quality T2w MRI from T1w images, effectively mitigating motion artifacts.
• This approach significantly enhances the proportion of analyzable subjects and improves statistical power for detecting neurodegenerative changes in hippocampal subfields.
• MRecover offers a promising solution for maximizing the utility of existing and future MRI datasets in clinical research and diagnostics.