CN-121981938-A - Implicit neural characterization regularization self-supervision depth unfolding magnetic resonance imaging method

Abstract

The invention discloses a self-supervision depth unfolding magnetic resonance imaging method for implicit neural characterization regularization. The method comprises the steps of obtaining undersampled k-space measurement data for a target part, performing inverse Fourier transform on the undersampled k-space measurement data after zero filling to obtain an input image, and obtaining a corresponding reconstructed image for the input image by utilizing a trained image reconstruction network, wherein the image reconstruction network comprises a regularization module and a data consistency module, the regularization module is used for capturing inherent priori information of the input image, and the data consistency module is used for ensuring consistency of the reconstructed image and a physical acquisition model. The invention utilizes the continuous representation capability of depth fusion implicit neural characterization and the physical guided expansion reconstruction framework, and can recover high-quality images under the condition of high-power undersampling.

Inventors

• ZHU YANJIE
• XU JINGRAN
• LIU YUANYUAN
• WANG YINING
• LIANG DONG
• ZHENG HAIRONG

Assignees

• 中国科学院深圳先进技术研究院

Dates

Publication Date

20260505

Application Date

20251205

Claims (10)

1. 1. Implicit neural characterization regularization self-supervision depth unfolding magnetic resonance imaging method Acquiring undersampled k-space measurement data for a target site; Performing inverse Fourier transform after zero filling on the undersampled k-space measurement data to obtain an input image; For the input image, obtaining an optimized reconstructed image using a trained image reconstruction network; The image reconstruction network comprises a regularization module and a data consistency module, wherein the regularization module is used for capturing inherent prior information of the input image, and the data consistency module is used for guaranteeing consistency of the reconstructed image and a physical acquisition model.
2. 2. The method of claim 1, wherein the image reconstruction network obtains an optimized reconstructed image based on the following optimization problem: Wherein, the Is an input image to be reconstructed which is, Is the undersampled k-space measurement, Representing a data-fidelity term, Is a regularized term that is used to determine the degree of regularization, Is a regularization parameter that balances the data fidelity term and regularization term.
3. 3. The method of claim 2, wherein the optimization problem is represented as two sub-problems: Wherein, the And Respectively represent the first The estimated image and intermediate variables in the multiple iterations, The representation acts on The regularization term on the top of the table, Representing the auxiliary variable.
4. 4. The method of claim 1, wherein a loss function training the image reconstruction network is set to: Wherein: Wherein, the Is a data-fidelity item and, Is the total variation loss term of the method, Is a regularization parameter which is a function of the data, Representing raw undersampled k-space measurement data, Is the reconstructed image of the object, Representing the gradient operator(s), , Is the forward measurement matrix.
5. 5. The method of claim 3, wherein the regularization module is implemented by a neural network based on implicit neural characterization, expressed as: Wherein, the Representing the function of the code of the coordinates, Representing the position of the coordinates in space, Is a multi-layer perceptron.
6. 6. The method of claim 5, wherein the multi-layer perceptron comprises two hidden layers, each hidden layer having 64 neurons.
7. 7. The method of claim 5, wherein the coordinate encoding function is implemented using a multi-resolution hash encoding function Instant-ngp.
8. 8. The method of claim 3, wherein the step of, The value of (2) is 1.
9. 9. A computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor realizes the steps of the method according to any of claims 1 to 8.
10. 10. A computer device comprising a memory and a processor, on which memory a computer program is stored which can be run on the processor, characterized in that the processor implements the steps of the method according to any of claims 1 to 8 when the computer program is executed.

Description

Implicit neural characterization regularization self-supervision depth unfolding magnetic resonance imaging method Technical Field The invention relates to the technical field of magnetic resonance imaging, in particular to a self-supervision depth unfolding magnetic resonance imaging method for implicit neural characterization regularization. Background Magnetic resonance imaging is a non-invasive, non-radiative imaging technique that has become an important tool in clinical practice. However, the long scan time remains a key technical bottleneck, limiting its wide clinical application. Therefore, a rapid imaging method based on k-space undersampling has attracted a great deal of attention. Conventional rapid imaging methods, such as compressed sensing and deep learning-based methods, generally require a large amount of full-sample data for training, and have a large computational burden, which still faces challenges in achieving high acceleration multiples. The existing rapid magnetic resonance imaging methods can be mainly divided into two types, namely methods based on traditional image reconstruction, such as compressed sensing and parallel imaging, and the methods are used for image reconstruction by utilizing the sparse or low-rank characteristic of signals and redundancy of multi-coil acquisition, but the methods have limited imaging quality under high acceleration multiple. Another is a deep learning-based approach that utilizes neural networks to learn an end-to-end mapping from undersampled k-space or artifact corrupted images to fully sampled k-space or artifact free images, but most still rely on large amounts of fully sampled data for training and have limited generalization ability in the event of imaging parameters or sampling patterns changes. Recently, a self-supervision learning method is widely focused, and the method only uses undersampled data as supervision signals to guide model optimization, so that the problem of data scarcity is effectively relieved. However, existing self-supervised learning methods tend to perform unstably under high-power acceleration conditions, resulting in limited reconstruction performance. The drawbacks of the prior art, analyzed, are mainly manifested in the following aspects: 1) Traditional image reconstruction methods rely on a priori of manual design and have limited acceleration ratios; 2) The deep learning-based method requires a large amount of completely sampled data for training, and has poor generalization capability when the sampling mode and imaging parameters are changed; 3) Under the condition of high-power undersampling, the self-supervision learning method cannot effectively keep the reconstruction quality, so that the quality of a reconstruction result is reduced. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide a self-supervision depth expansion magnetic resonance imaging method with implicit neural characterization regularization. The method comprises the following steps: acquiring undersampled k-space measurement data for a target site; Performing inverse Fourier transform after zero filling on the undersampled k-space measurement data to obtain an input image; For the input image, obtaining an optimized reconstructed image using a trained image reconstruction network; The image reconstruction network comprises a regularization module and a data consistency module, wherein the regularization module is used for capturing inherent prior information of the input image, and the data consistency module is used for guaranteeing consistency of the reconstructed image and a physical acquisition model. Compared with the prior art, the invention has the advantages that aiming at the problem that the quality of the reconstructed image is reduced under the condition of undersampling data in the existing rapid magnetic resonance imaging method, the self-supervision depth unfolding magnetic resonance rapid imaging method based on implicit neural characterization (INR) regularization is provided, the high-quality image can be recovered under the condition of high undersampling rate by only utilizing single undersampled data, an external training data set is not needed, and the high-quality image reconstruction is realized under high acceleration multiple. Other features of the present invention and its advantages will become apparent from the following detailed description of exemplary embodiments of the invention, which proceeds with reference to the accompanying drawings. Drawings The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description, serve to explain the principles of the invention. FIG. 1 is a flow chart of a self-supervised depth-expanded magnetic resonance imaging method of implicit neural characterization regularization, according to one embodiment of the invention; FIG. 2 is a gene