Click here to

Session: MRI: New Algorithms, Techniques, and Applications II [Return to Session]

Accelerated MRI Reconstruction Using Variational Feedback U-NET with Transfer Learning

Y Zhou1*, R Paul2, P Ding2, R Battle1, A Patel1, B Li2, (1) Mayo Clinic Arizona, Phoenix, AZ, (2) CIDAS, Arizona State University, Tempe, Az, United States

Presentations

SU-F-207-4 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room 207

Purpose: In this work, we propose a Variational Feedback U-NET (VFN) with Transfer Learning approach, aiming at speeding up AI model training and reconstructing high-quality images from accelerated MRI acquisition.

Methods: We use a Convolutional Neural Network (CNN) to learn the differences between the aliased images and the original images, employing a modification of the widely used U-Net architecture. Considering the peculiarity of the down-sampled k-space data, we introduce a new term to the loss function in learning, which effectively employs the given k-space data during training to provide additional regularization on the update of the network weights. In addition, the transfer learning is applied to the pre-trained model to improve the performance. The fastMRI brain datasets from NYU public database and 140 ACR phantom datasets collected with different slice thickness, orientations, and spatial resolutions were used to validate our newly developed AI based MRI reconstruction model and transfer learning approach. Peak signal to noise ratio (PSNR) and structural index similarity (SSIM) are used in image quality assessment.

Results: Three deep learning models were obtained: Model A was trained with fastMRI brain T1 training datasets with the training time about 100 hours; Model B was trained with 140 ACR Phantom datasets with the training time about 15 hours; Model C was a hybrid deep transfer learning model, which was fine tuned by ACR phantom based on the pre-trained model A. We used both the human brain images from fastMRI and ACR phantom datasets to evaluate the performance. Both PSNR and SSIM are improved significantly by using the transfer learning approach.

Conclusion: Our studies show promising results in reconstruction of 4x accelerated MRI human brain and phantom images. Transfer learning has shown the capability to train the AI model with significantly reduced training time while improving the image quality.

Keywords

MRI, Reconstruction, Image Processing

Taxonomy

IM- MRI : Image Reconstruction

Contact Email

Share: