Sampling Pattern Optimization for Multi-Contrast MRI with Fully Unrolled Reconstruction Network
Presentations
SU-E-207-3 (Sunday, 7/10/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Room 207
Purpose: To develop a fully unrolled deep learning (DL) framework for multi-contrast (T₁ and T₂) MR image reconstruction with joint optimization of k-space sampling patterns (SPs) of images to accelerate standard MRI acquisition.
Methods: Following a variable splitting formulation, a fully unrolled neural network (FU-net) consists of two unrolled Unets for each contrast reconstruction but with cross-links between the two Unets for multi-contrast mutual information learning. Eight fully sampled head and neck T₁ and T₂ weighted scans were retrospectively undersampled for training, 3 for validation, and 5 for testing. An image l₂ loss was used as the training loss function. The image quality was evaluated by peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM). Reconstructed T₁ and T₂ weighted images from FU-net were compared with ones from two separate unrolled Unets (S-net) and a single multi-channel unrolled Unet (MC-net). 1D and 2D SPs were optimized for both contrasts with FU-net by modeling them as i.i.d. samples from a learnable multivariate Bernoulli distribution, compared with 1D Gaussian, 2D variable density (VD) and 2D uniform random samplings. Acceleration rates of acquisition (R) were of 4, 8, and 12. The number of parameters of all networks were kept approximately the same.
Results: The FU-net outperformed S-net and MC-net for all SPs and Rs. The learned SPs further improved the performance of FU-net in both 1D and 2D settings and all R’s. The 2D learned SP showed up to 25% and 45% improvements in respective PSNR and SSIM compared with uniform random sampling.
Conclusion: This work showed effectiveness of FU-net in multi-contrast reconstruction and the advantage of data-driven SPs over empirical ones, which may further accelerate multi-contrast imaging. Deterministic and non-Cartesian SP optimization, and advanced multi-contrast representation learning such as deep dictionary learning will be investigated in the future.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH R01 EB016079.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email