US 11,988,733 B2
Cross-domain network based magnetic resonance imaging undersampling pattern optimization and reconstruction method
Ruiliang Bai, Hangzhou (CN); Zhaowei Cheng, Hangzhou (CN); and Xinyu Jin, Hangzhou (CN)
Assigned to ZHEJIANG UNIVERSITY, Hangzhou (CN)
Appl. No. 18/036,884
Filed by ZHEJIANG UNIVERSITY, Hangzhou (CN)
PCT Filed Dec. 13, 2022, PCT No. PCT/CN2022/138662
§ 371(c)(1), (2) Date May 14, 2023,
PCT Pub. No. WO2023/124971, PCT Pub. Date Jul. 6, 2023.
Claims priority of application No. 202111682080.X (CN), filed on Dec. 31, 2021; and application No. 202210176538.2 (CN), filed on Feb. 25, 2022.
Prior Publication US 2023/0324486 A1, Oct. 12, 2023
Int. Cl. G01R 33/56 (2006.01); G01R 33/48 (2006.01); G06T 7/00 (2017.01); G06T 7/168 (2017.01); G06T 7/174 (2017.01); G06T 11/00 (2006.01)
CPC G01R 33/5608 (2013.01) [G01R 33/4818 (2013.01); G06T 7/0012 (2013.01); G06T 7/168 (2017.01); G06T 7/174 (2017.01); G06T 11/005 (2013.01); G06T 11/006 (2013.01); G06T 2207/10088 (2013.01); G06T 2207/20056 (2013.01); G06T 2207/20076 (2013.01); G06T 2207/20084 (2013.01)] 10 Claims
OG exemplary drawing
 
1. A magnetic resonance imaging (MRI) undersampling and reconstruction method based on a cross-domain network, wherein the method comprises:
(1) acquiring and preprocessing head magnetic resonance (MR) images, and obtaining simulated full-sampled k-space data through Fourier Transform;
(2) separating real and imaginary parts of the simulated full-sampled k-space data obtained in step (1), and saving real and imaginary parts in two matrices with same dimensions independently, then merging the real and imaginary parts into two channels as input for the cross-domain network;
(3) constructing the cross-domain network, which includes an undersampling layer, an Inverse Fourier Transform layer, and a reconstruction network, by training the network with data obtained in step (2) as input, the undersampling layer simulating a process of k-space undersampling in real scenes, the Inverse Fourier Transform layer connecting the Fourier domain with an image domain and obtaining undersampled MR images via the Inverse Fourier Transform, the reconstruction network recovering details of the undersampled MR images to obtain a final image, after completing the training, and obtaining a trained cross-domain network;
(4) using the trained cross-domain network obtained in step (3) to undersample and reconstruct head MR images:
(4-1) setting different sampling rates of the undersampling layer in step (3), optimizing the trained cross-domain network to obtain probability matrices and corresponding reconstruction networks under the different sampling rates, generating optimal undersampling trajectories based on the probability matrices and regional sampling distance constraints;
(4-2) quantitatively analyzing the relationship between the probability matrices and the different sampling rates according to the probability matrices under the different sampling rates, obtaining functional expressions of a 3D probability curve Pface, a central probability curve Pcenter and a marginal probability curve Pmargin by data fitting;
(4-3) based on the functional expressions of Pface, Pcenter, Pmargin, and the regional sampling distance constraints, generating probability matrices and undersampling trajectories under the different sampling rates;
(4-4) undersampling k-space data based on the undersampling trajectories obtained in step (4-3) and generate undersampled MR images via the Inverse Fourier Transform layer, the reconstruction network reconstructing the undersampled MR images to recover the details.