Multi-contrast mr image reconstruction method and system based on dual-domain feature fusion

Abstract

The application discloses a multi-contrast MRI image reconstruction method and system based on double-domain feature fusion, and relates to the technical field of medical image processing and magnetic resonance imaging. The application sets a reference-guided cross-attention module RGCA, under the constraint of target modality features, carries out gate screening and attention enhancement on reference modality features, and then selectively injects the screened effective reference information into the target modality features; the application sets an energy-guided amplitude-phase decoupling double-domain mixed module EGAP-DM, inputs the features into a spatial domain enhancement branch and a frequency domain decoupling branch, models the amplitude spectrum and the phase spectrum in the frequency domain, and guides and adjusts the phase update by using the amplitude energy, so that the spatial domain local detail enhancement and the frequency domain global structure constraint can be simultaneously considered by the technical point, and the frequency information utilization accuracy and the stability of the reconstruction process are improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and magnetic resonance imaging technology, specifically to a method and system for multi-contrast MRI image reconstruction based on dual-domain feature fusion. Background Technology

[0002] Magnetic resonance imaging (MRI) is a non-invasive medical imaging technology widely used in clinical diagnosis. It has advantages such as no ionizing radiation, high soft tissue contrast, and multi-parameter multi-contrast imaging, and can provide rich anatomical and functional information for lesion detection, pathological analysis, and treatment follow-up.

[0003] Existing technologies for multi-contrast MRI reconstruction tasks typically use undersampled target modal images as the reconstruction object, while introducing fully sampled or higher-quality reference modal images as auxiliary inputs to utilize complementary structural information between different modalities to improve the reconstruction quality of the target modality.

[0004] In existing technologies, related methods typically construct two parallel pathways: a spatial domain branch and a frequency domain branch. The spatial domain branch is used to extract local texture, edge, and structural detail features of the image; the frequency domain branch uses Fourier transform to model the features to obtain global frequency information with a larger receptive field, thereby achieving dual-domain representation of the target modality image.

[0005] Building upon bi-branch feature extraction, existing technologies further incorporate a cross-modal fusion module to introduce useful information from the reference modality into the target modality feature reconstruction process. These methods typically perform selective or guided fusion of reference and target modality features in the spatial or frequency domains, respectively, to enhance the target modality's ability to utilize complementary information.

[0006] After completing spatial and frequency domain feature extraction and cross-modal information exchange, existing technologies typically include a spatial-frequency fusion module to further integrate the dual-domain features, forming a more comprehensive target modality representation, and ultimately outputting a reconstructed image of the target modality. The overall approach of this type of scheme is to enhance feature representation capabilities through dual-domain modeling and to achieve higher-quality MRI reconstruction with the assistance of a reference modality.

[0007] While existing technologies have introduced reference modalities to assist in target modality reconstruction and established cross-modal fusion mechanisms, most schemes primarily involve feature interaction within pre-defined branches or predetermined fusion paths. The distinction between which reference information should be retained and which information might constitute interference remains insufficiently precise. Because different modalities share anatomical structures but also exhibit inconsistencies due to differences in imaging mechanisms, the lack of more targeted guidance and screening mechanisms can easily lead to the introduction of modal information irrelevant or even conflicting with the target reconstruction, thereby affecting the final reconstruction quality.

[0008] While existing technologies employ a dual-branch structure in the spatial and frequency domains to extract local detail information and global frequency information respectively, the interaction between the two domains often involves parallel extraction followed by splicing, weighting, or staged fusion, lacking a closer two-way interaction and collaborative enhancement mechanism. This leads to insufficient mining of the complementary information relationship between spatial and frequency domain features, making it difficult to fully leverage the combined advantages of spatial detail modeling and frequency global modeling.

[0009] MRI reconstruction not only requires restoring the overall structure but also preserving as much detail as possible in terms of edges, textures, and lesion-related features. While existing techniques focus on frequency domain modeling, some methods do not adequately distinguish the roles of different frequency bands during frequency information utilization, or they are prone to weakening high-frequency details during fusion. As a result, in complex structural regions, edge transition regions, or fine-grained texture regions, there may still be issues such as unclear reconstruction and insufficient detail restoration.

[0010] For multimodal MRI reconstruction, it is necessary to model long-range dependencies and global structural relationships while controlling model complexity to adapt to practical applications. In existing technologies, convolution-based methods have limited global modeling capabilities, while Transformer-based or complex cross-modal attention methods, although enhancing global correlation modeling, are typically accompanied by high computational overhead. Even with the introduction of Mamba or frequency domain modeling to improve efficiency, a more unified and efficient solution has not yet been formed for cross-modal selective guidance, dual-domain joint collaboration, and detail recovery; therefore, there is still room for further optimization.

[0011] In summary:

[0012] Existing solutions have made some progress in cross-modal fusion, frequency domain modeling, spatial domain recovery, or state space modeling, but they often focus on a single direction and have not yet formed a unified framework that can simultaneously take into account effective selection of reference modes, deep collaboration between the spatial and frequency domains, enhancement of key features, and suppression of redundant information.

[0013] Therefore, existing technologies still have room for improvement in terms of structural consistency, detail fidelity, and model efficiency in target modality reconstruction. Summary of the Invention

[0014] The purpose of this invention is to provide a method and system for multi-contrast MRI image reconstruction based on dual-domain feature fusion, so as to solve the technical problems mentioned in the background art.

