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 following steps: acquiring an under-sampled target image; and for the target image, a trained image reconstruction model is utilized to obtain a 3D enhanced image, in the process of training the image reconstruction model, voxels of the 3D enhanced image are modeled into a learnable Gaussian kernel function mean value and variance, multiple voxels are modeled into a learnable Gaussian kernel function amplitude, and an overall 3D enhanced image is formed through an adaptive splitting method. According to the method, unified Gaussian continuous 3D expression is utilized, integrated modeling can be carried out on the plain-scanned image and the enhanced image only by relying on a group of plain-scans, the structure consistency is ensured, and the accuracy and efficiency of image reconstruction are improved.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, and more specifically, to a Gaussian expression-driven three-dimensional magnetic resonance enhancement imaging method. Background Technology

[0002] Compared to 2D imaging, three-dimensional magnetic resonance imaging (3D) offers isotropic spatial resolution and more continuous anatomical structures. Small lesions, fine blood vessels, and irregular enhancement boundaries are more complete in 3D, thus improving diagnostic accuracy. However, 3D MRI scans often require excessively long acquisition times and are prone to motion artifacts. K-space undersampling is often an effective acceleration method, but recovering a clean image from undersampled K-space is typically modeled as a reconstruction problem.

[0003] In existing technologies, one approach proposes a conditional diffusion model, OADiff, that moves from planar to enhanced images. It uses the contour information (Canny edges) of the planar image as a condition, strongly constraining the structure in each back-diffusion step. This maintains anatomical consistency when generating the enhanced image, avoiding skeletal deformation caused by noise disturbances during traditional diffusion denoising. Simultaneously, this approach incorporates a multi-band enhancement attention module (MFEA) to explicitly distinguish high-frequency textures from noise, preserving lesion edges and details. Combined with non-uniform time-step sampling, it reduces iterations and accelerates training and sampling while maintaining quality. Furthermore, a dedicated motion correction framework is designed to address motion artifacts easily generated in enhanced imaging. In a variational framework that jointly optimizes rigid body motion parameters and the target image, a structure-guided weighted TV (sTV) regularization is employed. By aligning the gradient field of the target image with the gradient field of the reference image, it leverages the clear structure of the reference contrast. Meanwhile, to address the problem of joint reconstruction of plain / enhanced images in high-acceleration 3D T1-weighted brain tumor imaging, a study proposed a multi-scale Transformer (Joint-Uformer). This transformer jointly encodes pre- / post-contrast images in the network input and attention mechanism, while utilizing inter-attention and intra-attention associations to learn the structural and contrast relationships between the two contrasts without explicitly introducing k-space data consistency.

[0004] While existing augmentation techniques have achieved some success, they all forcibly discretize human anatomical structures into a fixed mesh. This renders real biological tissues (such as blood vessel walls and nerve fibers) continuous and smooth, whereas voxel meshes are essentially jagged samples of reality. Once the mesh is determined, the image resolution is locked. When lesions are smaller than the voxel size, partial volumetric effects occur, leading to blurring of small blood vessels. Furthermore, deep learning methods heavily rely on augmentation datasets, limiting their clinical application. Gaussian representation, on the other hand, models 3D images by stacking multiple overlapping 3D Gaussian clouds (Kernels) with different shapes and brightness, providing a more continuous signal representation and possessing a natural advantage in 3D image modeling. Simultaneously, Gaussian representation provides continuous 3D augmented images without requiring any training set settings, solving the dataset dependency problem of deep learning methods. However, basic Gaussian representation still lacks design for contrast preservation and motion artifact elimination in MRI 3D augmentation, thus limiting its application in this field.

