Multi-contrast magnetic resonance image reconstruction method and device based on detail contour feature perception

Abstract

The invention discloses a multi-contrast magnetic resonance image reconstruction method and device based on detail contour feature perception. The method comprises the following steps: acquiring a target modal initial image and an auxiliary modal initial image; constructing an iterative network formed by alternately cascading image domain reconstruction units and data consistency layers, wherein each image domain reconstruction unit comprises an encoder and a decoder; in the first iteration, target modal initial images and auxiliary modal initial images are spliced and then input, an encoder extracts shared features firstly, then global contour features and high-frequency detail features are separated in parallel, and potential features are obtained through collaborative fusion; the decoder takes the potential features as input and outputs an image domain preliminary reconstruction result; the data consistency layer transforms the preliminary result into a k space, performs consistency correction on the preliminary result and a target modal sampling point, and then inversely transforms the preliminary result back to an image domain to complete one iteration; and splicing the current output and the auxiliary modal initial image again, inputting the spliced image into a next round of iteration, and repeating the process until a preset number of times to obtain a final target modal magnetic resonance image.

Description

Technical Field

[0001] This application relates to the fields of magnetic resonance image reconstruction and deep learning, and in particular to a multi-contrast magnetic resonance image reconstruction method and related apparatus based on detail contour feature perception. Background Technology

[0002] Multi-contrast magnetic resonance imaging (MRI), with its complementary tissue information provided by different modalities, plays an irreplaceable role in clinical diagnosis for accurately identifying anatomical structures and assessing the nature of lesions, especially in the diagnosis of neurological diseases and tumor screening, providing physicians with more comprehensive imaging evidence. However, traditional multi-contrast MRI requires complete data acquisition for each modality separately, resulting in a significant increase in scan time. This not only reduces equipment efficiency and increases patient costs but also easily leads to motion artifacts due to patients (such as infants, elderly patients, or critically ill patients) being unable to remain still for extended periods, severely impacting image quality and diagnostic accuracy.

[0003] To address these issues, the industry has proposed rapid reconstruction techniques based on undersampled k-space data. These techniques reduce scanning time by decreasing data acquisition and improve reconstruction accuracy through multimodal collaboration. However, existing multi-contrast MRI reconstruction methods still have significant limitations: some convolutional neural network-based schemes are limited by local receptive fields, making it difficult to effectively capture long-range structural associations across modalities, resulting in anatomical contour breaks and geometric distortions in the reconstructed images; while Transformer-based schemes can model global dependencies, they suffer from high computational complexity and insufficient capture of high-frequency detail features, easily leading to the loss of texture details (such as gray-white matter boundary details in brain tissue and vascular texture); a few methods combining state-space models have optimized long-range modeling efficiency, but they have failed to achieve precise separation and collaborative fusion of contour and detail features, often resulting in "coexistence of blurred contours and missing details" under high acceleration factors, failing to meet the clinical demand for high-resolution, high-structural-integrity images. Furthermore, in the process of multimodal data fusion, existing methods often suffer from information redundancy or loss due to insufficient feature alignment between modalities, further affecting the reconstruction effect and making it difficult to balance structural integrity and detail clarity in the image. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-contrast magnetic resonance image reconstruction method and apparatus based on detail contour feature perception. It aims to solve the problems of contour blurring, detail loss, and insufficient alignment of features between modalities in existing multi-contrast magnetic resonance image reconstruction methods, significantly improving the structural consistency and detail fidelity of the reconstructed image.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0007] 1. By utilizing undersampled k-space data of the target modality and auxiliary modality and inverse Fourier transforming them into initial images, the amount of sampling is reduced while introducing structural priors of the auxiliary modality. This helps the subsequent network to more fully recover the details and structure of the target modality, balancing scanning acceleration and reconstruction quality. This solves the problems of traditional multi-contrast MRI, which requires full sampling of each modality, resulting in long scanning time and numerous motion artifacts. At the same time, single-modality reconstruction is difficult to utilize the complementary structural information of multiple modalities.

[0008] 2. The complex image obtained after inverse transformation of the complex k-space data is split into two channels, one for the real part and one for the imaginary part, and input into the network. At the decoding end, it is returned to a two-channel output. This maintains the physical consistency of MR and preserves as much complete information as possible, allowing the network to improve image detail and contrast fidelity without increasing implementation complexity.

[0009] 3. A multi-level iterative network is constructed, consisting of alternating cascaded image domain reconstruction units and data consistency layers. Each reconstruction unit completes one image domain optimization and k-space correction. Through multiple rounds of iterative "network reconstruction + physical consistency correction," the reconstruction results are gradually enhanced in detail and structure while remaining consistent with the real sampled data, significantly reducing artifacts and erroneous structures.

[0010] 4. The first iteration uses the target modality initial image and the auxiliary modality initial image channels to be stitched together. Subsequent iterations consistently use the previous reconstruction result stitched together with the same auxiliary modality initial image. The auxiliary modality initial image provides stable and invariant geometric priors, while the iterative output provides progressively corrected target modality features. Both are utilized simultaneously in each iteration, which is beneficial for stable convergence and improves cross-modal structural consistency. This addresses the problems of insufficient modality alignment or repeated encoding of inputs in existing multimodal methods, leading to information redundancy, feature drift, or training instability.