[0015] To achieve the above objectives, the present invention provides the following technical solution: a multi-contrast MRI image reconstruction method and system based on dual-domain feature fusion, comprising the following steps:

[0016] S1: Shallow feature extraction and initial alignment. Shallow convolutional feature extraction is performed on the fully sampled reference modality image and the undersampled target modality image respectively. Through the bidirectional cross-attention interaction fusion block, i.e., the BICA module, the initial dual-domain interaction and feature pre-alignment of the spatial domain features and frequency domain features in their respective branches are completed to obtain the initial features of the reference modality and the initial features of the target modality.

[0017] S2: Multi-scale dual-domain collaborative coding performs multi-level coding processing on the initial features of the reference modality and the initial features of the target modality. Each level of coding processing includes downsampling operation, dual-domain modeling of EGAP-DM module, and bidirectional cross-domain interaction enhancement of BICA module, outputting high-level features of the reference modality and high-level features of the target modality.

[0018] S3: Target-led cross-modal information enhancement. The reference modality high-level features and the target modality high-level features are input into the RGCA module. The target modality high-level features are the main focus. The reference modality high-level features are gating and attention-enhanced. The effective reference information after screening is selectively injected into the target modality high-level features to obtain cross-modal enhanced target modality features.

[0019] S4: Latent space feature refinement. Input the target modal features of cross-modal enhancement into the LGBB block, and perform gating, compression and refinement of high-level features in the latent space to generate compact and discriminative latent representation features.

[0020] S5: Step-by-step decoding and recovery. Multi-level decoding processing is performed on the latent representation features. Each level of decoding processing includes upsampling operation, skip connection fusion with the corresponding coding layer features, dual-domain modeling of EGAP-DM module, and bidirectional cross-domain interaction enhancement of BICA module, gradually restoring the spatial resolution and structural details of the features.

[0021] S6: Reconstructed image output. The features recovered from the decoding end are mapped through a convolutional layer to generate the final reconstructed MRI image of the target modality.

[0022] Furthermore, the processing steps of the bidirectional cross-attention interaction fusion block BICA specifically include: processing the input spatial domain features... Frequency domain characteristics First, pre-alignment is performed using a feature mapping function to obtain the pre-aligned spatial features. With pre-aligned frequency characteristics The calculation formula is:

[0023]

[0024] in, and for The feature mapping function is constructed using convolutions; bidirectional cross-attention interaction is performed, using pre-aligned spatial features as keys and pre-aligned frequency features as queries, to complete the enhancement and update of frequency domain features, resulting in enhanced frequency features. Then, using the enhanced frequency feature as the key and the pre-aligned spatial feature as the query, the spatial domain feature enhancement and update are completed, resulting in the enhanced spatial feature. The calculation formula is:

[0025]

[0026]

[0027] in, For cross-attention operations, Q is the query matrix, K is the key matrix, and V is the value matrix;

[0028] Enhanced frequency characteristics Enhanced spatial features Perform channel concatenation and convolution fusion operations to obtain cross-domain joint features.

[0029] Furthermore, the energy-guided amplitude-phase decoupling dual-domain hybrid module EGAP-DM includes a spatial domain enhancement branch and a frequency domain decoupling enhancement branch, and the specific processing steps are as follows:

[0030] Spatial domain enhancement branch: Multi-scale local responses are extracted from the input features through depthwise convolution with 3×3 and 5×5 kernels. Adaptive fusion of multi-scale features is achieved through learnable gating weights, followed by channel recalibration via a channel attention mechanism to obtain the spatial domain enhanced features. The calculation formula is as follows:

[0031]

[0032]

[0033]

[0034]

[0035] in, and These represent depthwise convolutions with different kernel sizes. Represents a nonlinear mapping. For learnable parameters, This represents the Sigmoid function. For channel attention module, This represents a 1×1 convolutional layer;

[0036] Frequency domain decoupling enhancement branch: The input features are subjected to a two-dimensional Fourier transform to obtain a frequency representation, which is then decoupled into an amplitude spectrum and a phase spectrum. Global energy distribution enhancement is performed on the amplitude spectrum to obtain an enhanced amplitude spectrum. Guide weights are generated based on the energy distribution of the amplitude spectrum, and the update amplitude of the phase spectrum is adaptively adjusted through the guide weights to obtain an optimized phase spectrum. The enhanced amplitude spectrum and the optimized phase spectrum are combined and mapped back to the spatial domain by a two-dimensional inverse Fourier transform to obtain the frequency domain enhanced features.

[0037] Residual fusion is performed on the spatial domain enhancement features and the frequency domain enhancement features to output the dual-domain joint enhancement features.

[0038] Furthermore, in the frequency domain decoupling enhancement branch, the formula for generating the guiding weight based on the amplitude spectrum energy distribution is as follows:

[0039]

[0040] in, To guide weights, For amplitude spectrum, It is a numerically stable term;

[0041] The phase spectrum optimization formula is as follows:

[0042]

[0043] in, To optimize the phase spectrum, The original phase spectrum, To update the strength coefficient, For phase update amount, This is an element-wise multiplication operation.

