A three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method

Abstract

The application discloses a three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method, and relates to the technical field of medical imaging and artificial intelligence. The method first constructs a joint reconstruction optimization model based on WaveCAIPI coding and decoupling multi-contrast convolution dictionary prior, and explicitly decomposes each contrast image into common features and unique features. Then, an alternating minimization strategy is used to expand the iterative solving process into a multi-stage deep network MambaDDL, each stage containing a data consistency module, a unique feature Mamba proximal gradient module and a common feature Mamba proximal gradient module. Finally, a Mamba state space mechanism is used to realize global dependence modeling with low memory consumption. The application can effectively reduce aliasing artifacts, suppress noise amplification and maintain anatomical structure details under high acceleration undersampling conditions, and significantly reduces the memory consumption of three-dimensional reconstruction, thus having high clinical application value.

Description

Technical Field

[0001] This application relates to the fields of medical imaging and artificial intelligence technology, and in particular to a method for joint reconstruction of three-dimensional multi-contrast magnetic resonance imaging. Background Technology

[0002] Multi-contrast magnetic resonance imaging (MRI) can provide complementary tissue contrast information on the same anatomical structure, thus having important applications in brain lesion assessment, diagnosis of neurological diseases, and clinical follow-up. However, multi-contrast MRI usually requires sequential acquisition of multiple sequences such as T1, T2, and FLAIR, resulting in long scan times, which can easily cause patient discomfort and motion artifacts, further affecting image quality and diagnostic reliability.

[0003] In recent years, deep learning has made progress in multi-contrast MRI reconstruction, and related methods can be broadly divided into data-driven methods and depth unrolling methods. The former directly learns the nonlinear mapping of the reconstructed image from undersampled data, but its interpretability and robustness are relatively insufficient; the latter expands the traditional iterative optimization process into learnable modules, improving interpretability while maintaining model constraints. However, existing 3D multi-contrast reconstruction networks often suffer from high memory consumption, and many methods directly fuse different contrast features, which easily introduces contrast-specific information, reducing the interpretability of the representation and the stability of the reconstruction. Summary of the Invention

[0004] Therefore, it is necessary to provide a three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method to address the aforementioned technical problems.

[0005] The following technical solution is adopted in this specification: This specification provides a method for joint reconstruction of three-dimensional multi-contrast magnetic resonance imaging, including: Acquire undersampled multi-contrast magnetic resonance images; A joint reconstruction optimization model for multi-contrast magnetic resonance imaging is established; the joint reconstruction optimization model is solved by an alternating minimization method, and the iterative solution process is expanded into an end-to-end multi-stage deep network. The end-to-end multi-stage deep network includes an initialization module, a cascaded expansion module, and an output module. The cascaded expansion module includes multiple cascaded expansion stage processing units, and each expansion stage processing unit includes a data consistency module, a unique feature Mamba proximal gradient module, and a common feature Mamba proximal gradient module. Based on the multi-contrast magnetic resonance images, the initialization module obtains the common anatomical features shared across contrasts and the unique features of each contrast. For each contrast in each unfolding stage, the magnetic resonance image is updated using the data consistency module based on the current stage's magnetic resonance image, common features, and unique features. Based on the updated magnetic resonance image and common features, the unique features corresponding to the current contrast are updated using the Mamba proximal gradient module according to the Mamba state space mechanism. Based on the updated magnetic resonance image and unique features, the common features are updated using the Mamba proximal gradient module according to the Mamba state space mechanism. Based on the final common features updated after multiple unfolding stages and the unique features corresponding to each contrast, the output module obtains the reconstructed image corresponding to each contrast.

[0006] Furthermore, the specific feature Mamba proximal gradient module and the common feature Mamba proximal gradient module are respectively composed of a gradient descent module and a Mamba proximal mapping module; the Mamba proximal mapping block is used to serialize the input features, model the long-range dependencies in the sequence through the attention state space mechanism, deserialize, and convolutionally refine, and finally output the updated features.