[0011] 5. Before encoding, a feature transformation module (dual 3×3 convolution + BN + ReLU) is used to project the input onto the channels to increase the number of channels and form intermediate feature maps. This enhances the initial representation capability of the network, provides a richer feature base for the subsequent Transformer and state space modules, and improves the overall reconstruction ceiling.

[0012] 6. In the encoder, a TransformerBlock with layer normalization, multi-head self-attention, feedforward network, and pre-normalized residual structure is used to extract shared features. This achieves cross-modal global dependency modeling and structural alignment with low computational complexity, mitigating contour breaks and geometric distortion, and improving structural consistency between different modalities.

[0013] 7. BaseMambaBlock is used for contour feature extraction. Specifically, the shared feature map is divided into spatial patches and linearly projected into a token sequence. Dynamic visual state space modules are stacked, each integrating an efficient state space module, learnable descriptive convolution, and efficient channel attention. Long-range modeling with linear complexity is achieved through skip sampling, while compensating for local information and suppressing channel redundancy. Finally, the patch inverse embedding is used to restore the two-dimensional contour feature map. This achieves efficient capture of global geometric structure and coherent contours, suppresses long-range artifact propagation, and maintains clear organ boundaries and morphological integrity even under high acceleration factors.

[0014] 8. DetailCNNBlock employs a reversible neural network combined with MobileNetV2 bottleneck residual blocks to separate detail features. Specifically, the shared feature map is divided into two parts along the channel, and addition and multiplication transformations are performed through an affine coupling layer. The mapping function used for the addition / multiplication operations is implemented by the MobileNetV2 bottleneck residual block. Multiple reversible neural network blocks are stacked to finally obtain the detail feature map. This approach emphasizes high-frequency, non-stationary textures while maintaining complete spatial resolution and reversibility, effectively improving texture clarity and the discernibility of small structures, while reducing the risk of over-smoothing of details.

[0015] 9. Collaborative fusion of contour and detail features: Specifically, the contour and detail feature maps are concatenated channel by channel, semantically aligned using TransformerBlock, and the channel dimension is compressed using a feature projection module. The fused feature map is then divided into two equal parts along each channel and summed element-wise to obtain a latent feature map. Aligning contours and details within a unified semantic space and achieving complementary enhancement yields latent features that possess both global structure and local texture, providing high-quality input for decoding and reconstruction.

[0016] 10. The dense state space reconstruction decoder employs a dual-stream structure with dense state space blocks. Each dense state space block contains: a contour stream using BaseMambaBlock and a detail stream using DenseBlock. The output channels of the two streams are concatenated and compressed using 1×1 convolution + BN + ReLU. The contour stream emphasizes global geometric consistency, while the detail stream focuses on texture restoration. The two are closely integrated at each reconstruction level, resulting in a reconstructed image that is both accurate in contour and rich in detail.

[0017] 11. The multi-level outputs of multiple dense state space blocks are uniformly fused by the Global Feature Fusion Unit (GFF); this achieves cross-level and multi-scale global information aggregation, enhances the network's comprehensive representation ability of complex structures and multi-scale details, and further improves image structure consistency and noise suppression capabilities.

[0018] 12. By using two layers of 3×3 convolution + BN + ReLU, the number of channels in the globally fused features is reduced to 2, corresponding to the real and imaginary parts of the image. This maintains a complex image representation consistent with the input format, facilitating subsequent data consistency correction and also benefiting the network's end-to-end learning of physically interpretable mappings.

[0019] 13. In the data consistency layer, the Fourier transform of the preliminary reconstruction result is performed, and the frequency domain result is compared with the original target mode k-space data point by point. Inconsistent positions are replaced with the corresponding position values ​​of the target mode k-space data. This ensures that the reconstruction result is strictly consistent with the actual measurement data at the k-space sampling position, greatly reducing artifacts and erroneous textures, and improving diagnostic reliability.

[0020] 14. The present invention forms an iterative framework consisting of the above encoding, decoding, and data consistency layers. In each round, the structure and details are gradually improved, and errors are continuously corrected under the constraints of real data, so that the reconstructed image maintains high fidelity and low artifacts even with a small number of samples.

[0021] 15. During training, the total loss is calculated by weighting the pixel-level L1 loss, structural similarity loss (1-SSIM), and mean squared error loss. This comprehensively considers pixel accuracy, structural consistency, and overall error, resulting in a better reconstructed image across various performance metrics.

[0022] 16. Before performing the inverse Fourier transform, verify the integrity of the undersampled k-space data. Since the undersampled k-space data is complex domain data, its integrity must be verified before subsequent image domain transformation. If data is missing, it needs to be filled in through interpolation or invalid samples should be removed to ensure the validity of the input data. Ensuring the quality of data entering the network improves the robustness of model training and inference, and mitigates reconstruction failures caused by extreme outlier data.

[0023] 17. Multiple deployment methods for servers / terminals / clouds; reconstruction processing can be achieved independently or collaboratively by terminal devices, servers, or the cloud. Local or cloud-based reconstruction can be flexibly selected based on the actual scenario, achieving rapid reconstruction and seamless integration with clinical workflows.

