Self-supervised learning-based diffeomorphic non-rigid motion estimation for fast motion-compensated coronary MR...

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Artificial Intelligence in Medicine

Background:

• Coronary MR angiography (CMRA) is crucial for diagnosing heart conditions.
• Current CMRA reconstruction is lengthy due to motion correction requirements.
• Respiratory motion significantly degrades CMRA image quality.

Purpose of the Study:

• To accelerate non-rigid motion-corrected CMRA reconstruction.
• To develop a deep learning-based network for rapid non-rigid motion estimation.
• To integrate this network with an efficient undersampled motion-corrected reconstruction.

Main Methods:

• Introduced DiRespME-net, a self-supervised U-Net architecture for respiratory motion estimation.
• Coupled DiRespME-net with an efficient GPU-based iterative reconstruction for motion correction.
• Evaluated DiRespME-net's performance in 12 subjects, comparing vessel sharpness and length against state-of-the-art methods.

Main Results:

• No statistically significant difference in image quality (vessel sharpness and length) was found compared to traditional methods (MC:Nifty-reg).
• DiRespME-net achieved a 50-fold reduction in computation time, with total reconstruction time around 20 seconds.
• Comparable visible vessel length for RCA and LAD between MC:Nifty-reg and MC:DiRespME-net.

Conclusions:

• The proposed self-supervised learning approach enables fast, motion-corrected CMRA reconstruction.
• This method holds significant promise for integration into routine clinical practice.
• Accelerated CMRA acquisition could improve patient throughput and diagnostic capabilities.