Noise2noise Deep Learning Based Acceleration for MRI Echo-Planar Imaging
L Qin1*, C Lindsay2, A Konik1, G Young3, (1) Dana-Farber Cancer Institute, Boston, MA, (2) University Of Massachusetts Chan Medical School, Worcester, MA (3) Brigham And Women's Hospital, Boston, MA
Presentations
SU-E-207-1 (Sunday, 7/10/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Room 207
Purpose: Single-shot echo-planar imaging (EPI) has been widely used in functional MRI. Long echo train used in EPI leads to blurring and susceptibility artifacts. To ameliorate these artifacts, parallel imaging and compressed sensing techniques have been used to accelerate k-space traversal and shorten the echo train. Recently, deep learning (DL) approaches have been applied to MR image reconstruction to further accelerate MRI scan by acquiring less k-space data than has previously been required while maintaining image quality. One challenge in DL based reconstruction is the requirement of a large number of ground-truth images for training. In our study, we propose to use the Noise2Noise (N2N) DL training methodology, which can recover signals from noisy images without needing noise-free ground truth images.
Methods: The public dataset ACRIN-DSC-MR-Brain was downloaded from the TCIA website. EPI images from DSC acquisition were subsampled to remove 90% of the k-space to simulate noisy acquisitions. Each image was sampled twice with different noise realizations to produce an N2N training pair. The N2N was trained using a modified U-Net architecture, with 3722 training samples split at 80% for training and 20% for validation for 100 epochs.
Results: Our method shows that we can improve the overall quality of the EPI images through denoising (Fig.1). At a 90% reduction in k-space, we were able to improve the Peak Signal-to-noise Ratio by 3.5 decibels and the similarity (Structural Similarity Index Metric) by .286 (~60% improvement) relative to the noisy image.
Conclusion: Our preliminary results showed that subsampled EPI images can be highly improved using N2N DL network. The next step is to test the technique in a prospective trial including large number of real positive cases with various pathologies, to ensure that the denoised images do not introduce unexpected artifacts or decrease the conspicuity of various types of abnormalities.
Keywords
Echo Planar Imaging, Noise Reduction, SNR
Taxonomy
IM- MRI : Image Reconstruction
Contact Email