[0024] 18. The method of this invention can be used in various combinations of single-modality and auxiliary-modality scanning, such as T2 / PD, and is applicable to acceleration factors of 4x, 8x, and higher. It maintains detail fidelity and contour accuracy even under high acceleration conditions, significantly shortens scan time, reduces patient burden, and improves equipment utilization.

[0025] In summary, the method presented in this application fully utilizes the complementary information between multiple modalities and performs iterative optimization under the constraint of k-space data consistency, which significantly improves the detail clarity and contour accuracy of the reconstructed images. It can still obtain high-quality, high-detail, high-contour accuracy and structural consistency magnetic resonance images at low sampling rates, which has important clinical value and application prospects. Attached Figure Description

[0026] Figure 1 This is an application environment diagram of a multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception in one embodiment of this application;

[0027] Figure 2 A flowchart illustrating a multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception provided in an embodiment of this application;

[0028] Figure 3 A schematic diagram of the network structure of a multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception provided in an embodiment of this application;

[0029] Figure 4 A schematic diagram of BaseMambaBlock in a detail contour feature encoder provided in an embodiment of this application;

[0030] Figure 5 A schematic diagram of DetailCNNBlock in a detail contour feature encoder provided in an embodiment of this application;

[0031] Figure 6 A schematic diagram of each dense state space block in a dense state space reconstruction decoder provided in an embodiment of this application;

[0032] Figure 7 This is a schematic diagram of the multi-contrast magnetic resonance image reconstruction result provided in an embodiment of this application;

[0033] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0034] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of various embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1

[0037] The multi-contrast magnetic resonance image reconstruction method provided in this embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send current multi-contrast magnetic resonance image data to server 104. After receiving the current magnetic resonance image data, server 104 inputs the current multi-contrast magnetic resonance image data into a trained multi-contrast magnetic resonance image reconstruction model to obtain a target contrast magnetic resonance reconstructed image. Server 104 can then feed back the obtained target contrast magnetic resonance reconstructed image to terminal 102. Furthermore, in some embodiments, the multi-contrast magnetic resonance image reconstruction method can also be implemented independently by server 104 or terminal 102. For example, terminal 102 can directly perform reconstruction processing on the current multi-contrast magnetic resonance image data, or server 104 can obtain the current multi-contrast magnetic resonance image data from the data storage system and perform reconstruction processing on it.

[0038] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, etc. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0039] Example 2

[0040] like Figures 2-6 As shown, this embodiment provides a multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. This embodiment uses the application of this method to the server in Embodiment 1 as an example for illustration, including the following steps 201 to 205. The specific implementation process of each step is as follows:

[0041] Step 201: Obtain undersampled k-space data from multi-contrast magnetic resonance imaging.

[0042] The acquired multi-contrast magnetic resonance data are undersampled k-space data, which contain two types of core modal data: target modal undersampled k-space data (denoted as...). ) and auxiliary modal undersampled k-space data (denoted as The target mode is the core mode to be reconstructed, and the auxiliary mode is the mode used to provide complementary structural information. Both are derived from magnetic resonance scans of the same anatomical structure.

[0043] Undersampled k-space data is complex domain data, so the data integrity must be verified before subsequent image domain conversion. If there is missing data, it is necessary to fill in the missing data through interpolation or remove invalid samples to ensure the validity of the input data.

[0044] Step 202: Convert the undersampled k-space data into an initial image in the image domain.

[0045] Through the inverse Fourier transform (denoted as...) The target mode k-space data and the auxiliary mode k-space data are mapped to the image domain respectively to obtain the corresponding initial images. The specific formulas are as follows:

[0046]

[0047] in, For the initial image of the target modality, As auxiliary modal initial images, both are complex domain images (with 2 channels, corresponding to the real and imaginary parts respectively), in the following format: R represents the set of real numbers, and H and W represent the height and width, respectively.

[0048] The initial image obtained in this step is a "zero-filled reconstructed image". Because undersampling in the k-space introduces artifacts and blurring, it needs to be further optimized by subsequent iterative reconstruction units.

[0049] Step 203: Construct iterative reconstruction units and perform the first image domain reconstruction.

[0050] In one exemplary embodiment, step 203 specifically includes:

[0051] Step 2031: Construct the overall structure of the iterative reconstruction unit.

[0052] Iterative reconstruction unit (denoted as) ) is the core reconstruction module, each It comprises two main components: an image domain reconstruction unit and a data consistency layer, wherein:

[0053] The image domain reconstruction unit is responsible for optimizing the detail and contour features of the initial image from the image domain. The core sub-modules include the "detail contour feature encoder" and the "dense state space reconstruction decoder".

[0054] The data consistency layer is responsible for ensuring the consistency between the reconstruction results and the original k-space data, preventing image domain optimization from deviating from the real data distribution.

[0055] The number of iterative reconstruction units (denoted as T) is a preset hyperparameter. Through multiple iterations, artifacts are gradually reduced and image quality is improved.

[0056] Step 2032: Determine the input for the first iteration and perform channel splicing.

[0057] The input data for the first iteration is the channel stitching result of "target modality initial image + auxiliary modality initial image", denoted as Specifically, this is achieved through channel splicing operations (denoted as...). )accomplish:

[0058]

[0059] in, The aim is to integrate the initial structural information of multiple modalities to provide a cross-modal foundation for subsequent feature extraction.