[0005] Analysis reveals that current 3D magnetic resonance imaging (MRI) with contrast enhancement (such as for tumors, liver, or brain blood vessel walls) often requires a second scan after contrast agent injection. This second scan must maintain the same spatial position as the first, and the interval between scans can be several tens of minutes, making it difficult for patients to endure and easily introducing motion artifacts. Furthermore, the two signal acquisitions may cause spatial position changes, affecting the final diagnosis. Therefore, accelerating the scanning (or signal acquisition) process is crucial. However, accelerated scanning often results in severe aliasing artifacts, necessitating image reconstruction to recover a clean image from the aliasing artifacts. Currently, deep learning technology can achieve optimal reconstruction quality. Typical deep learning methods for 3D MRI with contrast enhancement include supervised and self-supervised methods. Supervised methods provide some acceleration by extracting intra-layer structural priors from large datasets, but obtaining fully sampled 3D data is often difficult clinically, especially with enhanced data after 3D drug administration. Self-supervised methods optimize network parameters using undersampled data, offering a solution; however, they often lack detailed modeling of inter-layer relationships, leading to inter-layer misalignment and affecting diagnosis. Furthermore, these two methods lack a unified model for enhancing the similarity of imaging structures and preserving contrast, thus failing to induce contrast changes and resulting in enhancement failures.

[0006] In summary, the existing technology has the following main drawbacks: (1) Traditional models are usually unable to accurately and continuously model 3D images in 3D augmented imaging scenarios, resulting in the inability to accurately depict the interlayer continuity.

[0007] (2) Existing methods usually lack a unified structure sharing mechanism, which causes changes in the contrast of lesions in the enhanced image, thus making it impossible to enhance the appearance of lesions.

[0008] (3) Mainstream models usually rely entirely on massive training data or only rely on a single scan to optimize the model, resulting in a lack of prior data distribution.

[0009] (4) The mainstream model lacks a specific design for motion artifacts, which may cause motion artifacts to affect the anatomical structure during the enhanced scanning process. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a Gaussian expression-driven three-dimensional magnetic resonance imaging method. This method includes the following steps: Acquire the undersampled target image; For the target image, a 3D enhanced image is obtained using a trained image reconstruction model; In the process of training the image reconstruction model, the voxels of the 3D enhanced image are modeled as learnable Gaussian kernel function mean and variance, and multiple voxels are modeled as learnable Gaussian kernel function amplitude. The overall 3D enhanced image is then composed through an adaptive splitting method.

[0011] Compared with existing technologies, the advantages of this invention are that the Gaussian expression-driven 3D magnetic resonance imaging enhancement method, based on the Gaussian expression framework, models a unified enhancement imaging model, enabling the modeling of both 3D MRI plain scans and enhanced images without requiring any training set settings. This invention designs a learnable amplitude to construct a lesion region enhancement contrast preservation mechanism, achieving contrast preservation in enhanced lesion regions. Furthermore, it designs an adaptive learnable method and covariance to adaptively eliminate motion artifacts, preventing potential motion artifacts from interfering with lesion diagnosis during acquisition.

[0012] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0014] Figure 1 This is a flowchart of a Gaussian expression-driven three-dimensional magnetic resonance imaging method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of a Gaussian expression-driven three-dimensional magnetic resonance imaging method according to an embodiment of the present invention. Detailed Implementation

[0015] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0016] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0017] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0018] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0020] To suppress aliasing artifacts, noise, and motion artifacts in rapid MRI-enhanced imaging, this invention provides a Gaussian-expression-driven 3D MRI-enhanced imaging scheme. It designs a continuous representation method for MRI-enhanced images. First, driven by the physical properties of MRI Fourier transform, a continuous Gaussian kernel function representation model of the 3D enhanced image is constructed. The voxels of the 3D enhanced image are modeled as learnable Gaussian kernel function means and variances, and multiple voxels are modeled as learnable Gaussian kernel function amplitudes. An adaptive splitting method is designed to assemble the overall 3D enhanced image. Furthermore, to refine the anatomical structure of lesions in the enhanced image, iteratively optimized intermediate enhanced images are aligned with distributed priors, constructing a data-driven prior integration. Overall, this invention proposes a self-supervised, basic framework-based continuous modeling method for MRI 3D enhanced images, while also incorporating a motion artifact suppression module, providing a high-performance solution for rapid 3D MRI-enhanced imaging.

[0021] Specifically, in combination Figure 1 and Figure 2 As shown, the Gaussian expression-driven three-dimensional magnetic resonance imaging method includes the following steps: Step S1: Construct an image reconstruction model framework, which is used to learn the correspondence between the flat scan image and the enhanced image.

[0022] Image reconstruction models are used to learn the correspondence between plain scanned images and enhanced images, enabling the reconstruction of high-quality enhanced images from undersampled data. In one embodiment, the image reconstruction model includes functional modules such as a data fidelity module (or model), an image thinning module, a contrast preservation module, and a motion correction module. The construction process of the image reconstruction model framework will be described in detail below.

[0023] Step S11: Construct 3D models of scanned and enhanced images For example, we first approximate the grayscale distribution of the entire brain volume (i.e., the target area) using a set of three-dimensional Gaussian kernels. Assume there are N Gaussian kernels, each with three types of parameters: Indicates the first The basic intensity of a Gaussian controls the contribution of this Gaussian to the overall brightness; Indicates the center position; The Gaussian kernel covariance matrix controls the extent, anisotropy, and rotation of the Gaussian kernel in three directions, determining the shape of the ellipsoid. The expression for the basic Gaussian kernel in 3D voxel space is:

[0024] The superposition of all Gaussians yields a continuous field:

[0025] in, Indicates the first n A Gaussian kernel function, r This indicates the spatial location of the Gaussian kernel during the splitting process. n It is a Gaussian kernel index. Let T represent the unknown set of the entire space, and let T denote the transpose. Represents the learnable parameters of all Gaussian kernel functions. This represents a 3D brain volume modeled using a Gaussian continuous representation, unlike voxel modeling, which allows for arbitrary... Both are defined and naturally continuous. In practical solutions, this continuous field needs to be sampled on a regular 3D mesh, i.e. Figure 2 Mid-rasterization step. After rasterization, each voxel center Corresponding to a grayscale value i, j, and k represent the three-dimensional spatial indices, thus obtaining the modeled 3D image. The significance of this step is that subsequent scans and enhancements are performed using the same set of indices. In Gaussian geometry, the intensity is scheduled only once for invariant parameters, while parameters that change (such as contrast) are represented separately. This approach is both physically consistent and interpretable.

[0026] The model in formula (2) above only constructs a basic 3D planar scan image. Subsequently, by designing and combining geometric sharing, the idea of ​​constructing an enhanced 3D image by only making changes in the intensity space is adopted. For example, a set of baseline intensity parameters is introduced. The comparative variations involved in this invention are written as follows:

[0027] In other words, the first sweep under flat sweep The brightness of a Gaussian unit is... Therefore, planar 3D is represented as (denoted as ). ):

[0028] The corresponding enhancements are:

[0029] in, This represents a plain magnetic resonance imaging (MRI) image. This represents an enhanced image of a drug injected via magnetic resonance imaging. It is the global contrast scaling factor. This represents a spatially dependent local enhancement term.

[0030] In this invention, the designed method satisfies the requirement that if there is no contrast agent (or a certain area is not enhanced), then... ,Right now . It is the global contrast scaling factor, which simulates situations where the overall brightness changes by a uniform proportion, such as the baseline effect caused by changes in system gain or overall T1. This is a spatially correlated local enhancement term, which only needs to be significantly positive in areas where the contrast agent is actually distributed, such as lesions or blood vessels; in most normal tissue areas... It will be optimized to be close to 0, ensuring that the enhancement occurs only on a few Gaussians. The final enhancement. Represented as:

[0031] Can be recorded as At this point, the planar scan and the enhanced scan share the same set of spatial positions and variances, resulting in natural structural alignment and a unified structural representation.

[0032] Step S12: Construct the data fidelity module have and After creating two 3D representations, it is necessary to ensure that they truly correspond to the actual acquired data and conform to the forward measurement process of magnetic resonance physics, such as... Figure 2 The first data fidelity item at the bottom is specifically modeled as follows:

[0033] in, This indicates two contrast methods (plain scan and enhancement). Used for marking flat scans Used for identification enhancement, imposing the same type of constraint on each; The multi-channel k-space measurement data under this comparison are the raw observations given by the scanner; A This is the forward measurement operator for multi-channel magnetic resonance imaging. This term is to ensure that, during the parameter tuning process, all images modeled with the parameters must ultimately match the actual acquired data when projected into k-space, thus guaranteeing that the reconstruction does not deviate from actual observations.

[0034] Step S13: Construct the image thinning module It is difficult to fully recover fine structures from a single measurement signal. This invention refines images by constructing coupled high-dimensional distributed priors in an unsupervised manner. Figure 2 The refined priors are shown below:

[0035] Here It is a reference image, an intermediate variable after denoising by Tweedie during the iterative process of the reverse diffusion model. The weights are used to balance the weighting of the generative priors. The core of model (8) is to align the modeled plain and enhanced 3D images with the pre-trained 2D diffusion priors to ensure that the modeled image structure is gradually refined. Since the 3D images modeled by the Gaussian kernel function ensure the continuity of the interlayer relationships, there will be no interlayer bias during the model refinement process. Under the constraints of model (8), the modeled 3D images not only capture the structural data of the current image, but also fall near the brain image manifold learned by the diffusion model, resulting in a more accurate structure.

[0036] Step S14: Construct the contrast preservation module Although at the parameter level While the enhancement method is defined, unnecessary brightness differences may still occur over large areas at the voxel level; or the enhanced contrast may be pulled back a priori, making the actual lesions less prominent. Therefore, in one embodiment, a contrast preservation method is designed, making it a contrast constraint term, expressed as:

[0037] in, It is a planar scan 3D image representation. It is an enhanced 3D image representation. This indicates a contrast constraint.

[0038] In constraint (9), Specifically or This invention does not impose specific limitations on this. The former encourages sparse difference maps, meaning that only a few locations show significant differences, which better reflects the fact that enhancement lesions usually account for a small proportion; while the latter controls the overall difference energy, preventing large-scale shifts in the enhancement map across the entire brain. Combined with contrast transformation, this can be understood as: in most normal tissues, Suppressed to 0, Approaching 1, at this point The contrast is preserved in this estimation. In the lesion area, It can be significantly positive, making the local The contrast retention term is relatively large, and it does not prohibit such local differences, but only prevents the differences from spreading to the entire brain. Therefore, model (9) can highlight the contrast of the lesion area to prevent it from being suppressed during the optimization process and highlight the enhancement significance.

[0039] Step S15: Constructing a motion suppression module Even with well-contrast and prior processing, the difference map will still be filled with artifacts caused by misalignment if the motion between the flat scan and enhancement is not modeled. Therefore, in one embodiment, a motion suppression module is constructed, represented as follows:

[0040] here and This indicates that under standardized anatomical space, the first The static position and shape of a Gaussian kernel. This represents a learnable parameter that controls changes in overall contrast. This indicates a learnable bias that controls the contrast of highlighted areas. ( The special orthogonal group representing three-dimensional space represents the global rigid body parameters of the c-th contrast (plain / enhanced scan), describing the rotation and translation of the brain relative to the normal space during scanning. and Its function is to determine the true center and covariance of the same anatomical Gaussian after motion during the c-th scan. Specifically, when rendering the c-th comparison image, the motion-motion Gaussian is used:

[0041] in, To perform a flat scan or enhance the corresponding intensity (flat scan) Enhancement = From a physical perspective, model (11) explains that what remains unchanged is the anatomical structure of the brain. During the scan, the entire brain rotates and translates relative to the magnetic field coordinate system. This method applies the motion directly to each Gaussian element, rather than interpolating onto the reconstructed voxel image. Through this process, the invention unifies motion, structure, and contrast within the same optimization framework, eliminating the need for additional reconstruction and registration steps and removing potential errors introduced by these extra steps. Furthermore, applying a rigid body transformation to the Gaussian element is analytical, avoiding the blurring of edges that can occur with voxel interpolation.

[0042] Step S16: Construct the motion correction module To suppress motion errors that may be introduced in enhanced imaging, simply add Gaussian... , That's not enough; there must be an optimization objective to drive these parameters to converge in the correct alignment direction. Therefore, in one embodiment, a motion suppression module is designed, denoted as: (12) in, It is a motion correction constraint.

[0043] First, construct the difference plot. If the two images are aligned well, then in the non-lesion area... It should be close to 0, with significant deviations only in the lesion area; if translational or rotational misalignment exists, strong superposition of positive and negative textures will appear in the edge area. It will exhibit high-frequency oscillations. Then, the gradient of the difference map is taken; penalizing the gradient is equivalent to encouraging a smoother difference map, which updates the rigid body parameters. , let and Geometrically coincident, thus allowing In non-lesion regions, the parameters approach a constant value. Therefore, by optimizing the learnable parameters... and Until the flat scan and enhancement overlap spatially, most motion artifacts are eliminated at the source.

[0044] Step S2: Train the image reconstruction model with the goal of minimizing the set total loss function to obtain optimized learnable parameters.

[0045] Merge the modules in step S1, and denote all learnable parameters as:

[0046] In one embodiment, the constructed total loss function is expressed as:

[0047] in, , and These are the weighting coefficients of the relevant terms, which can be determined based on actual needs or simulation. The loss function described above is linearly weighted, but it can also be in exponentially weighted form.

[0048] Ultimately, within this unified framework, the present invention yields a pair of... and The former is a structurally sound, physically and a priori Gaussian 3D representation of a plain scan. The latter, on the same structural basis, is an enhanced representation obtained through parametric contrast transformation and motion transformation, which can highlight the contrast of lesions and is strictly aligned with the plain scan.

[0049] Step S3: For the undersampled target image, the corresponding 3D enhanced image is obtained using the trained image reconstruction model.

[0050] Once the image reconstruction model is trained, optimized learnable parameters can be obtained, which can then be applied to practical image reconstruction. See [link to model application process] for details. Figure 2 In the testing phase, specifically for the actual acquired undersampled data (or low-quality images) of the target, a trained image reconstruction model is used to obtain the corresponding 3D enhanced image. This invention achieves fast, artifact-free 3D enhanced magnetic resonance imaging by designing a Gaussian continuous representation method for 3D enhanced images, driven by a high-dimensional distribution prior.

[0051] It should be noted that, without departing from the spirit and scope of this invention, those skilled in the art can make appropriate changes or modifications to the above embodiments. For example, in addition to enhancing image reconstruction, after appropriate modifications, it can also be used for other three-dimensional imaging categories, such as unsupervised cardiac imaging, three-dimensional brain structure imaging, and three-dimensional blood vessel wall imaging, simply by adjusting the modeling parameters according to the specific anatomical location. This invention can also be used in conjunction with other prior knowledge, such as introducing relaxation priors such as T1, T2, and proton density in quantitative magnetic resonance imaging to form a joint reconstruction framework of Gaussian continuous representation plus quantitative physical priors. Furthermore, the Gaussian continuous representation of this invention is not limited to Gaussian kernel functions; it can also be replaced with other continuous basis function families such as spline functions, radial basis functions, and multi-resolution wavelet bases, as long as it can achieve continuous parameterized expression of three-dimensional volume data. Moreover, with appropriate replacement of the forward physical model, this invention is not only applicable to magnetic resonance imaging but can also be extended to other three-dimensional imaging modalities, such as CT, PET, and photoacoustic tomography. By replacing the magnetic resonance forward operator with the corresponding modality's physical forward operator, the same approach of rapid, high-quality three-dimensional reconstruction can be achieved.

