Reconstruction of multicontrast MR images through deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Long magnetic resonance (MR) scan times degrade image quality due to patient motion, discomfort, and increased costs.
• Rapid MR imaging techniques are crucial for improving patient experience and reducing healthcare expenses.
• Deep learning offers a promising avenue for accelerating MR image acquisition and reconstruction.

Purpose of the Study:

• To develop and evaluate a novel deep-learning network for the joint reconstruction of multicontrast brain MR images from down-sampled data.
• To accelerate the MR imaging data acquisition process while maintaining diagnostic image quality.
• To compare the performance of the proposed network against conventional reconstruction methods.

Main Methods:

• Development of two deep-learning networks, X-net and Y-net, for reconstructing T1- and T2-weighted MR images from down-sampled data.
• Investigation of optimal sampling patterns, patch sizes, and acceleration factors for network training.
• Comparison with single- and joint-reconstruction parallel-imaging, compressed-sensing algorithms, and a conventional U-net.
• Evaluation using metrics such as structural similarity (SSIM), normalized mean square error (NMSE), and Fréchet inception distance (FID).

Main Results:

• The proposed X-net and Y-net demonstrated statistically significant improvements in image quality compared to a single subnetwork.
• Uniform down-sampling patterns yielded better image quality than random or central patterns.
• The developed networks outperformed U-net, compressed-sensing, and parallel-imaging algorithms in SSIM and NMSE.
• A generative adversarial network (GAN)-enhanced Y-net produced more realistic images with a better FID score.

Conclusions:

• The proposed X-net and Y-net networks effectively reconstruct full MR images from down-sampled data, surpassing conventional methods.
• The integration of a GAN further enhanced image realism.
• These deep learning approaches hold significant potential for accelerating multicontrast anatomical MR imaging in clinical settings.