[0007] Furthermore, the construction of the joint reconstruction optimization model for multi-contrast magnetic resonance imaging includes: Construct a WaveCAIPI-encoded image domain forward model, which includes a k-space domain model and an image domain equivalent model; Based on the fully sampled 3D image to be reconstructed, undersampled observation data of the k-th contrast is acquired through the k-space domain model; through the image domain equivalent model, the magnetic resonance images of each contrast are mapped by the coded forward operator to obtain the aliased multi-coil image corresponding to the n-th contrast. A decoupled multi-contrast convolution dictionary prior is introduced to decompose each contrast image into common features and unique features of each contrast. The features are then mapped back to the image space by convolving and summing the common features and unique features. The aliased multi-coil image obtained through the image domain equivalence model, the common features obtained through the prior of the decoupled multi-contrast convolution dictionary, the unique features of each contrast, and the approximation terms of the decoupled convolution dictionary are combined to obtain the joint reconstruction optimization model.

[0008] Furthermore, the formula for calculating the aliased multi-coil image corresponding to the nth contrast is: ; in, This represents the aliasing operator caused by CAIPIRINHA sampling; Indicates the first k Coil sensitivity diagrams corresponding to each contrast ratio; This represents a one-dimensional Fourier transform along the readout direction; This represents a one-dimensional inverse Fourier transform along the readout direction; This represents the point spread function operator introduced by Wave modulation.

[0009] Furthermore, the updating of the magnetic resonance image through the data consistency module includes: For the first contrast in the t-th unfolding stage, the current magnetic resonance image is mapped to the observation domain and compared with the actual aliased observation image; at the same time, the magnetic resonance image is updated by combining the image priors corresponding to common features and unique features.

[0010] Furthermore, the process of updating the specific features corresponding to the current contrast using the specific feature Mamba proximal gradient module is represented as follows: ; in, Indicates the unfolding iteration phase; Indicates the first One contrast ratio; and These represent the contrast-specific latent variables and the common latent variables shared by multiple contrast ratios in the previous stage, respectively. Used to characterize the anatomical information shared between different contrasts; This indicates the number of data points obtained after the current stage has undergone data consistency updates. One contrast-reconstructed image; Indicates the use of updating specific variables The gradient descent module.

[0011] Furthermore, the process of updating the common features through the Mamba proximal gradient module is represented as follows: in, Indicates the unfolding iteration phase; Indicates the first One contrast ratio; and These represent the contrast-specific latent variables and the common latent variables shared by multiple contrast ratios in the previous stage, respectively. Used to characterize the anatomical information shared between different contrasts; This indicates the number of data points obtained after the current stage has undergone data consistency updates. One contrast-reconstructed image; Indicates the use of updating public variables The gradient descent module.

[0012] Furthermore, the multiple contrast ratios include T1, T2, and FLAIR.

[0013] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method.

[0014] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method.

[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: In the three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method provided in this specification, a joint reconstruction optimization model is constructed based on Wave-CAIPI encoding and decoupled multi-contrast dictionary priors, and solved using an alternating minimization algorithm and expanded into a multi-stage deep network; the reconstructed image, common features and unique features are alternately updated through the data consistency module and the Mamba proximal gradient module; finally, the reconstructed images of each contrast are output.

[0016] This invention introduces Wave-CAIPI 3D spatial coding constraints into the multi-contrast joint reconstruction process, and constructs a decoupled multi-contrast dictionary prior model to explicitly decompose each contrast image into common features and unique features. This allows for more full utilization of cross-contrast shared information, while avoiding the erroneous mixing of contrast-specific information during joint modeling, thereby improving image detail preservation and reconstruction stability.

[0017] Furthermore, by introducing the Mamba state space mechanism into the feature update module, long-distance dependencies in 3D multi-contrast MRI can be modeled with lower memory overhead, overcoming the problem that existing 3D depth reconstruction methods have high memory requirements and are difficult to apply to high-dimensional medical images. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This document provides a schematic flowchart of a three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method. Figure 2 This is a schematic diagram of the Wave-CAIPI forward model provided in this specification in the k-space domain and the image domain; Figure 3 This is a schematic diagram of the overall structure of the Mamba-DDL multi-contrast joint reconstruction network provided by the present invention; Figure 4 A schematic diagram of the update process of each unfolding stage of the Mamba-DDL multi-contrast joint reconstruction network provided by the present invention; Figure 5 The present invention provides in A schematic diagram of the reconstruction results under undersampling conditions and the comparison method. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0021] To shorten the scan time of multi-contrast MRI, existing techniques typically employ parallel imaging methods to undersample the k-space and recover the image using coil sensitivity information and reconstruction algorithms. For two-dimensional imaging techniques, since coil sensitivity primarily varies along a single phase coding direction, its coding capability is limited, restricting the achievable speedup. In contrast, three-dimensional imaging techniques can utilize sensitivity variations along both the phase coding direction and the partition coding direction, offering greater speedup potential. Wave-CAIPI technology further introduces readout direction waveform modulation on top of two-dimensional CAIPIRINHA (rapid MRI scanning technique), resulting in a more uniform distribution of aliasing in three-dimensional space, thereby improving coding efficiency and reconstruction stability under high-speed sampling conditions.

