CN-122023712-A - Gaussian expression driven three-dimensional magnetic resonance enhanced imaging method

Abstract

The invention discloses a Gaussian expression driven three-dimensional magnetic resonance enhanced imaging method. The method comprises the steps of obtaining an undersampled target image, and obtaining a 3D enhanced image by utilizing a trained image reconstruction model for the target image, wherein in the process of training the image reconstruction model, voxels of the 3D enhanced image are modeled into a mean value and a variance of a learnable Gaussian kernel function, a plurality of voxels are modeled into a magnitude of the learnable Gaussian kernel function, and the whole 3D enhanced image is formed by a self-adaptive splitting method. The invention utilizes the unified Gaussian continuous 3D expression, can carry out integrated modeling on the flat scanning image and the enhanced image only by relying on a group of flat scanning, ensures the consistency of the structure and improves the accuracy and the efficiency of image reconstruction.

Inventors

• LIANG DONG
• Cui Zhuoxu
• LIU CONGCONG
• JIA SEN

Assignees

• 中国科学院深圳先进技术研究院

Dates

Publication Date

20260512

Application Date

20251224

Claims (8)

1. 1. A Gaussian expression driven three-dimensional magnetic resonance enhanced imaging method comprises the following steps: acquiring an undersampled target image; For the target image, obtaining a 3D enhanced image by using a trained image reconstruction model; in the process of training the image reconstruction model, 3D enhanced image voxels are modeled into a mean value and a variance of a learnable Gaussian kernel function, a plurality of voxels are modeled into a magnitude of the learnable Gaussian kernel function, and an integral 3D enhanced image is formed through an adaptive splitting method.
2. 2. The method of claim 1, wherein the total loss function for training the image reconstruction model is set to: Wherein: Wherein, the Representing the data-fidelity constraint term(s), Represents a refinement of the a priori constraint terms, Representing the contrast preservation constraint terms, Representing the motion correction constraint term(s), 、 And Is the coefficient of the corresponding constraint term, Representing a magnetic resonance multichannel forward measurement operator, Showing the sweep and the enhanced contrast, For the multi-channel k-space measurement data under this comparison, Is the mark of the horizontal scanning image, Is the identification of the enhanced image(s), The weight coefficient is represented by a number of weight coefficients, Is a reference image of the object to be processed, Representing a 3D pan-scan image, A 3D enhanced image is represented and, , A learnable parameter representing all gaussian kernel functions, r representing the spatial position of the gaussian kernel during splitting, Representing the c-th contrast image.
3. 3. The method of claim 2, wherein the 3D pan image and the 3D enhancement image are obtained according to the following formula: Wherein: Wherein, the Representing the nth gaussian kernel function, n being the gaussian kernel index, Representing the entire set of spatial unknowns, The learnable parameters representing all gaussian kernel functions, Representing a 3D pan-scan image, A 3D enhanced image is represented and, Is a global contrast scaling factor that is used to determine the quality of the image, Representing spatially correlated local enhancement terms, Is the baseline intensity parameter.
4. 4. The method of claim 2, wherein a c-th contrast image is rendered When using a gaussian after motion, it is expressed as: Wherein, the In order to sweep or enhance the corresponding intensity, Is the true center after the motion at the c-th scan, Is the covariance after the motion at scan c.
5. 5. The method of claim 1, wherein the target image comprises cardiac imaging, three-dimensional brain structure imaging, three-dimensional vessel wall imaging.
6. 6. The method of claim 4, wherein the step of determining the position of the first electrode is performed, And Calculated according to the following formula: Wherein, the And Expressed under the normative anatomical space, the first The resting position and shape of the individual gaussian kernels, Representing a learnable parameter controlling the change of contrast, Representing a learnable bias controlling the contrast of the enhanced region.
7. 7. 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 6.
8. 8. 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 6 when the computer program is executed.

