Deep Learning Applications for MRI Image Processing and Artifact Reduction

Magnetic resonance imaging (MRI) is a non-invasive medical imaging modality. It offers the capability to produce high-quality multi-contrast diagnostic images without the use of ionizing radiation. The resulting images not only provide detailed anatomical information but also capture functional process, making MRI invaluable for clinical diagnosis, treatment planning, and biomedical research. However, the full potential of MRI remains partially obscured by persistent challenges, notably the presence of imaging artifacts that compromise the quality of acquired images. The advent of potent graphics processing unit (GPU)-based computational platforms and the availability of open-access datasets provide previously unachievable opportunities to address these issues through deep learning methods.

For the signal-to-noise ratio (SNR) issues, a complex-valued convolutional network (CNN) incorporating the noise level map (non-blind ℂDnCNN) was trained with ground truth and simulated noise-corrupted image pairs. The network was validated using both simulated and in vivo data collected from low-field scanners. The non-blind ℂDnCNN showed superior quantitative metrics and significantly improved the SNR and visual quality of the image. By incorporating the noise level map, the method showed better performance on dealing with spatially varying parallel imaging noise.

For the motion artifacts, a multi-task conditional generative adversarial network (MT-cGAN) was developed for simultaneous motion detection and compensation. The training images were generated with a realistic artifact simulation process, incorporating comprehensive rigid motion profiles, noise addition, and parallel imaging acquisitions. Performance was evaluated using both simulated and real in-vivo data. For the motion detection task, MT-cGAN achieved the best classification accuracy on simulated and real in-vivo dataset. For the motion compensation task, the outputs of MT-cGAN showed less visual blurring, fewer residual artifacts, and better preservation of fine structures compared to other models.

For off-resonance artifacts, a deep-learning-based method (AutofocusNet) was developed to correct both field inhomogeneities and concomitant files in spiral MRI. By training the network using images with simulated field inhomogeneity and concomitant field effect, AutofocusNet showed superior performance compared to the conventional autofocus method. It offers a practical and effective solution for off-resonance correction in spiral MRI without using the field map or computing the concomitant fields during the reconstruction.

For undersampled cardiac MRI reconstruction, a novel complex-valued cascading cross-domain CNN was proposed, named C3-Net, for improved balance between computation demands and image quality for accelerated CMR. C3-Net outperformed other comparison methods, especially at high acceleration rates (> 8). The short-axis results from C3-Net showed reduced residual artifact and improved temporal fidelity of cardiac motion.