[0044] Furthermore, the RGCA module is a reference-guided cross-attention module;

[0045] The specific processing steps of the RGCA module are as follows:

[0046] For the input target modal features With reference modal features Channel splicing is performed, and a cross-modal injection gating graph is generated using a mapping function and a sigmoid activation function. ;

[0047]

[0048] in, Indicates channel splicing. This represents a mapping function.

[0049] Apply attention filtering to the reference modality features:

[0050]

[0051] in, This indicates the reference feature enhancement operator.

[0052] Finally, the reference information is injected into the target modal features in a controlled manner:

[0053]

[0054] Obtain the target modal features for cross-modal enhancement .

[0055] Furthermore, the processing steps of the latent space-gated bottleneck block are as follows: For the input bottleneck feature X, lightweight compression and transformation are performed through two 3×3 convolutional layers in the backbone transformation branch to generate a compact intermediate feature representation Fm, calculated using the following formula:

[0056]

[0057] in, This represents the bottleneck transformation function.

[0058] Gated branches generate a gated graph and modulate the backbone features:

[0059]

[0060]

[0061] in, This represents the gating mapping function.

[0062] Finally, the bottleneck output is obtained through output refinement and residual connection:

[0063]

[0064] in, To output a refined function.

[0065] Furthermore, the number of levels of multi-level encoding processing in S2 is 3-5. After each level of encoding processing, the spatial size of the feature map is compressed to 1 / 2 of the input, and the number of channels is expanded to twice the input.

[0066] In S5, the number of levels of multi-level decoding processing is the same as the number of levels of encoding processing. After each level of decoding processing, the spatial size of the feature map is restored to twice the input, and the number of channels is compressed to half the input.

[0067] Furthermore, the shallow convolution feature extraction is implemented using a 3×3 convolutional layer, and the final convolution mapping in S6 is implemented using a 3×3 convolutional layer. The number of output channels of the convolutional layer is consistent with the number of channels of the target modality MRI image.

[0068] Furthermore, the reference modal image and the target modal image are MRI images of the same imaging object with different contrasts, including any two of T1-weighted, T2-weighted, PD-weighted, FLAIR, and DWI.

[0069] The system for multi-contrast MRI image reconstruction based on dual-domain feature fusion includes an input module, a shallow feature processing module, a dual-domain co-coding module, a cross-modal guided enhancement module, a latent space bottleneck refinement module, a step-by-step decoding and recovery module, and an output module.

[0070] The input module is used to acquire fully sampled reference modal MRI images and undersampled target modal MRI images;

[0071] The shallow feature processing module, connected to the input module, is used to perform shallow convolutional feature extraction on the two input images, and to perform initial dual-domain interaction and pre-alignment on spatial domain features and frequency domain features through a bidirectional cross-attention interaction fusion block.

[0072] The dual-domain collaborative coding module, connected to the shallow feature processing module, is used to perform multi-level downsampling, dual-domain modeling and cross-domain interaction enhancement on the two features, and output two high-level features.

[0073] The cross-modal guidance enhancement module is connected to the dual-domain co-coding module and has a built-in reference-guided cross-attention module, which is used to filter and inject effective information of the reference modality in a controlled manner, with the target modality features as the main focus.

[0074] The latent space bottleneck refining module is connected to the cross-modal guidance enhancement module and has a built-in latent space gating bottleneck block for gating, filtering, compressing and refining high-level features.

[0075] The step-by-step decoding and recovery module, connected to the latent space bottleneck refining module and the dual-domain collaborative coding module, is used to perform multi-level upsampling, skip connection fusion, dual-domain modeling and cross-domain interaction enhancement on the refined features to gradually restore feature details.

[0076] The output module, connected to the step-by-step decoding and recovery module, is used to map the recovered features into the target modality reconstructed MRI image and output it.

[0077] Compared with the prior art, the beneficial effects of the present invention are:

[0078] 1. This invention employs a reference-guided cross-attention module (RGCA) to perform gating and attention enhancement on reference modal features under target modal feature constraints. The selected effective reference information is then selectively injected into the target modal features. This technique enables structural priors in the reference modality to participate in target modality recovery in a controlled manner, reducing the risk of introducing irrelevant or conflicting information. It is one of the core technical means by which this invention achieves accurate cross-modal guidance.

[0079] 2. This invention employs an energy-guided amplitude-phase decoupling dual-domain hybrid module, EGAP-DM, which feeds input features into both a spatial domain enhancement branch and a frequency domain decoupling branch. In the frequency domain, it distinguishes and models the amplitude and phase spectra, and uses amplitude energy to guide and adjust phase updates. This technique simultaneously addresses both local detail enhancement in the spatial domain and global structural constraints in the frequency domain, improving the accuracy of frequency information utilization and the stability of the reconstruction process.

[0080] 3. This invention establishes a bidirectional information exchange relationship between spatial domain features and frequency domain features through the bidirectional cross-attention interaction fusion block BICA, enabling effective information from one domain to supplement the other. With this technique, detailed information such as local textures and edges can form a closer synergy with global structure and consistency information, helping to improve the joint representation capability of the two domains.

[0081] 4. This invention sets up a latent space-gated bottleneck block (LGBB) at the network bottleneck stage. Through a combination of backbone compression, gating modulation, and output refinement, it filters, enhances, and compactly represents high-level latent space features. This technique can highlight potential responses that contribute more to reconstruction and suppress the propagation of redundant features, thereby providing a more stable and discriminative high-level representation for subsequent decoding and recovery.

