A medical image reconstruction method and device based on structural constraint and spectrum collaborative modulation

Abstract

The application relates to a medical image reconstruction method and device based on structural constraint and spectrum collaborative modulation. First, a target medical image is acquired, and a reference medical image is acquired when the reference medical image exists; input images are subjected to spatial alignment, intensity normalization, initial upsampling, frequency domain mapping and flow rearrangement processing; subsequently, gradient response, edge response and tissue structure response of the target medical image are extracted through structural gradient coding to generate a structural constraint result; different frequency components are subjected to band decoupling, response analysis and adaptive weighted modulation through spectrum collaborative modulation to generate a spectrum collaborative modulation result; the structural constraint result and the spectrum collaborative modulation result are further fused for space-frequency coupling reconstruction to output a reconstructed medical image; in the training stage, model parameters are updated through local frequency domain consistency calibration. The application improves the structural consistency, boundary definition, high-frequency detail recovery capability and spectrum stability of medical image reconstruction, and improves the reconstruction efficiency.

Description

Technical Field

[0001] This application relates to the field of medical image processing and computer-aided imaging, and in particular to a method and apparatus for medical image reconstruction based on structural constraints and spectral co-modulation. Background Technology

[0002] Medical images are crucial for clinical diagnosis, lesion localization, treatment planning, and postoperative evaluation. Common medical images include magnetic resonance imaging (MRI), computed tomography (CT), positron emission tomography (PET), and ultrasound. In practical applications, factors such as acquisition time, scan dose, imaging equipment, patient movement, and sampling density can easily lead to problems like insufficient resolution, loss of detail, blurred edges, or increased noise in medical images. Therefore, achieving high-quality medical image reconstruction under limited sampling or low-resolution input conditions is a significant challenge in the field of medical image processing.

[0003] Existing medical image reconstruction methods mainly include interpolation reconstruction, compressed sensing reconstruction, and deep learning-based reconstruction methods. Interpolation methods are computationally simple but struggle to recover true high-frequency details; compressed sensing methods rely on sparse priors and iterative optimization processes, making reconstruction quality susceptible to parameter settings and sampling methods; while deep learning methods can learn the nonlinear mapping relationship between low-quality and high-quality images, some methods primarily rely on image spatial feature fusion, failing to adequately utilize image structural constraints, spectral distribution characteristics, and frequency domain consistency, easily leading to problems such as overly smoothed edges, texture artifacts, or local structural discontinuities.

[0004] For multimodal, multi-sequence, or multi-contrast medical image reconstruction tasks, reference medical images typically contain anatomical structural information related to the target medical image. Existing methods often simply use the reference medical image as an additional input channel or auxiliary feature for fusion, lacking explicit constraints on the structural information of the reference image and adaptive modulation mechanisms for different frequency components. This results in limited recovery performance of the target image in tissue boundaries, fine structures, and local texture regions.

[0005] Furthermore, visual similarity in image space does not necessarily guarantee consistency in frequency domain distribution. Without constraints on local frequency domains or K-space regions, reconstructed images may exhibit problems such as abnormal spectral energy distribution, spurious enhancement of details, or unstable edge structures. Therefore, it is necessary to provide a medical image reconstruction method and apparatus based on structural constraints and spectral co-modulation to improve the structural consistency, boundary sharpness, spectral stability, and high-frequency detail recovery capability of reconstructed medical images. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this application provides a medical image reconstruction method and apparatus based on structural constraints and spectral co-modulation.

[0007] The medical image reconstruction method based on structural constraints and spectral coordinated modulation provided in this application adopts the following technical solution: A medical image reconstruction method based on structural constraints and spectral co-modulation includes the following steps: S1. Acquire the target medical image, and if a reference medical image exists, acquire at least one reference medical image. Perform spatial alignment, intensity normalization, initial upsampling, frequency domain mapping, and streaming rearrangement on the target medical image and the reference medical image to obtain a structured input data stream and the corresponding frequency domain spectral data. S2. Perform structural gradient encoding on the structured input data stream, extract the gradient response, edge response and tissue structure response of the target medical image, and generate structural constraint results; S3. Perform frequency domain mapping, frequency band decoupling, and spectral response analysis on the structured input data stream and / or frequency domain spectral data to generate adaptive spectral weights, and perform collaborative modulation on different frequency components to obtain spectral collaborative modulation results; S4. Fuse the structural constraint results and the spectral co-modulation results, perform space-frequency coupling reconstruction, and output the reconstructed medical image; S5. During the training phase, the reconstructed medical image is mapped to the frequency domain along with the corresponding real high-resolution medical image. Frequency domain consistency error is calculated in multiple local spectral regions, and the reconstruction model parameters are updated based on the frequency domain consistency error.