[0052] In summary, this invention utilizes a unified Gaussian continuous 3D representation, enabling integrated modeling of both plain and enhanced images based on only one set of plain scans, ensuring structural consistency and preventing distortion of the enhanced structure. A 2D diffusion model is introduced as a generative prior to refine the 3D enhanced image, improving detail fidelity and suppressing noise and artifacts. Simultaneously, a contrast preservation mechanism prevents contrast drift between the enhanced and plain scans, and motion during acquisition is modeled as a learnable parameter to adaptively suppress motion artifacts, resulting in more stable and reliable 3D enhanced imaging results. This invention provides a magnetic resonance 3D enhanced imaging method that does not require clean enhancement training data, enabling refined modeling of 3D inter-slice correlations and providing a more continuous 3D enhanced image modeling process to deliver detail-preserving 3D enhanced imaging results. Furthermore, to unify the plain and enhanced scan structures while maximizing the preservation of enhanced anatomical structures, a modeling method unifying structure and contrast is designed to prevent contrast shifts in the enhanced structure that could affect subsequent diagnosis, thereby promoting the development of faster and more accurate magnetic resonance 3D enhancement.

[0053] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0054] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0055] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0056] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0057] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0058] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0059] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0060] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0061] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A Gaussian expression-driven three-dimensional magnetic resonance imaging method, comprising the following steps: Acquire the undersampled target image; For the target image, a 3D enhanced image is obtained using a trained image reconstruction model; In the process of training the image reconstruction model, the voxels of the 3D enhanced image are modeled as learnable Gaussian kernel function mean and variance, and multiple voxels are modeled as learnable Gaussian kernel function amplitude. The overall 3D enhanced image is then composed through an adaptive splitting method.

2. The method according to claim 1, characterized in that, The total loss function for training the image reconstruction model is set to: in: in, This indicates a data fidelity constraint. This indicates a refinement of prior constraints. This indicates that the comparison retains the constraint terms. Indicates motion correction constraint terms. , and These are the coefficients of the corresponding constraint terms. This represents the multi-channel forward measurement operator for magnetic resonance imaging. This indicates the contrast between plain scan and enhanced scan. This is the multi-channel k-space measurement data under this comparison. It is a flat scan image identifier. It is to enhance image identification. Indicates the weighting coefficient. This is a reference image. Represents a 3D planar scan image. Represents 3D enhanced images. , Represents the learnable parameters of all Gaussian kernel functions. r This indicates the spatial location of the Gaussian kernel during the splitting process. This represents the c-th comparison image.

3. The method according to claim 2, characterized in that, The 3D flat scan image and the 3D enhanced image are obtained according to the following formula: in: in, Indicates the first n A Gaussian kernel function, n It is a Gaussian kernel index. Represents an unknown set in the entire space. Represents the learnable parameters of all Gaussian kernel functions. Represents a 3D planar scan image. Represents 3D enhanced images. It is the global contrast scaling factor. This represents a spatially dependent local enhancement term. It is the baseline intensity parameter.

4. The method according to claim 2, characterized in that, Render the c-th comparison image When using the Gaussian after motion, it is expressed as: in, To flatten or enhance the corresponding intensity, It is the true center after the motion during the c-th scan. It is the covariance after motion during the c-th scan.

5. The method according to claim 1, characterized in that, The target images include cardiac imaging, three-dimensional brain structure imaging, and three-dimensional blood vessel wall imaging.

6. The method according to claim 4, characterized in that, and Calculate using the following formula: in, and This indicates that under standardized anatomical space, the first The static position and shape of a Gaussian kernel. This represents a learnable parameter that controls the change in contrast. This represents a learnable bias that controls the contrast of the enhanced region.

7. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

Images