Description

Gaussian expression driven three-dimensional magnetic resonance enhanced imaging method Technical Field The invention relates to the technical field of magnetic resonance imaging, in particular to a three-dimensional magnetic resonance enhanced imaging method driven by Gaussian expression. Background Three-dimensional magnetic resonance enhanced (3D) imaging can provide isotropic spatial resolution, more continuous anatomy, relative to 2D imaging. Small lesions, thin blood vessels, and irregularly strengthened boundaries are more complete in 3D, thereby improving disease diagnosis accuracy. However, 3D magnetic resonance enhanced imaging scans often require acquisition for lengthy periods of time, are prone to motion artifacts, and the like. K-space undersampling is often an effective acceleration means, but recovering a clean image from undersampled K-space is often modeled as a reconstruction problem. In the prior art, a conditional diffusion model OADiff from flat scanning to enhancement is proposed, which uses contour information (Canny edges) of flat scanning images as a condition, and strong constraint structures are adopted in each step of back diffusion, so that anatomical structure consistency is maintained when an enhancement map is generated, and skeleton deformation caused by noise disturbance in the traditional diffusion denoising process is avoided. Meanwhile, the scheme also designs a multi-frequency band enhanced attention Module (MFEA) for explicitly distinguishing high-frequency textures from noise so as to protect focus edges and details, and the multi-frequency band enhanced attention module is matched with non-uniform time step sampling to reduce iteration and accelerate training and sampling while guaranteeing quality. In addition, a special motion correction frame is designed for motion artifacts which are easy to generate in enhanced imaging, and in a variation frame for jointly optimizing rigid motion parameters and a target image, a structure-guided weighted TV (sTV) regularization is adopted, and a clear structure of reference contrast is borrowed by aligning a target image gradient field with a reference image gradient field. Meanwhile, for the problem of flat scan/enhanced Joint reconstruction in high acceleration 3d T1 weighted brain tumor imaging, there is a study to propose a multi-scale transducer (Joint-Uformer) which jointly encodes pre-/post-contrast images in a network input and attention mechanism, and simultaneously utilizes cross-contrast (inter-attention) and co-contrast (intra-attention) association to learn the structure and contrast relationship between two contrasts without explicitly introducing k-space data consistency. While existing enhanced imaging schemes achieve a certain effect, the human anatomy is forcibly discretized into a fixed grid, which keeps the real biological tissue (e.g., vessel walls, nerve fibers) continuous and smooth, while voxel grids essentially sample the real saw tooth, once the grid is determined, the image resolution is locked. When the lesion is smaller than the voxel size, partial volume effects may occur, resulting in a blurring of the tiny blood vessels. Furthermore, deep learning methods rely heavily on enhanced data sets, which limits the clinical application of the model. Whereas the idea of gaussian representation is to model a 3D image by stacking multiple overlapping 3D gaussian clouds (Kernels) with different shapes and intensities, providing a more continuous signal representation, with natural advantages for 3D image modeling. Meanwhile, the Gaussian expression provides continuous dependence of 3D enhanced images without any training set setting, and the dependence of a deep learning method on a data set is solved. However, the basic gaussian expression still lacks design for contrast preservation and elimination of motion artifacts in magnetic resonance 3D enhanced imaging, thereby limiting its application in magnetic resonance 3D enhanced imaging. Through analysis, in the current 3D magnetic resonance enhanced imaging (such as tumor, liver or brain blood vessel wall enhanced imaging), since the contrast medium is often required to be injected to perform a secondary scan, the same spatial position is required to be unchanged for the secondary scan and the first time, and since the interval is required to be tens of minutes, the patient is difficult to adhere, motion artifact is easy to introduce, and the spatial position change may be caused by two signal acquisitions to affect the final diagnosis. Therefore, it is important to accelerate the scanning (or signal acquisition) process. However, accelerated scanning tends to have severe aliasing artifacts in the image, where it is necessary to perform an image reconstruction task, i.e., to recover a clean image from the image with aliasing artifacts. Currently, deep learning techniques enable optimal reconstruction quality. Among them, typical deep learning methods for 3D