[0008] By adopting the above technical solution, the target medical image and the reference medical image are first preprocessed and frequency domain mapped to construct a unified structured data input; structural gradient coding is used to extract tissue boundary and anatomical structure information to form structural constraints; combined with the spectrum co-modulation mechanism, different frequency components are adaptively enhanced and optimized; at the same time, local frequency domain consistency calibration is introduced in the training stage to improve the consistency between the reconstruction results and the real high-resolution image in terms of spectrum distribution, thereby improving the structural continuity, edge clarity, high-frequency detail recovery ability and frequency domain stability of medical image reconstruction, and improving reconstruction accuracy and reliability.

[0009] Optionally, in step S1, the target medical image is initially upsampled to obtain the initial representation of the target image Xt0=U(N(It)); Where It is the target medical image, N(·) is the normalization operator, and U(·) is the upsampling operator; In step S2, the horizontal and vertical gradient responses of the target medical image are obtained through gradient analysis; And extract edge response and tissue structure response based on the horizontal gradient response and vertical gradient response; When a reference medical image is present, the edge response of the target medical image is fused with the structural response of the reference medical image to generate a structural prior: Cs = σ(Wt × Et + Wr × Er + bs); in: Cs represents structural prior; Et represents the structural response of the target medical image; Er represents the structural response to the reference medical image; Wt, Wr, and bs are learnable parameters; σ(·) is a nonlinear mapping function.

[0010] By adopting the above technical solution, the target medical image is first normalized and initially upsampled to obtain a target image representation at a uniform scale. Then, the horizontal and vertical gradient information of the image is extracted through gradient analysis to further obtain edge response and tissue structure response, thereby constructing a structural prior that reflects the anatomical structure features. When a reference medical image is present, the structural prior of the target image is enhanced by fusing the structural response information in the reference image, which more accurately preserves tissue boundaries, texture details and structural continuity, providing reliable structural constraints for the subsequent reconstruction process, reducing edge blurring, structural breakage and detail loss, and improving the quality of medical image reconstruction and diagnostic reliability.

[0011] Optionally, step S3 includes: mapping intermediate features in the spatial domain to the frequency domain; Frequency domain features are decoupled by using multiple spectral region masking functions; Generate corresponding adaptive spectral weights based on the response descriptors of each spectral region; The frequency domain features are weighted and modulated using the adaptive spectral weights. The modulated frequency domain features are inversely mapped to the spatial domain and fused with the spatial domain features; The adaptive spectrum weights are obtained in the following manner: ωb=exp(ρbqb) / Σexp(ρjqj); Where: ωb represents the spectral weight corresponding to the b-th spectral region; qb represents the response descriptor for the b-th spectral region; ρb represents the learnable response coefficient.

[0012] By adopting the above technical solution, after mapping the intermediate features of the spatial domain to the frequency domain, multiple spectral region mask functions are used to separate and analyze different frequency components. Spectral weights are adaptively generated based on the response intensity of each spectral region, and low-frequency structural information and high-frequency detail information are differentially modulated and enhanced. Subsequently, the modulated frequency domain features are inversely mapped back to the spatial domain and fused with the original spatial features. This allows the reconstruction process to simultaneously maintain the overall tissue structure and restore local texture details, improving the efficiency of spectral information utilization, enhancing the ability to express high-frequency details, reducing edge over-smoothing, texture distortion, and abnormal spectral energy distribution, and further improving the quality and frequency domain stability of medical image reconstruction.

[0013] Optionally, in step S5, the reconstructed medical image and the real high-resolution medical image are mapped to the frequency domain respectively, and divided into multiple local spectral regions. The frequency domain consistency error is calculated in each local spectral region to generate a local frequency domain consistency constraint signal. The local frequency domain consistency constraint signal is obtained by the following function: Llfc=Σλb‖Mb⊙KX-Mb⊙KHR‖ 2 2 / (‖Mb⊙KHR‖ 2 2+ε); Where: KX is the frequency domain representation of the reconstructed image; KHR is the frequency domain representation of true high-resolution medical images; Mb is a local spectral region masking function; λb is the corresponding weight coefficient; ε is the stability constant.

[0014] By employing the aforementioned technical solution, the reconstructed medical image is mapped to the frequency domain along with the real high-resolution medical image. Frequency domain consistency error is calculated within multiple local spectral regions to generate local frequency domain consistency constraint signals, thereby guiding the model training process. This approach enhances the matching degree between the reconstructed image and the real image in terms of spectral distribution, improves the accuracy of local frequency component recovery, and reduces problems such as spectral distortion, false texture enhancement, and loss of edge details. Ultimately, this improves the detail fidelity, frequency domain stability, and overall reconstruction quality of the medical image.

[0015] Optionally, the space-frequency coupling reconstruction in step S4 adopts a joint optimization objective function: Ltotal=Lrec+β1Lstr+β2Lfcm+β3Llfc; in: Lrec represents the image reconstruction error; Lstr represents structural constraint terms; Lfcm is the spectrum modulation constraint term; Llfc is a local frequency domain consistency constraint term; β1, β2 and β3 are weighting coefficients.

[0016] By adopting the above technical solution, image reconstruction error, structural constraints, spectral modulation constraints, and local frequency domain consistency constraints are jointly optimized to achieve collaborative modeling and comprehensive constraints of spatial and frequency domain information. This can take into account overall structural restoration, edge detail enhancement, and spectral distribution consistency, reduce structural distortion, detail loss, and frequency domain deviation during the reconstruction process, and improve the accuracy, stability, and visual quality of medical image reconstruction.

[0017] Optionally, the medical image is one or more of magnetic resonance imaging, computed tomography (CT) images, positron emission tomography (PET) images, or ultrasound images; preferably, it is a multi-contrast magnetic resonance imaging, and the frequency domain spectral data is K-space spectral data.

[0018] By adopting the above technical solutions, the proposed reconstruction method can be applied to various medical imaging scenarios such as magnetic resonance, CT, PET and ultrasound, and has good versatility and scalability. In particular, the combination of multi-contrast magnetic resonance images and K-space spectral data can make full use of frequency domain information and intermodal structural correlation information to improve the accuracy of tissue structure restoration, the ability to reconstruct details and the quality of medical image reconstruction.

[0019] In a second aspect, a medical image reconstruction device based on structural constraints and spectral co-modulation includes: a multi-source streaming acquisition and adaptive preprocessing system, a structural gradient coding engine, a spectral co-modulation engine, a local frequency domain consistency calibration engine, a space-frequency coupling reconstruction engine, an on-chip cache array, and a programmable data stream computing unit. Each module works in concert to perform the medical image reconstruction method as described in any one of claims 1 to 6.

[0020] By adopting the above technical solutions, the integrated processing of medical image reconstruction can be achieved through the collaborative work of modules such as multi-source streaming acquisition, adaptive preprocessing, structural coding, spectrum modulation, frequency domain calibration, and space-frequency coupling reconstruction. This can improve data processing efficiency and reconstruction accuracy, reduce computational latency, and enhance the structural consistency, detail fidelity, and system stability of the reconstruction results.

[0021] Optionally, the multi-source streaming acquisition and adaptive preprocessing system includes a data interface unit, a multi-source spatiotemporal alignment unit, an intensity adaptive normalization unit, an initial upsampling unit, a frequency domain mapping unit, and a streaming rearrangement control unit. The structural gradient coding engine includes a gradient parsing subunit, an edge response extraction subunit, and a structural constraint generation subunit; The spectrum collaborative modulation engine includes a frequency domain mapping subunit, a frequency band decoupling subunit, a spectrum response analysis subunit, and an adaptive spectrum weighting unit. The local frequency domain consistency calibration engine includes a frequency domain spectrum partitioning subunit, a local frequency domain error calculation subunit, and a frequency domain closed-loop calibration subunit. The on-chip cache array is used to cache the target medical image, the reference medical image, structural feature data, frequency domain spectral data, and intermediate feature data during the reconstruction process; The programmable data stream computing unit is used to complete image preprocessing, structure coding, spectrum modulation, frequency domain calibration, and image reconstruction calculations according to a pipeline mechanism. The programmable dataflow computing unit includes at least one of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), and a field-programmable gate array (FPGA).

