Learning the Sampling Pattern for MRI.

Area of Science:

• Medical Imaging
• Signal Processing
• Machine Learning

Background:

• Compressed sensing theory allows solving inverse problems with incomplete measurements.
• Long acquisition times in Magnetic Resonance Imaging (MRI) limit its clinical utility.
• Optimizing MRI acquisition involves balancing scan time and image reconstruction quality.

Purpose of the Study:

• To develop a supervised learning framework for learning sparse MRI sampling patterns.
• To achieve an optimal trade-off between MRI acquisition time and reconstructed image quality.
• To demonstrate the generalizability of learned sampling patterns across different MRI data.

Main Methods:

• Utilizing a supervised learning approach with training data of measurements and ground-truth images.
• Implementing a framework capable of learning arbitrary sampling patterns, including Cartesian, spiral, and radial.
• Evaluating the performance of learned sparse patterns on a test set of brain images.

Main Results:

• A learned sampling pattern utilized only 35% of k-space for 192x192 brain images.
• Reconstructions achieved a high mean SSIM of 0.914 on a test dataset.
• The supervised learning approach proved effective even with a small training set (7 pairs).

Conclusions:

• Learned sparse sampling patterns can significantly accelerate MRI acquisition without compromising image quality.
• The proposed framework offers a flexible method for optimizing MRI data acquisition.
• Supervised learning is a viable strategy for discovering efficient MRI sampling strategies.