[0022] Based on the shortcomings of the existing technology, this invention provides a multi-contrast MRI joint reconstruction method based on Wave-CAIPI encoding and decoupled multi-contrast dictionary priors. This method solves the problems of insufficient parallel imaging encoding capability, large memory overhead of existing three-dimensional depth reconstruction methods, and insufficient mixed modeling of shared and unique information among multiple contrasts in high-acceleration three-dimensional multi-contrast MRI under undersampling conditions, which makes it difficult to achieve high-quality reconstruction and efficient computation at the same time.

[0023] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0024] Figure 1This is a schematic diagram of a three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method described in this specification, which specifically includes the following steps: S1: A joint reconstruction optimization model is constructed based on Wave-CAIPI encoding and decoupled multi-contrast dictionary priors. The Wave-CAIPI model utilizes 3D coil sensitivity information to mitigate aliasing artifacts under extreme acceleration, and the decoupled multi-contrast convolutional dictionary priors decompose each contrast image into common and unique features to capture the correlation between contrasts and avoid inconsistent feature leakage.

[0025] S2: Solve the joint reconstruction model using an alternating minimization strategy, and expand the solution process into a multi-stage deep network. The network is preferably Mamba-DDL (Mamba-Empowered Decoupled Dictionary Learning for Highly Accelerated 3D Multi-Contrast MRI with Wave-CAIPI), including an initialization module, multiple expansion stages, and an output module; each expansion stage includes a data consistency module, a specific feature Mamba proximal gradient module, and a common feature Mamba proximal gradient module.

[0026] S3: In the process of updating unique features and common features, Mamba near-end mapping blocks are introduced. The linear complexity of long-range dependencies is modeled through the attention state space mechanism, which reduces the memory usage in 3D reconstruction tasks while maintaining the ability of global context interaction.

[0027] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects: This invention improves physical coding capabilities under high-speed undersampling conditions by introducing Wave-CAIPI 3D spatial coding constraints into the multi-contrast joint reconstruction process, effectively mitigating aliasing artifacts and noise amplification. By constructing a decoupled multi-contrast dictionary prior model, images of each contrast level are explicitly decomposed into common and unique features, thereby making fuller use of cross-contrast shared information while avoiding erroneous mixing of contrast-specific information during joint modeling, thus improving image detail preservation and reconstruction stability. By unfolding the alternating minimization process into a multi-stage interpretable deep network, this invention combines the physical consistency of model-driven methods with the representational capabilities of data-driven methods.

[0028] Furthermore, by introducing the Mamba state space mechanism into the feature update module, long-range dependencies in 3D multi-contrast MRI can be modeled with lower memory overhead, overcoming the problem that existing 3D depth reconstruction methods have high memory requirements and are difficult to apply to high-dimensional medical images.

[0029] Therefore, this invention can achieve high-quality, low-memory consumption multi-contrast MRI combined reconstruction under high-acceleration sampling conditions, and has high clinical application value and promotion prospects.

[0030] When applying the three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method provided in this manual, it is not necessary to consider... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0031] Furthermore, each part of the present invention will be described in detail below.

[0032] This invention is first based on, as Figure 2 The Wave-CAIPI encoding and decoupled multi-contrast dictionary prior construction jointly reconstruct an optimization model, providing a physical consistency basis for subsequent joint reconstruction and avoiding information leakage and detail loss problems caused by mixed modeling of different contrast features in existing methods. The specific steps are as follows:

[0033] Step 1: Mapping of Wave-CAIPI encoding in the k-space domain.