[0082] 5. This invention does not involve isolated improvements to a single module, but rather integrates the reference modality guidance mechanism, the spatial and frequency domain collaborative modeling mechanism, and the bottleneck stage refinement mechanism into a unified reconstruction framework, forming a complete processing flow from input, encoding, cross-modal enhancement, bottleneck refinement to decoding and recovery. This overall flow ensures functional continuity between different technical means, enabling this invention to achieve comprehensive improvements in structural consistency, detail recovery capability, and reconstruction stability. Attached Figure Description

[0083] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0084] Figure 1This is a diagram illustrating the overall network framework of the reconstruction method described in this invention.

[0085] Figure 2 This is a schematic diagram of the structure of the Bidirectional Cross-Attention Interaction Fusion Block (BICA) described in this invention;

[0086] Figure 3 This is a schematic diagram of the structure of the energy-guided amplitude-phase decoupling dual-domain hybrid module EGAP-DM described in this invention;

[0087] Figure 4 This is a schematic diagram of the Reference Guided Cross Attention Module (RGCA) of the present invention;

[0088] Figure 5 This is a schematic diagram of the latent space gated bottleneck block LGBB described in this invention. Detailed Implementation

[0089] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0090] Example 1:

[0091] This embodiment proposes a system for multi-contrast MRI image reconstruction based on dual-domain feature fusion, including an input module, a shallow feature processing module, a dual-domain co-coding module, a cross-modal guided enhancement module, a latent space bottleneck refinement module, a step-by-step decoding and recovery module, and an output module.

[0092] The reference modality input branch receives the fully sampled reference modality image Ref(T1), and the target modality input branch receives the undersampled target modality image Tar(T2). Both inputs first pass through a Conv3×3 convolutional layer to extract shallow local texture and basic structural responses. After the convolutional layers, the two feature paths enter the BICA module to achieve initial interactive alignment between spatial and frequency domain features, obtaining feature representations more suitable for subsequent dual-domain modeling.

[0093] During the encoding phase, the features of the reference mode branch and the target mode branch are sequentially processed by the downsampling unit to compress the spatial size and expand the receptive range. The downsampled features are then fed into the EGAP-DM module to jointly model the local details in the spatial domain and the global prior in the frequency domain. Subsequently, the BICA module further enhances the interaction between the two domains. The above encoding process can be repeated multiple times according to the network depth to form multiple coding blocks Block1(N1).

[0094] At the end of the encoding process, the high-level features output from the reference modality branch and the high-level features output from the target modality branch are jointly input into the RGCA module. The RGCA module, with the target modality features as the primary driver, performs gating and attention enhancement on the reference modality features, and selectively injects the filtered effective reference information into the target modality features to enhance the high-level structural representation of the target modality across modalities.

[0095] The target modality enhancement features processed by the RGCA module are further fed into the LGBB module. The LGBB module is set up at the network bottleneck stage to compress, filter, and refine the high-level semantic features in the latent space, highlighting key structural responses that are more relevant to the reconstruction task and suppressing the propagation of redundant or irrelevant features.

[0096] The refined features from the bottleneck stage are fed into the decoding module. The decoding module includes an upsampling unit, an EGAP-DM module, a BICA module, and a SkipConnection. The upsampling unit is used to progressively restore the spatial resolution of the feature map; the skip connection is used to pass features from the corresponding layer in the encoding stage to the decoding stage to supplement shallow structure and local texture information; the EGAP-DM and BICA modules in the decoding stage continue to collaboratively model and interactively enhance the dual-domain features during the restoration process. Finally, the restored target modality features are mapped to the output image through a Conv3×3 convolutional layer, yielding the target modality reconstruction result Recon(T2).

[0097] exist Figure 1 In the diagram, blue arrows represent the reference modal feature flow, orange arrows represent the target modal feature flow, and black arrows represent skip connection paths. Through the above structural connection method, this invention forms an overall reconstruction framework that includes reference modal guidance, dual-domain collaborative modeling, bottleneck refinement, and step-by-step recovery.

[0098] Example 2:

[0099] This embodiment proposes a multi-contrast MRI image reconstruction method based on dual-domain feature fusion, based on the above embodiments.

[0100] S1: Obtain the fully sampled reference modal image Ref(T1) and the undersampled target modal image Tar(T2), and input them into the reference modal branch and the target modal branch respectively.

[0101] S2: Perform shallow convolutional feature extraction on the reference modality image and the target modality image respectively. Through the BICA module, perform initial dual-domain interaction and pre-alignment on the spatial domain features and frequency domain features to form input features suitable for subsequent coding modeling.

[0102] S3: In the encoding stage, downsampling processing is performed on the two features respectively, and the EGAP-DM module is used in each level of encoding block to jointly enhance the local details information in the spatial domain and the global structure information in the frequency domain; then the BICA module is used to establish bidirectional information interaction between the spatial domain and the frequency domain to output a more complete dual-domain joint representation.

[0103] S4: In the high-level encoding stage, the reference modal features and the target modal features are input into the RGCA module. The target modality guides the selection and injection of reference modal information, and the structural priors that are more relevant to the target reconstruction are enhanced into the target modal features.

[0104] S5: The target modal features after cross-modal enhancement are fed into the LGBB module, where key responses are gated, compressed, and refined in the latent space to generate a more compact and discriminative high-level latent representation.

