(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: To compare the fast MRI performance using the Cartesian undersampling scheme against that using a non-Cartesian undersampling scheme by deep learning in terms of preserving image quality and biological information for diffusion tensor imaging (DTI).
Methods: Seventeen DTI brain scans from the TCIA archive were used, each with twelve diffusion-weighted data sets with b = 1,000 s/mm2 and one with b = 0 s/mm2. To simulate the fast MR acquisition, two Cartesian undersampling schemes by a quarter and one-sixth, and two non-Cartesian undersampling schemes by one-sixth and one-eighth retrospectively sampled the full k-space to derive the undersampled k-space data, which will be reconstructed by the DC-CNN network. For each DTI data set, three slices were randomly selected for evaluation, and the remaining data were used for the training reconstruction models. The reconstructed images using the above four undersampling schemes were quantitatively analyzed by total relative error (TRE) and mean structure similarity (MSSIM) for image quality, as well as Fractional Anisotropy (FA), scaled Relative Anisotropy (sRA) maps and principle direction maps for biological information.
Results: For image quality, the reconstructed images by the DC-CNN network using the 16.7% and 12.5% non-Cartesian undersampling scheme are measured to have TRE values of 0.068 and 0.088, MSSIM values of 0.60 and 0.51; in contrast, the 25% and 16.7% Cartesian undersampling yielded TRE values of 0.097 and 0.042, MSSIM values of 0.60 and 0.47. For biological information, ADC and FA maps derived by non-Cartesian undersampling and Cartesian undersampling all show comparable to the reference ADC and FA maps derived from the fully sampled data.
Conclusion: This early work demonstrated the feasibility of fast DTI and indicates that non-Cartesian undersampling outperformed the Cartesian undersampling on accelerating the DTI acquisition while preserving comparable image quality and biological information.
Funding Support, Disclosures, and Conflict of Interest: The DTI data used in this project referred to Barboriak, Daniel. (2015). Data From RIDER_NEURO_MRI. The Cancer Imaging Archive. http://doi.org/10.7937/K9/TCIA.2015.VOSN3HN1.
Keywords
Taxonomy
IM- MRI : Image Reconstruction
Contact Email