[0034] set up Indicates the first The k-space observation data with varying contrast is denoted as the corresponding fully sampled image. ,in, For contrast ratio, Number of receiving coils , , These represent the readout direction and the phase encoding direction, respectively. ) and partition coding direction ( The size of the matrix. To determine the size of the readout direction oversampling matrix, and This represents the acceleration factor in the phase coding direction and the partition coding direction. The acquisition process can then be modeled as:

[0035] in, This represents the coil sensitivity diagram corresponding to the nth contrast ratio. To measure noise, Indicates in CAIPIRINHA undersampling operator on the plane This represents a one-dimensional Fourier transform along the readout direction. This represents the two-dimensional Fourier transform along the phase coding and partition coding directions. This represents the point spread function operator introduced by Wave modulation. Wave-CAIPI applies a sinusoidal gradient during readout, making the aliasing more uniformly distributed in three-dimensional space, thereby improving coding capability under high-speed sampling.

[0036] Step 2: Mapping of Wave-CAIPI encoding in the image domain.

[0037] Considering that direct optimization on complete 3D volume data would result in high computational and memory overhead, this invention further derives its image domain equivalent model. Performing an inverse Fourier transform along the spatial dimension on the forward model yields the aliased multi-coil image corresponding to the nth contrast:

[0038] in, For convolution operations, Represents the point spread function in the image domain. Indicates the first Altered images with varying contrast, by It is obtained through three-dimensional inverse Fourier transform.

[0039] Step 3: Multi-contrast decoupling convolution dictionary prior.

[0040] To simultaneously utilize the highly correlated shared anatomical structures among different contrasts while preserving the unique tissue features of each contrast, this invention introduces a decoupled multi-contrast convolution dictionary prior. Specifically, the first... Each contrast image is represented as the convolution sum of common features and unique features:

[0041] in, Indicates the first Common features shared across contrast channels. Indicates the first The unique characteristics of contrast. and These represent common filters and unique filters, respectively. The number of feature channels is represented by the above decoupling representation. This allows for the explicit separation of anatomical information shared between different contrast levels from contrast-specific diagnostic information, thus avoiding inconsistent feature leakage that occurs in traditional direct feature fusion methods.

[0042] Step 4: Multi-contrast joint reconstruction model.

[0043] Based on the aforementioned image domain coding model and the prior knowledge of the decoupled multi-contrast convolution dictionary, this invention constructs a joint reconstruction optimization model as follows: The first item is the multi-contrast Wave-CAIPI data consistency item. Forward operators in the image domain encoded with Wave-CAIPI, the last three terms are prior terms of the decoupled convolution dictionary, mainly including decoupled convolution dictionary approximation terms and unique features. With public characteristics Implicit regularization terms and , and These are the weighting coefficients. and These are convolution kernels representing common features and unique features, respectively.

[0044] The present invention further proposes, as follows: Figure 3 The iterative process illustrated uses an alternating minimization method to unfold into an end-to-end multi-stage deep network (Mamba-DDL) to improve model interpretability and computational efficiency. This network consists of an Initialization Module (IM), 12 unfolding stages (cascaded unfolding modules), and an Output Module (OM). Each unfolding stage includes three core modules: a Data Consistency Module (DCM), which introduces Wave-CAIPI encoding constraints to ensure the reconstruction results are consistent with the actual undersampled measurement data; a Unique Mamba-based Proximal Gradient Module (U-Mamba-PGM), used to restore and refine the unique local structures and diagnostic features of each contrast level; and a Common Mamba-based Proximal Gradient Module (C-Mamba-PGM), used to jointly update and enhance shared anatomical information across different contrast levels. Through this multi-stage iterative unfolding, end-to-end joint optimization reconstruction of multi-contrast MRI images is achieved. The specific steps are as follows:

[0045] 1. Initialization The initialization module is used to initialize the input data based on undersampled data. The initial reconstruction results of the generator network are obtained, and common features are further derived. and unique characteristics The initial estimate provides initial conditions for subsequent multi-stage iterative updates. Its initialization formula is:

[0046] 2. Data consistency update For the nth contrast in the nth unfolding stage, the current reconstructed image is mapped to the observation domain and mixed with the actual observed image. The reconstructed image is updated by comparing and combining image priors corresponding to common and unique features. The optimization objective is:

[0047] 3. Unique Feature Update like Figure 4 As shown, for the first The first of the development phases Each contrast factor in the currently reconstructed image and public characteristics Under known conditions, this invention introduces U-Mamba-PGM to the first The unique features corresponding to each contrast Updates are performed to restore the unique texture, edge, and diagnostic information in each contrast image. U-Mamba-PGM consists of a Gradient Descent Block (GDB) and a Mamba Proximal Mapping Block (Mamba-PMB). The process is defined as follows:

[0048] 4. Update of common features like Figure 4 As shown, for the nth unfolding stage, the reconstructed image at all contrast levels... and corresponding unique features Under known conditions, this invention introduces C-Mamba-PGM for common features shared across contrast ratios. Updates are performed to enhance the representation of common anatomical structures across different contrast levels. C-Mamba-PGM consists of a gradient descent block (GDB) and a Mamba proximal mapping block (Mamba-PMB). The process is defined as follows:

[0049] 5. Output module After completing the preset After each development phase, the final public features will be... Unique features corresponding to each contrast ratio Perform convolutional decoding to output the reconstructed image corresponding to each contrast level. The expression for this is: This invention further introduces Mamba-PMB into U-Mamba-PGM and C-Mamba-PGM to maintain global information interaction capabilities without increasing training memory, thereby balancing reconstruction performance and computational efficiency in high-acceleration 3D multi-contrast reconstruction tasks. The specific steps are as follows:

[0050] Input feature serialization: The input feature map is patchified, which divides it into multiple local patches and further maps them into a token sequence to convert the spatial domain features into a representation suitable for sequence modeling.

[0051] State space modeling: The token sequence is sequentially input into multiple Attentive State-Space Blocks (ASSBs). The long-range dependencies between tokens are modeled through the Attentive State-Space Mechanism (ASSM), thereby realizing the interaction of global context information.

[0052] Feature reconstruction: The token sequence after global feature modeling is remapped back to the spatial domain to restore the corresponding feature map representation.

[0053] Local refinement output: Convolution refinement is performed on the recovered feature map to enhance local texture and edge details, resulting in an updated feature output.

[0054] The complete process of this invention is as follows: Input data: The input to this invention is undersampled magnetic resonance imaging (MRI) data under multiple contrast conditions, specifically including but not limited to T1-weighted (T1), T2-weighted (T2), FLAIR, and other image contrasts. This invention can still achieve joint reconstruction of complete multi-contrast MRI even when the input image is undersampled at high magnification.

[0055] Model structure and processing flow: The overall framework of this invention consists of three core modules: 1. Establish a joint reconstruction optimization model based on Wave-CAIPI encoding and decoupled multi-contrast dictionary priors, which is used to mitigate aliasing artifacts under extreme acceleration by using 3D spatial encoding and to prevent inconsistent feature leakage while utilizing cross-contrast correlation. 2. An alternating minimization algorithm is used to solve the problem and expand it into a multi-stage deep network; the reconstructed image, common features, and unique features are updated alternately through a data consistency module and a Mamba proximal gradient module; finally, reconstructed images with varying contrast are output. 3. A Mamba-based feature update module for efficiently capturing global dependencies without increasing memory overhead; Output: Through the above processing steps, the final output of this invention is a complete set of multi-contrast magnetic resonance images. Even with a significantly reduced sampling rate, this invention can still reconstruct multi-contrast images that are consistent with the full sampling case, effectively overcoming the limitations of traditional methods in terms of input integrity and speedup.

[0056] In numerical experiments, the method was validated using a private brain dataset, uMR. This dataset was acquired using a 3T MRI scanner and a 32-channel head coil, and includes full-sample k-space data from 51 subjects, along with additional prospective data and lesion data; each subject's data includes T1, T2, and FLAIR contrast values. The data from the 51 subjects were divided into training, validation, and test sets in a 40:1:10 ratio.

[0057] Table 1 presents the quantitative comparison results between the proposed method and the comparative method under different undersampling conditions. The evaluation metrics include peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and normalized root mean square error (NRMSE). Experimental results show that the proposed method exhibits superior reconstruction performance compared to the comparative method under different undersampling rates.