[0105] S6: During the decoding stage, the bottleneck output features are upsampled stepwise and fused with the skip connection features from the encoding stage. During the decoding and recovery process, the EGAP-DM and BICA modules are used to perform dual-domain modeling and cross-domain interaction to gradually restore the spatial resolution and structural details of the target modality image.

[0106] S7: At the end of the decoding process, a target modality reconstruction image Recon(T2) is generated through convolution mapping. Through the above steps, the present invention can simultaneously utilize the structural priors provided by the reference modality, the global information provided by the frequency domain, and the detailed information provided by the spatial domain during the reconstruction process, thereby obtaining high-quality target modality reconstruction results.

[0107] Based on the above, the following design is proposed:

[0108] Bidirectional Cross-Attention Interaction Fusion Block (BICA)

[0109] See Figure 2 The bidirectional cross-attention interaction fusion block (BICA) is set at the beginning of the encoding stage, inside the encoding block, and in the decoding and recovery stage to establish a bidirectional information exchange relationship between spatial domain features and frequency domain features.

[0110] Let the spatial domain characteristics be Frequency domain characteristics are First, pre-aligned features are obtained through lightweight mapping:

[0111]

[0112] in, This represents the feature mapping function.

[0113] Subsequently, a bidirectional cross-attention interaction is performed:

[0114]

[0115]

[0116] in, This indicates cross-attention operations.

[0117] Finally, the fusion yields cross-domain joint features:

[0118]

[0119] In existing technologies, the processing of spatial and frequency domain features typically focuses on parallel extraction followed by concatenation, summation, or unified mapping, making it difficult to explicitly model the correspondence between the two types of features. Therefore, the complementary utilization of local detail information and global structural information remains insufficient. This invention, through a BICA module, uses one domain feature as query information and the other domain feature as key-value information to perform bidirectional cross-attention interaction. In this way, the global context in the frequency domain can supplement the local recovery process in the spatial domain, and the fine-grained texture and edge information in the spatial domain can also inversely constrain the update of the frequency domain features.

[0120] Furthermore, the BICA module can perform lightweight mapping or pre-alignment on the two types of features before bidirectional interaction to reduce the difference in feature distribution; after completing the bidirectional attention interaction, a joint representation is obtained through the fusion operator. This design makes the spatial domain and frequency domain no longer just a simple parallel relationship, but can continuously form synergistic enhancement during the reconstruction process, thus improving the problems of insufficient bi-domain synergy and weak correlation between global structure and local details in existing technologies.

[0121] Energy-guided amplitude-phase decoupled dual-domain hybrid module EGAP-DM

[0122] See Figure 3 The energy-guided amplitude-phase decoupled dual-domain hybrid module EGAP-DM is positioned at the dual-domain modeling location in both the encoding and decoding stages to simultaneously enhance local structural details and global spectral consistency.

[0123] Spatial domain section:

[0124] Given input features Spatial domain branching extracts local responses under different receptive fields:

[0125]

[0126] in, and These represent depthwise convolutions with different kernel sizes. This represents a nonlinear mapping.

[0127] Then, multi-scale features are fused using a gating method:

[0128]

[0129] in, For learnable parameters, This represents the Sigmoid function.

[0130] Further channel recalibration yields the spatial domain augmented output:

[0131]

[0132] Frequency domain section:

[0133] Input features The frequency representation is obtained after Fourier transform:

[0134]

[0135] in, and These represent the amplitude spectrum and the phase spectrum, respectively.

[0136] Amplitude spectrum enhancement:

[0137]

[0138] in, This represents the amplitude enhancement function.

[0139] For the phase spectrum, a guiding weight generated from the amplitude energy is introduced:

[0140]

[0141] in, This represents the phase update function. As a stable term, To update the strength coefficient, This indicates element-wise multiplication.

[0142] Finally, mapping back to the spatial domain:

[0143]

[0144] In the spatial domain, EGAP-DM extracts local responses under different receptive fields through multi-scale convolution and uses a gated fusion mechanism to adaptively adjust features at different scales, enabling the network to dynamically select local structural information that is more conducive to the current reconstruction task based on the input content. After completing multi-scale local fusion, the response to edges, textures and high-discrimination regions can be further enhanced through a channel recalibration mechanism, thereby improving the ability to model local details.

[0145] In the frequency domain, EGAP-DM maps the input features to the frequency domain and decouples the frequency representation into amplitude and phase spectra. Since the amplitude and phase spectra carry different types of information, this invention employs differentiated enhancement strategies for both. For the amplitude spectrum, its global energy distribution response is enhanced to improve the overall structural expressiveness; for the phase spectrum, a gating weight derived from the amplitude energy is used to adaptively adjust the phase update amplitude, thereby highlighting the phase information relevant to key structures while suppressing unstable phase perturbations that may arise from low-energy regions. Subsequently, the enhanced amplitude and updated phase spectra are remapped back to the spatial domain to form a frequency domain enhanced output with global spectral prior constraints.

[0146] In existing technologies, the processing of frequency domain features often lacks fine-grained modeling that distinguishes between amplitude and phase information, or it only performs a uniform transformation on the entire frequency domain, resulting in insufficient utilization of frequency information, especially in terms of high-frequency detail recovery and global structural stability. This invention, through decoupled modeling of the amplitude and phase spectra and an energy-guided phase update mechanism, makes the utilization of frequency domain information more targeted. Finally, the spatial domain enhancement results are fused with the frequency domain enhancement results to obtain a dual-domain output that combines local detail fidelity with global structural consistency. This improves upon the problems of coarse frequency domain modeling and fragmented use of local and global features in existing technologies.

[0147] Reference-guided cross-attention module RGCA

[0148] See Figure 4 The reference-guided cross-attention module RGCA is positioned between the reference modality branch and the target modality branch, and is used to introduce effective structural information from the reference modality into the target modality features in a controlled manner.

[0149] Let the target modal features be The reference modal features are Then, first generate a cross-modal injection gating graph:

[0150]

[0151] in, Indicates channel splicing. This represents a mapping function.

[0152] Apply attention filtering to the reference modality features:

[0153]

[0154] in, This indicates the reference feature enhancement operator.

[0155] Finally, the reference information is injected into the target modal features in a controlled manner:

[0156]

[0157] In multi-contrast MRI reconstruction, the reference modality typically contains more complete structural priors, while the target modality is more susceptible to undersampling artifacts and loss of detail. Current techniques often fuse reference modality information using direct stitching, simple addition, or indiscriminate cross-modal attention. This approach can easily introduce useful information while simultaneously injecting redundant information unrelated to the target reconstruction, or even cross-modal discrepancies, thus affecting reconstruction stability.

[0158] To address the aforementioned issues, this invention utilizes the RGCA module, centered on the target modal features. First, it generates a cross-modal injection gating map based on the joint representation of the target and reference modal features to determine the injection location and intensity of the reference information. Subsequently, it applies joint attention filtering to the reference modal features, highlighting structural regions more relevant to target reconstruction and suppressing responses in inconsistent or redundant regions. Finally, it injects the gated and attention-filtered reference enhancement information into the target modal features using a residual approach, resulting in cross-modal enhancement output.

[0159] Through the above design, the target mode remains dominant throughout the fusion process, while the reference mode only provides controlled supplementation. This improves the accuracy of complementary information utilization, reduces the interference risk caused by indiscriminate fusion in existing technologies, and enhances the structural recovery capability and result stability in target mode reconstruction.

[0160] Latent space gated bottleneck block LGBB

[0161] The latent space gated bottleneck block (LGBB) is set in the network bottleneck stage and is used to compress, filter, and refine high-level latent space features.

[0162] Given the bottleneck stage input features The main branches first generate the middle part:

[0163]

[0164] in, This represents the bottleneck transformation function.

[0165] Gated branches generate a gated graph and modulate the backbone features:

[0166]

[0167]

[0168] in, This represents the gating mapping function.

[0169] Finally, the bottleneck output is obtained through output refinement and residual connection:

[0170]

[0171] Features in the bottleneck stage typically have low spatial resolution but contain richer high-level semantic and global structural information, making this stage ideal for feature compression and important response filtering. Current techniques often employ conventional convolution stacking or simple residual mapping for bottleneck stage features, lacking an adaptive filtering mechanism for key latent space features. This results in redundant responses potentially propagating in subsequent decoding, impacting recovery quality and computational efficiency.

[0172] The LGBB module of this invention includes a backbone transformation branch, a gated modulation branch, and an output refinement branch. The backbone transformation branch performs lightweight compression and transformation on the input features to generate a compact intermediate representation; the gated modulation branch generates a gating graph conditioned on the input features and modulates the backbone features element-wise to enhance the latent space response more relevant to the reconstruction task and suppress irrelevant or redundant components; the output refinement branch performs channel reshaping and representation refinement on the gated features; finally, the original high-level semantic information in the input is preserved through residual connections, and the enhanced latent space representation is output.

[0173] Through this design, the present invention can improve the compactness, stability and discriminativeness of latent space feature representation with lower computational cost, providing a higher quality high-level representation for subsequent decoding and recovery, thereby improving the problems of low efficiency in bottleneck feature utilization and excessive redundant propagation in the prior art.

[0174] In summary:

[0175] Compared with existing technologies, this invention introduces reference guidance, dual-domain collaborative modeling, and latent space gating refinement mechanisms into the multi-contrast MRI reconstruction framework, which enables more full utilization of reference modal priors, spatial domain local detail information, and frequency domain global structural information during the target modality reconstruction process, thus achieving better reconstruction results.

[0176] Specifically, in terms of cross-modal information utilization, existing technologies commonly employ direct splicing, simple addition, or unified attention fusion methods, which often fail to distinguish which information in the reference modality truly contributes to the recovery of the target modality and which may be cross-modal differences or redundant components. Based on this, this invention establishes a reference-guided cross-attention module (RGCA), which generates gating constraints from target modality features and filters and enhances the reference modality features before injecting them into the target modality. Since the injection location and intensity of the reference information are constrained by the target modality, it can preferentially retain structural priors relevant to target reconstruction while suppressing irrelevant or conflicting information from entering the reconstruction process. Therefore, this invention can improve the targeting of cross-modal complementary information utilization, thereby enhancing the structural recovery capability and reconstruction stability of the target modality.

[0177] In dual-domain feature modeling, while existing technologies offer parallel modeling schemes for the spatial and frequency domains, many methods still lack in-depth exploration of the relationship between the two domains, making it difficult to achieve effective synergy between local detail modeling and global structural constraints. This invention addresses this by establishing a bidirectional cross-attention interaction fusion block (BICA), enabling bidirectional information exchange between spatial and frequency domain features during both encoding and decoding. Since spatial domain features are better suited for representing edges, textures, and local structures, while frequency domain features are better suited for representing global distribution and overall structural consistency, bidirectional interaction allows effective information from one domain to supplement the other, thereby improving the completeness of the joint feature representation. Based on this mechanism, this invention can simultaneously address both local detail recovery and overall structural consistency, improving the visual quality and structural accuracy of the reconstruction results.

[0178] Furthermore, in terms of frequency domain modeling, this invention utilizes the energy-guided amplitude-phase decoupled dual-domain hybrid module EGAP-DM to perform more refined differentiation and processing of frequency information. Since the amplitude spectrum and phase spectrum carry different types of information, processing them uniformly can easily lead to coarse utilization of frequency information, even affecting high-frequency detail recovery and spectral stability. This invention first decouples the amplitude spectrum and phase spectrum, then performs enhancement and updating separately. The amplitude spectrum is used to characterize the overall energy distribution, while the phase spectrum undergoes adaptive correction under amplitude energy guidance. This enhances the frequency domain's ability to express global structure and high-frequency response, while avoiding excessive perturbation of phase updates by low-energy regions. Therefore, it can be inferred that this invention is more effective in preserving detailed information of image edges, textures, and structural transition regions during reconstruction, and improves the stability of global spectral constraints.

[0179] Furthermore, in the processing of latent space features in the higher layers of the network, this invention also sets up a latent space gated bottleneck block (LGBB). Although the features at the bottleneck stage have a small spatial size, they contain a lot of high-level semantic and structural information. Without a filtering mechanism, redundant responses will continue to be passed to subsequent decoding processes, affecting the recovery quality and increasing invalid computations. This invention adaptively filters latent space features by combining backbone compression, gated modulation, and output refinement. Since the gated graph can highlight feature responses that are more relevant to the target reconstruction and suppress unimportant latent components, the subsequent decoding stage receives a more compact and discriminative higher-level representation. Based on this, this invention can improve feature transmission efficiency and enhance the stability and effectiveness of the decoding and recovery stage.

[0180] In summary, this invention does not merely improve the multi-contrast MRI reconstruction process at a single level, but rather constructs a complete technical solution from four aspects: cross-modal information guidance, spatial and frequency domain synergy, amplitude-phase decoupling enhancement, and latent space refinement. Because these modules function in concert, effective structural priors from the reference modality can be controlledly introduced into the target modality reconstruction process, complementary information in the spatial and frequency domains can be jointly modeled, and high-level latent representations are further filtered and enhanced. Therefore, compared to existing technologies, this invention achieves higher structural consistency, better local detail recovery capabilities, and more stable reconstruction output.

[0181] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A multi-contrast MRI image reconstruction method based on dual-domain feature fusion, characterized in that: Includes the following steps: S1: Shallow feature extraction and initial alignment. Shallow convolutional feature extraction is performed on the fully sampled reference modality image and the undersampled target modality image respectively. The initial dual-domain interaction and feature pre-alignment of spatial domain features and frequency domain features in their respective branches are completed through the bidirectional cross-attention interaction fusion block, i.e., the BICA module, to obtain the initial features of the reference modality and the initial features of the target modality. S2: Multi-scale dual-domain collaborative coding performs multi-level coding processing on the initial features of the reference modality and the initial features of the target modality. Each level of coding processing includes downsampling operation, dual-domain modeling of EGAP-DM module, and bidirectional cross-domain interaction enhancement of BICA module, outputting high-level features of the reference modality and high-level features of the target modality. S3: Target-led cross-modal information enhancement. The reference modality high-level features and the target modality high-level features are input into the RGCA module. The target modality high-level features are the main focus. The reference modality high-level features are gating and attention-enhanced. The effective reference information after screening is selectively injected into the target modality high-level features to obtain cross-modal enhanced target modality features. S4: Latent space feature refinement. The target modal features for cross-modal enhancement are input into the latent space gated bottleneck block LGBB. The high-level features are gated, filtered, compressed and refined in the latent space to generate compact and discriminative latent representation features. S5: Step-by-step decoding and recovery. Multi-level decoding processing is performed on the latent representation features. Each level of decoding processing includes upsampling operation, skip connection fusion with the corresponding coding layer features, dual-domain modeling of EGAP-DM module, and bidirectional cross-domain interaction enhancement of BICA module, gradually restoring the spatial resolution and structural details of the features. S6: Reconstructed image output. The features recovered from the decoding end are mapped through a convolutional layer to generate the final reconstructed MRI image of the target modality.

2. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 1, characterized in that: The specific processing steps of the BICA module are as follows: spatial domain features of the input With frequency domain characteristics First, pre-alignment is performed using a feature mapping function to obtain the pre-aligned spatial features. With pre-aligned frequency characteristics ; Perform bidirectional cross-attention interaction, using pre-aligned spatial features as keys and pre-aligned frequency features as queries, to complete the enhancement and update of frequency domain features, resulting in enhanced frequency features. ; Then, using the enhanced frequency feature as the key and the pre-aligned spatial feature as the query, the spatial domain feature enhancement and update are completed, resulting in the enhanced spatial feature. ; Enhanced frequency characteristics Enhanced spatial features Perform the fusion operation to obtain cross-domain joint features. .

3. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 1, characterized in that: The EGAP-DM module is an energy-guided amplitude-phase decoupling dual-domain hybrid module, which includes a spatial domain enhancement branch and a frequency domain decoupling enhancement branch. The specific processing steps of the EGAP-DM module are as follows: Spatial domain enhancement branch: Extract multi-scale local responses from input features through deep convolution with different kernel scales, perform adaptive fusion of multi-scale features through learnable gating weights, and then obtain spatial domain enhanced features through channel recalibration; Frequency domain decoupling enhancement branch: The frequency representation is obtained by Fourier transforming the input features, and the frequency representation is decoupled into amplitude spectrum and phase spectrum; Global energy distribution enhancement is performed on the amplitude spectrum to obtain the enhanced amplitude spectrum; guiding weights are generated based on the energy distribution of the amplitude spectrum, and the update amplitude of the phase spectrum is adaptively adjusted through the guiding weights to obtain the optimized phase spectrum; The enhanced amplitude spectrum and the optimized phase spectrum are combined and then mapped back to the spatial domain via inverse Fourier transform to obtain the frequency domain enhancement features. The spatial domain enhancement features and the frequency domain enhancement features are fused to output a dual-domain joint enhancement feature.

4. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 3, characterized in that: In the frequency domain decoupling enhancement branch, the formula for generating the guiding weight based on the amplitude spectrum energy distribution is as follows: in, To guide weights, For amplitude spectrum, It is a numerically stable term; The phase spectrum optimization formula is as follows: in, To optimize the phase spectrum, The original phase spectrum, To update the strength coefficient, For phase update amount, This is an element-wise multiplication operation.

5. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 1, characterized in that: The RGCA module is a reference-guided cross-attention module; The specific processing steps of the RGCA module are as follows: For the input target modal features With reference modal features Channel splicing is performed, and a cross-modal injection gating graph is generated using a mapping function and a sigmoid activation function. ; For reference modal features Perform attention enhancement and filtering to obtain enhanced reference features. ; Enhance reference features using residuals Gated chart Modulated injection of target modal features ; Obtain the target modal features for cross-modal enhancement .

6. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 1, characterized in that: The LGBB block is the latent space gate bottleneck block; The specific processing steps for the LGBB block are as follows: For the bottleneck features of the input, lightweight compression and transformation are performed through the main transformation branch to generate a compact intermediate feature representation; Gated graphs are generated by using the input bottleneck features as conditions through a gated modulation branch. The intermediate feature representations are then modulated element-wise using the gated graphs to enhance the feature responses relevant to the reconstruction task and suppress redundant features. The modulated features are refined by output refinement, and then fused with the input bottleneck features via residual connection to output the refined latent space features.

7. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 1, characterized in that: The S2 multi-level encoding process has 3-5 levels. After each level of encoding process, the spatial size of the feature map is compressed to 1 / 2 of the input, and the number of channels is expanded to twice the input. In S5, the number of levels of multi-level decoding processing is the same as the number of levels of encoding processing. After each level of decoding processing, the spatial size of the feature map is restored to twice the input, and the number of channels is compressed to half the input.

8. The multi-contrast MRI image reconstruction method based on dual-domain feature fusion according to claim 1, characterized in that: The shallow convolutional feature extraction is implemented using a 3×3 convolutional layer, and the final convolutional mapping in S6 is implemented using a 3×3 convolutional layer. The number of output channels of the convolutional layer is consistent with the number of channels of the target modality MRI image.

9. A system for multi-contrast MRI image reconstruction based on dual-domain feature fusion, used in the multi-contrast MRI image reconstruction method based on dual-domain feature fusion as described in any one of claims 1-8, characterized in that: It includes an input module, a shallow feature processing module, a dual-domain co-coding module, a cross-modal guided enhancement module, a latent space bottleneck refinement module, a step-by-step decoding and recovery module, and an output module; The input module is used to acquire fully sampled reference modal MRI images and undersampled target modal MRI images; The shallow feature processing module, connected to the input module, is used to perform shallow convolutional feature extraction on the two input images and to perform initial dual-domain interaction and pre-alignment on the spatial domain features and frequency domain features within each input feature. The dual-domain collaborative coding module, connected to the shallow feature processing module, is used to perform multi-level downsampling, dual-domain modeling and cross-domain interaction enhancement on the two features, and output two high-level features. The cross-modal guidance enhancement module is connected to the dual-domain co-coding module and has a built-in reference-guided cross-attention module, which is used to filter and inject effective information of the reference modality in a controlled manner, with the target modality features as the main focus. The latent space bottleneck refining module is connected to the cross-modal guidance enhancement module and has a built-in latent space gating bottleneck block for gating, filtering, compressing and refining high-level features. The step-by-step decoding and recovery module, connected to the latent space bottleneck refining module and the dual-domain collaborative coding module, is used to perform multi-level upsampling, skip connection fusion, dual-domain modeling and cross-domain interaction enhancement on the refined features to gradually restore feature details. The output module, connected to the step-by-step decoding and recovery module, is used to map the recovered features into the target modality reconstructed MRI image and output it.

Images