[0060] Step 2033: Feature processing of the detail contour feature encoder.

[0061] The core function of the detail contour feature encoder is to: extract details from the concatenated input. The shared feature map is extracted and separated into contour feature map and detail feature map. Finally, the latent feature map is obtained through collaborative fusion (denoted as ). Specifically, it will be implemented in three phases, such as Figure 3 As shown.

[0062] Phase 1: Extracting shared feature maps .

[0063] Shared features are low-level structural information common to multimodal images (such as edges and anatomical shapes). The extraction process involves two steps:

[0064] (1) Projection onto the channel: through two stacked The "feature transformation module" (denoted as ) consists of convolutional layers, batch normalization (BN) layers, and ReLU activation layers. ), for iterative input (First iteration is) Perform channel dimension enhancement (e.g., from 4 channels to C channels, where C is a preset parameter) to obtain an intermediate feature map;

[0065] (2) TransformerBlock processing: TransformerBlock with pre-normalized residual design performs global structure alignment on intermediate feature maps to capture long-range dependencies across modalities.

[0066] The specific formula for extracting shared feature maps is as follows:

[0067]

[0068] in, To share feature maps, the internal structure of TransformerBlock follows pre-normalized residual logic:

[0069]

[0070] In the formula, Representation layer normalization, This indicates a multi-head self-attention mechanism. This represents a multilayer perceptron, where residual connections are used to preserve low-level structural information and prevent feature degradation. X represents a feature map.

[0071] Phase 2, Separating Contour Feature Maps .

[0072] Contour feature maps are extracted using the BaseMambaBlock module, such as... Figure 4 As shown. This module is specifically designed to capture "long-range geometric continuity" and processes shared feature maps in three steps. :

[0073] (1) Patch embedding: A patch embedding module (denoted as Patch Embedding Module) is used. ),Will The space is divided into non-overlapping fixed-size patches, and each patch is projected as a one-dimensional dense token through a linear layer to obtain a token sequence. :

[0074]

[0075] in, (S is the patch size) is the token sequence length, and D is the dimension of each token;

[0076] (2) Dynamic state space aggregation: for First, layer normalization is performed, then multiple "Dynamic Visual State Space Modules (DVSS)" are stacked for processing. Low-level spatial cues are preserved through residual connections to obtain the aggregated token sequence. :

[0077]

[0078] In the formula, s is a learnable scaling parameter. The dynamic state space module integrates "Efficient State Space Module (ESSM)," "Learnable Descriptive Convolution (LDC)," and "Efficient Channel Attention (ECA)," with the following specific structure:

[0079]

[0080] Among them, the efficient state space module achieves long-range dependency capture with linear complexity through skip sampling with a stride of 2, learnable descriptive convolution compensates for local feature loss, and efficient channel attention suppresses channel redundancy.

[0081] (3) Patch inverse embedding: A patch inverse embedding module (denoted as Patch inverse embedding module) is used. ),Will Convert back to a two-dimensional feature map, i.e., a contour feature map :

[0082]

[0083] Encode globally consistent geometric structural information (such as organ outlines and tissue boundaries).

[0084] Phase 3, Separating Detail Feature Maps .

[0085] Detailed features are extracted using the DetailCNNBlock module, such as... Figure 5 As shown, this module is based on a reversible neural network (INN) design, which can capture high-frequency texture information while maintaining full spatial resolution.

[0086] Stacking multiple reversible neural network blocks to process shared feature maps Each reversible neural network block employs an affine coupling layer structure, dividing the input features into two parts along the channel axis (denoted as...). and The invertible transformation is performed on one part as a condition for the other part, and the specific formula is as follows (taking the nth invertible neural network block as an example):

[0087]

[0088]

[0089]

[0090] in, , , For continuous mapping functions, the bottleneck residual block (BRB) of MobileNetV2 is used for implementation. Ensure the multiplication transformation is invertible; the feature maps obtained after processing all invertible neural network blocks are the detail feature maps. It encodes high-frequency, non-stationary texture information (such as the microstructures inside tissues) without information bottleneck loss.

[0091] Phase 4: Collaborative fusion yields potential feature maps.

[0092] The contour feature map is processed through three steps: channel stitching, feature alignment, and segmentation addition. With detailed feature map Fusion into latent feature map First, the channels are spliced ​​together, along the channel axis. and The spliced ​​feature map is obtained. :

[0093]

[0094] Then feature alignment and dimensionality reduction are performed: TransformerBlock processing is used, consistent with "shared feature extraction". Semantic alignment of details and outlines is achieved through MSA; then, through two stacked... The "feature projection module" (denoted as ) consists of convolutional layers, batch normalization (BN) layers, and ReLU layers. The channel dimension is compressed to C to obtain the fused feature map. :

[0095]

[0096] Finally, segmentation and addition are performed to enhance the result. Divide into two sub-feature maps equally along the channel axis and (Each contains a C channel), element-wise addition is used to enhance complementary information, resulting in a latent feature map. :

[0097]

[0098] This provides an optimized basis for the dense state space reconstruction decoder.

[0099] Step 2034: Decoding process of dense state space reconstruction decoder.

[0100] Dense state space reconstruction decoder with latent feature map As input, global contours and local details are aggregated through multi-stream dense state space blocks (DSSB), and the initial reconstruction result P in the image domain is output, which is implemented in three stages:

[0101] Phase 1: Initialization and dense state space block stacking.

[0102] by As the initial input features for the decoder, multiple dense state space blocks (DSSBs) are stacked for processing. Each DSSB employs a two-stream structure to optimize contours and details separately. Contour optimization is performed first. A module consistent with the BaseMambaBlock structure in step 2033 is used to process the input features (the input of the nth DSSB). Output the streaming features that encode the global contour. Then, detailed optimizations are performed, using densely connected blocks for processing. By aggregating fine-grained details through dense interlayer connections, it outputs flowing features that encode local details. Dual-stream fusion and dimensionality reduction: and Channel splicing, through The convolutional layer, BN layer, and ReLU layer compress the channels to C, resulting in the output of the nth dense state space block. :

[0103]

[0104] in, This is the convolution module, where each dense state space block uses independent parameters to adapt to the optimization needs of different iteration stages.

[0105] Phase 2, global feature fusion.

[0106] The outputs of all dense state space blocks are concatenated along the channel axis to obtain the multi-stage feature aggregation result:

[0107]

[0108] Will The input is a Global Feature Fusion Unit (GFF), which reduces the channel dimension and fuses local context through convolution to obtain the global fused features. .

[0109] In stage 3, channel regression yielded preliminary reconstruction results.

[0110] Through two stacks The projection module, composed of convolutional layers, batch normalization (BN) layers, and ReLU layers, globally fuses features. The channel dimension is reduced to 2 (corresponding to the real and imaginary parts of the image), resulting in a preliminary reconstruction of the image domain:

[0111]

[0112] in, This refers to the projection module mentioned above.

[0113] Step 204: Data consistency layer processing and first iteration completed.

[0114] In one exemplary embodiment, step 204 specifically includes:

[0115] Step 2041: Frequency domain comparison and correction.

[0116] Perform a positive Fourier transform on the preliminary reconstruction result P Preliminary results in the frequency domain are obtained; these preliminary results are then compared with the target mode k-space data. Position-by-position comparison: If the preliminary frequency domain result at a certain position is consistent with... If the data is consistent, retain the preliminary frequency domain result value for that position; if the data is inconsistent, use... The value at this position replaces the corresponding value in the initial frequency domain result, ensuring that the frequency domain result matches the actual sampled data.

[0117] Step 2042: Inverse transformation to obtain the current Output.

[0118] Perform inverse Fourier transform on the corrected frequency domain data The output of the current iterative reconstruction unit is obtained, denoted as... That is, the output of the first iteration:

[0119]

[0120] in, This indicates the correction operation at the data consistency layer. This completes one iteration of the reconstruction process.

[0121] Step 205: Perform subsequent iterations and output the final reconstructed image.

[0122] In one exemplary embodiment, step 205 specifically includes:

[0123] Step 2051: Construction of inputs for subsequent iterations.

[0124] Starting from the second iteration (let's denote the iteration number as t), the initial image of the target modality is no longer used. Instead, it combines the output of the previous reconstruction unit with the initial image of the auxiliary modality. Perform channel concatenation as input for the current iteration. The specific formula is as follows:

[0125]

[0126] in, This is the output of the reconstruction unit in the t-th iteration. The core of this design is to use the structural information optimized in the previous iteration as a basis, while retaining the stable complementary information of the auxiliary modes to avoid feature drift during the iteration process.

[0127] Step 2052: Repeat the iterative reconstruction process.

[0128] Will Input the reconstruction unit of the current iteration, and repeat the detailed contour feature encoder processing in step 2033, the dense state space reconstruction decoder processing in step 2034, and the data consistency layer processing in step 204 to obtain the output of the reconstruction unit of the current iteration. The formula can be uniformly expressed as:

[0129]

[0130] in, This represents the feature optimization process of the image domain reconstruction unit, where each iteration corrects artifacts, enhances details, and... Layer-based constraints ensure data consistency, guaranteeing that the reconstruction results gradually approximate the real image.

[0131] Step 2053: Determine if the iteration has terminated and output the final result.

[0132] When the number of iterations reaches the preset total number T, the iteration stops, and the output of the last reconstruction unit is used as the final magnetic resonance reconstruction image of the target mode.

[0133]

[0134] It can be converted into a clinically readable grayscale image through subsequent post-processing (such as amplitude calculation and grayscale normalization) for subsequent diagnostic analysis.

[0135] Additional information: Model training process.

[0136] To ensure the effectiveness of the above reconstruction method, the model needs to be trained in advance. The core of the training process is to optimize the model parameters through a loss function. Specifically, this includes:

[0137] (1) Constructing the training dataset. Collect multi-contrast magnetic resonance imaging (MRI) scan data, including fully sampled k-space data of the target mode and auxiliary mode (as the source of ground truth) and corresponding undersampled k-space data (simulating the actual scanning scenario); perform inverse Fourier transform on the fully sampled k-space data to obtain the ground truth image. The undersampled k-space data is converted into an initial image according to step 202. and The input-true value sample pairs are formed; the sample set is divided into training set, validation set and test set in a ratio of 8:1:1.