[0058] Table 1 Comparison of different methods in uMR brain data Specifically, this manifests as follows: Peak Signal-to-Noise Ratio (PSNR): PSNR is a commonly used metric to measure the error between the reconstructed image and the real image; a higher value indicates better reconstruction quality. As shown in Table 1, the method of this invention achieves the best or one of the best PSNR results under different undersampling conditions and contrast ratios. Under the undersampling condition of R=4×4, the PSNR of the method of this invention reaches 34.13 dB, 35.37 dB, and 31.30 dB at T1, T2, and FLAIR contrast ratios, respectively. T1 and T2 show improvements of approximately 1.28 dB and 2.02 dB compared to the strong baseline method JVN-Net, respectively. Under the undersampling condition of R=4×5, the PSNR of the method of this invention still reaches 32.85 dB, 34.54 dB, and 30.78 dB, indicating strong reconstruction robustness even under higher acceleration conditions. Particularly under FLAIR contrast ratios, the method of this invention still achieves a significant advantage under high undersampling conditions, demonstrating better recovery capabilities for complex tissue contrasts and weakly structured regions.

[0059] Structural Similarity (SSIM): SSIM measures the similarity of the reconstructed image and the real image in terms of structural information. A value closer to 1 indicates greater structural similarity. According to Table 1, the method of this invention maintains high SSIM values ​​under both high-speed undersampling conditions. Under R=4×4, the SSIM values ​​for T1, T2, and FLAIR are 0.9214, 0.9432, and 0.8904, respectively; under R=4×5, the SSIM values ​​are 0.9183, 0.9328, and 0.8862, respectively. These results demonstrate that the method of this invention can still maintain the consistency and edge information of brain tissue structure well under high-speed sampling conditions, especially in contrast ratios sensitive to structural details such as T2 and FLAIR, effectively reducing structural blurring and local distortion.

[0060] Normalized Root Mean Square Error (NRMSE): NRMSE measures the overall error level between the reconstructed image and the real image. A smaller NRMSE value indicates a smaller difference between the reconstructed and real images, and higher reconstruction accuracy. According to Table 1, the NRMSE values ​​for T1, T2, and FLAIR contrast ratios of the method in this invention are 0.0883, 0.0849, and 0.1167 under R=4×4 conditions; and 0.1020, 0.0944, and 0.1226 under R=4×5 conditions. It can be seen that even under higher acceleration conditions, the method in this invention can maintain a low error, indicating that the combination of Wave-CAIPI coding consistency and decoupled multi-contrast dictionary priors can effectively suppress the propagation of reconstruction errors and improve overall reconstruction accuracy.

[0061] Acceleration Factor: This invention breaks through the limitations of traditional methods in terms of acceleration factor. Even with a multi-contrast input that is 20 times undersampled, this invention can still recover a complete multi-contrast image, significantly exceeding the 20-fold upper limit of traditional single-contrast methods.

[0062] Detail Preservation: The images reconstructed by this invention can clearly preserve the detailed information of tissue structures, especially the anatomical structures and pathological areas of the brain, and are highly consistent with the full-sampled images.

[0063] Noise suppression: No obvious noise artifacts were found in the reconstructed image, indicating that the method can effectively suppress noise and improve image quality when processing undersampled data.

[0064] Combining the quantitative analysis in Table 1 and Figure 5Qualitative analysis clearly demonstrates that the method proposed in this invention has significant advantages in multi-contrast MRI joint reconstruction tasks. This method not only outperforms existing algorithms in numerical metrics but also excels in detail preservation and noise suppression, providing high-quality reconstructed images for clinical diagnosis. Furthermore, the method exhibits stable performance across different contrast levels, demonstrating broad applicability and practicality.