[0138] (2) Loss Function Design. The training process employs a multi-component joint loss function, comprehensively constraining pixel-level accuracy, structural similarity, and numerical error. The total loss function... The definition is as follows:

[0139]

[0140] in, , , The specific calculation method for the loss of each component, using weighted coefficients, is as follows: Pixel-level L1 loss The formula for measuring the pixel-level absolute error between the reconstructed image and the ground truth image, avoiding gradient explosion, is as follows:

[0141]

[0142] Structural similarity loss To measure the structural consistency between the reconstructed image and the ground truth image, and to avoid the problem of "pixel-level accuracy but structural distortion," the formula is as follows:

[0143]

[0144] Among them, SSIM is the structural similarity index, and the closer the structural similarity index is to 1, the better the structural consistency.

[0145] Mean square error loss The squared error between the reconstructed image and the ground truth image is used to further constrain numerical accuracy, and the formula is as follows:

[0146]

[0147] (3) Training optimization process and optimizer selection. The Adam optimizer is used, and the initial learning rate is set to A stepped learning rate scheduler is used, which decays the learning rate to 0.5 times the current value every 20 training rounds; 100 training rounds are preset, and the loss is calculated on the validation set after each training round; the model parameters with the lowest validation loss are saved as the final trained model used for inference.

[0148] according to Figure 7 A comparison of the recorded magnetic resonance image reconstruction results shows that implementing the above-mentioned multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception can significantly improve the quality of the reconstructed image while maintaining a low sampling factor. This application fully utilizes the complementary information of multi-contrast magnetic resonance images, using undersampled k-space data of the target and auxiliary modes as input. It designs an image domain reconstruction unit consisting of a detail contour feature encoder and a dense state space reconstruction decoder, which are alternately cascaded with a data consistency layer to form an iterative reconstruction architecture. The detail contour feature encoder achieves the extraction of shared features and accurate separation of contour and detail features. It uses TransformerBlock to capture global structural dependencies, BaseMambaBlock to efficiently model long-range contour relationships, and DetailCNNBlock to maintain the integrity of high-frequency details through a reversible neural network. These features are then collaboratively fused to achieve feature complementarity. Simultaneously, the dense state space reconstruction decoder uses a dual-stream structured dense state space block to focus on global contour reconstruction and local detail recovery, respectively. Finally, global feature fusion yields a structurally consistent and detail-rich reconstruction result.

[0149] The multi-contrast magnetic resonance imaging reconstruction method based on detail contour feature perception presented in this application is highly applicable. It can be used not only for single-modality reconstruction tasks, such as reconstructing T2-weighted images from T2-weighted undersampled images and reconstructing PD images from PD-weighted undersampled images, but also for reconstruction scenarios involving auxiliary modalities, such as reconstructing PD-weighted images using both T2-weighted and PD-weighted undersampled images simultaneously. This method maintains good reconstruction performance even at acceleration factors of 4x, 8x, or even higher, making it particularly suitable for clinical applications requiring high fidelity in image detail and contour accuracy.

[0150] This application also provides an application scenario in which the above-described multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception is applied. Specifically, the multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception provided in this embodiment can be applied to the scenario of reconstructing high-quality multi-contrast magnetic resonance images from undersampled k-space data. This scenario includes a rapid multi-contrast magnetic resonance image reconstruction stage and a reconstruction result display stage. The rapid multi-contrast magnetic resonance image reconstruction stage is used to obtain a high-quality target modality magnetic resonance image through iterative processing of undersampled k-space data of the target modality and auxiliary modality using a reconstruction network that perceives detail contour features. The reconstruction result display stage is used to display the reconstructed target modality magnetic resonance image for medical personnel to refer to for diagnosis. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception provided in this embodiment belongs to the rapid multi-contrast magnetic resonance image reconstruction stage.

[0151] Example 3

[0152] Based on the same inventive concept, this application also provides a multi-contrast magnetic resonance image reconstruction apparatus based on detail contour feature perception, which can be used to implement the method described in Embodiment 2 above, as described in the following embodiments. Since the principle of the multi-contrast magnetic resonance image reconstruction apparatus based on detail contour feature perception is similar to that of the multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception, the implementation of this apparatus can refer to the aforementioned method implementation, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0153] Embodiments of the present invention provide a specific implementation of a reconstruction apparatus capable of realizing a multi-contrast magnetic resonance image reconstruction method based on detailed contour feature perception, specifically including the following:

[0154] An image domain processing unit is used to acquire undersampled k-space data of multi-contrast magnetic resonance imaging, wherein the undersampled k-space data includes target mode k-space data and auxiliary mode k-space data; and to convert the target mode k-space data and auxiliary mode k-space data into target mode initial image and auxiliary mode initial image in the image domain, respectively, through inverse Fourier transform.

[0155] Each reconstruction unit includes an image domain reconstruction unit and a data consistency layer; the image domain reconstruction unit includes a detail contour feature encoder and a dense state space reconstruction decoder.

[0156] The initial iteration unit is used to perform channel concatenation of the initial image of the target modality and the initial image of the auxiliary modality during the first iteration to obtain the first iteration input. The first iteration input is then passed to the image domain reconstruction unit of the current reconstruction unit: the detailed contour feature encoder extracts the shared feature map and separates it into a contour feature map and a detail feature map, which are then collaboratively fused to obtain a latent feature map; the latent feature map is decoded by the dense state space reconstruction decoder to obtain the preliminary reconstruction result of the image domain.

[0157] The data consistency processing unit is used to perform a Fourier transform on the preliminary reconstruction result by the data consistency layer to obtain a preliminary result in the frequency domain, compare the preliminary result in the frequency domain with the target mode k-space data position by position, replace the inconsistent positions with the corresponding position values ​​of the target mode k-space data to obtain corrected frequency domain data, and perform an inverse Fourier transform on the corrected frequency domain data to obtain the output of the current reconstruction unit.

[0158] Subsequent iteration units are used to stitch the output of the previous reconstruction unit with the initial image of the auxiliary modality through channels, and then use the result as the input of the next reconstruction unit to repeat the processing of the reconstruction unit.

[0159] The output unit is used to take the output of the last reconstruction unit as the final magnetic resonance reconstructed image of the target mode after completing a preset number of iterations.

[0160] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores multi-contrast magnetic resonance imaging data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When executed by the processor, the computer program implements a multi-contrast magnetic resonance imaging reconstruction method based on detail contour feature perception.

[0161] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0162] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0163] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0164] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0166] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0167] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0168] 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.

[0169] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception, characterized in that, include: Acquire undersampled k-space data from multi-contrast magnetic resonance imaging, wherein the undersampled k-space data includes target mode k-space data and auxiliary mode k-space data; The target mode k-space data and auxiliary mode k-space data are converted into the target mode initial image and auxiliary mode initial image in the image domain by inverse Fourier transform, respectively. An iterative reconstruction unit is constructed, each of which includes an image domain reconstruction unit and a data consistency layer; the image domain reconstruction unit includes a detail contour feature encoder and a dense state space reconstruction decoder. In the first iteration, the target modality initial image and the auxiliary modality initial image are concatenated by channels to obtain the first iteration input. The first iteration input is then passed to the image domain reconstruction unit of the current reconstruction unit: the detail contour feature encoder extracts the shared feature map and separates it into a contour feature map and a detail feature map, which are then fused together to obtain a latent feature map; the latent feature map is then decoded by the dense state space reconstruction decoder to obtain the preliminary reconstruction result of the image domain. The data consistency layer performs a Fourier transform on the preliminary reconstruction result to obtain a preliminary result in the frequency domain. The preliminary result in the frequency domain is compared with the target mode k-space data position by position. Inconsistent positions are replaced with the corresponding position values ​​of the target mode k-space data to obtain corrected frequency domain data. An inverse Fourier transform is performed on the corrected frequency domain data to obtain the output of the current reconstruction unit. In subsequent iterations, the output of the previous reconstruction unit and the initial image of the auxiliary modality are stitched together via channels and used as the input of the next reconstruction unit, and the processing of the reconstruction unit is repeated. After a preset number of iterations, the output of the last reconstruction unit is used as the final magnetic resonance reconstruction image of the target mode.

2. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception according to claim 1, characterized in that, The process of extracting shared feature maps by the detailed contour feature encoder specifically includes: The channel projection operation is performed on the iterative input: the channel dimension of the iterative input is increased by a feature transformation module consisting of two stacked 3×3 convolutional layers, a batch normalization layer and a ReLU activation layer to obtain an intermediate feature map; Shared feature extraction is performed on the intermediate feature map: the intermediate feature map is processed by a TransformerBlock with pre-normalized residual design. The TransformerBlock includes a layer normalization layer, a multi-head self-attention mechanism and a feedforward network, and retains low-level structural information through residual connections. The feature map obtained after processing is the shared feature map, which encodes globally aligned, modality-independent structural information.

3. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception according to claim 1 or 2, characterized in that, The process of separating the contour feature map by the detail contour feature encoder is implemented through BaseMambaBlock, specifically including: Patch embedding operation: The shared feature map is divided into non-overlapping fixed-size spatial patches using a patch embedding module, and each patch is projected into a dense token through a linear layer to obtain a one-dimensional token sequence; Dynamic state space aggregation: The one-dimensional token sequence is first normalized by layer, and then multiple dynamic visual state space modules are stacked for processing. Each dynamic visual state space module includes an efficient state space module, a learnable descriptive convolution, and an efficient channel attention. The efficient state space module achieves long-range dependency capture with linear complexity through skip sampling with a stride of 2. The learnable descriptive convolution compensates for local feature loss, and the efficient channel attention suppresses channel redundancy. During processing, low-level spatial cues are preserved through residual connections to obtain the aggregated token sequence. Patch inverse embedding operation: The patch inverse embedding module is used to convert the aggregated token sequence into a two-dimensional feature map to obtain a contour feature map, which encodes globally consistent geometric structure information.

4. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception according to claim 1 or 2, characterized in that, The process of separating detail features by the detail contour feature encoder is implemented through DetailCNNBlock, specifically including: Multiple reversible neural network blocks are stacked to process shared feature maps. Each reversible neural network block adopts an affine coupling layer structure. It is used to divide the input features into two parts along the channel axis, and perform a reversible transformation on the other part based on one part. The reversible transformation includes additive transformation and multiplicative transformation, and both use the bottleneck residual block of MobileNetV2 as a continuous mapping function. The feature map obtained after processing by all reversible neural network blocks is the detail feature map. The detail feature map captures high-frequency, non-stationary texture information while maintaining the complete spatial resolution, and there is no information bottleneck loss.

5. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception according to claim 1 or 2, characterized in that, The collaborative fusion process of the detailed contour feature encoder includes: Channel stitching: The contour feature map and the detail feature map are stitched together along the channel axis to obtain a stitched feature map; Feature alignment and dimensionality reduction: The stitched feature map is processed by TransformerBlock with pre-normalized residual design. The semantic alignment of detail features and contour features is achieved through a multi-head self-attention mechanism. Then, the channel dimension of the stitched feature map is compressed to be consistent with the shared feature map through a feature projection module consisting of two stacked 3×3 convolutional layers, a batch normalization layer and a ReLU activation layer, to obtain the fused feature map. Segmentation and addition enhancement: The fused feature map is divided into two sub-feature maps along the channel axis. An element-wise addition operation is performed on the two sub-feature maps to obtain the latent feature map.

6. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception according to claim 1, characterized in that, The decoding process of the dense state space reconstruction decoder includes: Initialization and block stacking: The latent feature map is used as the initial input feature; multiple dense state space blocks are stacked for processing, and each dense state space block adopts a two-stream structure; In each of the dense state space blocks, the first stream uses BaseMambaBlock to process the input features and outputs a stream feature encoding the global contour; the second stream uses DenseBlock to process the input features, aggregates fine-grained details through dense inter-layer connections, and outputs a stream feature encoding local details; the outputs of the above two streams are concatenated along the channel axis, and the channel dimension is compressed through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the output of the current dense state space block; Global feature fusion: The outputs of multiple dense state space blocks and the initial input features are concatenated along the channel axis and input to the global feature fusion unit; the global feature fusion unit first reduces the channel dimension through a 1×1 convolutional layer, and then fuses the local context through a convolutional layer to obtain the global fused features; Channel regression: By using two stacked 3×3 convolutional layers, a batch normalization layer, and a ReLU activation layer, the channel dimension of the global fusion feature is reduced to 2 to correspond to the real and imaginary parts of the image. The resulting feature map is the preliminary reconstruction result of the image domain.

7. The multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception according to claim 1, characterized in that, It also includes the model training process, in which the total loss function used in the model training process includes pixel-level L1 loss, structural similarity loss and mean squared error loss; The pixel-level L1 loss is the L1 norm distance between the final magnetic resonance reconstructed image and the ground truth image; The structural similarity loss is 1 minus the structural similarity index between the final magnetic resonance reconstructed image and the ground truth image; The mean squared error loss is the squared L2 norm of the final magnetic resonance reconstructed image and the ground truth image; The total loss function is the sum of the pixel-level L1 loss, structural similarity loss, and mean squared error loss multiplied by their respective weight coefficients. The model parameters are updated by minimizing the total loss function.

8. A multi-contrast magnetic resonance image reconstruction device based on detail contour feature perception, characterized in that, include: An image domain processing unit is used to acquire undersampled k-space data of multi-contrast magnetic resonance imaging, wherein the undersampled k-space data includes target mode k-space data and auxiliary mode k-space data; The target mode k-space data and auxiliary mode k-space data are converted into the target mode initial image and auxiliary mode initial image in the image domain by inverse Fourier transform, respectively. Each reconstruction unit includes an image domain reconstruction unit and a data consistency layer; the image domain reconstruction unit includes a detail contour feature encoder and a dense state space reconstruction decoder. The initial iteration unit is used to perform channel concatenation of the initial image of the target modality and the initial image of the auxiliary modality during the first iteration to obtain the first iteration input. The first iteration input is then passed to the image domain reconstruction unit of the current reconstruction unit: the detailed contour feature encoder extracts the shared feature map and separates it into a contour feature map and a detail feature map, which are then collaboratively fused to obtain a latent feature map; the latent feature map is decoded by the dense state space reconstruction decoder to obtain the preliminary reconstruction result of the image domain. The data consistency processing unit is used to perform a Fourier transform on the preliminary reconstruction result by the data consistency layer to obtain a preliminary result in the frequency domain, compare the preliminary result in the frequency domain with the target mode k-space data position by position, replace the inconsistent positions with the corresponding position values ​​of the target mode k-space data to obtain corrected frequency domain data, and perform an inverse Fourier transform on the corrected frequency domain data to obtain the output of the current reconstruction unit. Subsequent iteration units are used to stitch the output of the previous reconstruction unit with the initial image of the auxiliary modality through channels, and then use the result as the input of the next reconstruction unit to repeat the processing of the reconstruction unit. The output unit is used to take the output of the last reconstruction unit as the final magnetic resonance reconstructed image of the target mode after completing a preset number of iterations.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the multi-contrast magnetic resonance image reconstruction method based on detail contour feature perception as described in any one of claims 1 to 7.

Images