[0065] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0066] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for joint reconstruction of three-dimensional multi-contrast magnetic resonance imaging, characterized in that, include: Acquire undersampled multi-contrast magnetic resonance images; A joint reconstruction optimization model for multi-contrast magnetic resonance imaging is established; the joint reconstruction optimization model is solved by an alternating minimization method, and the iterative solution process is expanded into an end-to-end multi-stage deep network. The end-to-end multi-stage deep network includes an initialization module, a cascaded expansion module, and an output module. The cascaded expansion module includes multiple cascaded expansion stage processing units, and each expansion stage processing unit includes a data consistency module, a unique feature Mamba proximal gradient module, and a common feature Mamba proximal gradient module. Based on the multi-contrast magnetic resonance images, the initialization module obtains the common anatomical features shared across contrasts and the unique features of each contrast. For each contrast in each unfolding stage, the magnetic resonance image is updated using the data consistency module based on the current stage's magnetic resonance image, common features, and unique features. Based on the updated magnetic resonance image and common features, the unique features corresponding to the current contrast are updated using the Mamba proximal gradient module according to the Mamba state space mechanism. Based on the updated magnetic resonance image and unique features, the common features are updated using the Mamba proximal gradient module according to the Mamba state space mechanism. Based on the final common features updated after multiple unfolding stages and the unique features corresponding to each contrast, the reconstructed image corresponding to each contrast is obtained through the output module.

2. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 1, characterized in that, The specific feature Mamba proximal gradient module and the common feature Mamba proximal gradient module are respectively composed of a gradient descent module and a Mamba proximal mapping module; the Mamba proximal mapping block is used to serialize the input features, model the long-range dependencies in the sequence through the attention state space mechanism, deserialize, and convolutionally refine, and finally output the updated features.

3. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 1, characterized in that, The construction of the joint reconstruction optimization model for multi-contrast magnetic resonance imaging includes: Construct a WaveCAIPI-encoded image domain forward model, which includes a k-space domain model and an image domain equivalent model; Based on the fully sampled 3D image to be reconstructed, undersampled observation data of the k-th contrast is acquired through the k-space domain model; through the image domain equivalent model, the magnetic resonance images of each contrast are mapped by the coded forward operator to obtain the aliased multi-coil image corresponding to the n-th contrast. A decoupled multi-contrast convolution dictionary prior is introduced to decompose each contrast image into common features and unique features of each contrast. The features are then mapped back to the image space by convolving and summing the common features and unique features. The aliased multi-coil image obtained through the image domain equivalence model, the common features obtained through the prior of the decoupled multi-contrast convolution dictionary, the unique features of each contrast, and the approximation terms of the decoupled convolution dictionary are combined to obtain the joint reconstruction optimization model.

4. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 3, characterized in that, The formula for solving the aliased multi-coil image corresponding to the nth contrast is: ； in, This represents the aliasing operator caused by CAIPIRINHA sampling; Indicates the first k Coil sensitivity diagrams corresponding to each contrast ratio; This represents a one-dimensional Fourier transform along the readout direction; This represents a one-dimensional inverse Fourier transform along the readout direction; This represents the point spread function operator introduced by Wave modulation.

5. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 1, characterized in that, The step of updating the magnetic resonance image through the data consistency module includes: For the first contrast in the t-th unfolding stage, the current magnetic resonance image is mapped to the observation domain and compared with the actual aliased observation image; at the same time, the magnetic resonance image is updated by combining the image priors corresponding to common features and unique features.

6. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 1, characterized in that, The process of updating the unique feature corresponding to the current contrast using the Mamba proximal gradient module is represented as follows: ； in, Indicates the unfolding iteration phase; Indicates the first One contrast ratio; and These represent the contrast-specific latent variables and the common latent variables shared by multiple contrast ratios in the previous stage, respectively. Used to characterize the anatomical information shared between different contrasts; This indicates the number of data points obtained after the current stage has undergone data consistency updates. One contrast-reconstructed image; Indicates the use of updating specific variables The gradient descent module.

7. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 1, characterized in that, The process of updating the common features using the Mamba proximal gradient module is represented as follows: in, Indicates the unfolding iteration phase; Indicates the first One contrast ratio; and These represent the contrast-specific latent variables and the common latent variables shared by multiple contrast ratios in the previous stage, respectively. Used to characterize the anatomical information shared between different contrasts; This indicates the number of data points obtained after the current stage has undergone data consistency updates. One contrast-reconstructed image; Indicates the use of updating public variables The gradient descent module.

8. The three-dimensional multi-contrast magnetic resonance imaging joint reconstruction method as described in claim 1, characterized in that, The multiple contrast ratios include T1, T2, and FLAIR.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 8.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 8.

Images