[0022] By adopting the above technical solutions, the multi-source streaming acquisition and adaptive preprocessing system can complete the reception, alignment, normalization, upsampling, and frequency domain mapping of medical images, providing standardized input data for subsequent reconstruction. The structural gradient coding engine is used to extract tissue boundary and structural feature information to form structural constraints. The spectrum cooperative modulation engine realizes the analysis and adaptive enhancement of different frequency components. The local frequency domain consistency calibration engine is used to establish frequency domain constraints and optimize the model training process. The on-chip cache array is used to cache image data and intermediate features, reducing data handling and repeated access, and improving system operating efficiency and data processing stability. The programmable dataflow computing unit schedules the collaborative operation of each functional module according to the pipeline mechanism to realize continuous processing of image preprocessing, structural coding, spectrum modulation, frequency domain calibration, and reconstruction calculation. At the same time, by combining at least one computing resource from CPU, GPU, NPU, DSP, and FPGA for parallel acceleration, the system can effectively reduce computational latency and storage access overhead, and improve the real-time processing capability, computational efficiency, and overall operating performance of medical image reconstruction.

[0023] Thirdly, an electronic device includes a processor, a memory, and a computer program stored in the memory and executable by the processor; When the processor executes the computer program, it implements the medical image reconstruction method according to any one of claims 1 to 6.

[0024] By adopting the above technical solution, the electronic device executes a computer program stored in the memory through a processor to realize structural constraint analysis, spectral co-modulation, and space-frequency coupling reconstruction processing of medical images. This enables high-quality medical image reconstruction on general-purpose or dedicated hardware platforms, improving reconstruction efficiency, structural restoration capability, and system deployment flexibility, thus meeting the application needs of different medical imaging scenarios.

[0025] Fourthly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical image reconstruction method according to any one of claims 1 to 6.

[0026] By adopting the above technical solution, the medical image reconstruction method is stored in a computer-readable storage medium in the form of a computer program, enabling the processor to quickly call and execute the corresponding reconstruction process. This facilitates algorithm deployment, upgrading, and porting, and improves the application flexibility and ease of promotion of the medical image reconstruction system.

[0027] In summary, this application includes at least one of the following beneficial technical effects: 1. This application performs structural gradient coding on the target medical image and combines it with a spectral co-modulation mechanism to jointly model spatial domain structural information and frequency domain feature information, thereby achieving the fusion reconstruction of structural constraint results and spectral co-modulation results. This can improve the structural consistency, boundary clarity and high-frequency detail recovery capability of medical images, and obtain higher quality reconstructed images. 2. This application extracts the gradient response, edge response, and tissue structure response of the target medical image, and fuses the structural response information of the reference medical image when a reference medical image is present, to construct an enhanced structural prior. This can effectively preserve tissue boundaries and anatomical structure features, reduce problems such as edge blurring, structural breakage, and loss of local details, and improve the structural fidelity of the reconstructed image. 3. This application, through frequency band decoupling, spectral response analysis, and adaptive spectral weighted modulation of frequency domain features, can dynamically allocate spectral weights according to the importance of different frequency components, thereby achieving synergistic optimization of low-frequency structural information and high-frequency detail information, and improving the efficiency of spectral information utilization and the effect of detail reconstruction. 4. This application introduces a local frequency domain consistency calibration mechanism during the training phase to establish frequency domain constraint relationships in multiple local spectral regions, thereby ensuring that the reconstructed image and the real high-resolution medical image maintain a high degree of consistency in spectral distribution, thus reducing spectral distortion, false texture enhancement, and local frequency domain anomalies, and improving the realism and stability of the reconstruction results.

[0028] 5. This application constructs a joint optimization objective function that includes image reconstruction error, structural constraints, spectral modulation constraints, and local frequency domain consistency constraints. This achieves coordinated optimization of spatial and frequency domain information, taking into account overall structure restoration, local detail enhancement, and spectral stability, thereby improving model training performance and final reconstruction accuracy. 6. This application achieves pipelined processing and efficient caching of target images, reference images, structural feature data, and frequency domain spectral slice data by setting up a multi-source streaming acquisition and adaptive preprocessing system, an on-chip cache array, and a programmable data stream computing unit. This reduces the overhead of repeated reading and writing of intermediate data and storage access, lowers system computational latency and energy consumption, and improves the real-time processing capability and engineering application value of medical image reconstruction. Attached Figure Description

[0029] Figure 1 This is a structural block diagram of the medical image reconstruction device according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of a medical image reconstruction method according to an embodiment of the present invention, wherein, Figure 2 In this context, S10, S21, S22, S23, and S24 are process node numbers in the specific embodiments. Their correspondence with S1-S5 in the claims and invention description is as follows: S10 corresponds to S1, S21 corresponds to S2, S22 corresponds to S3, S24 corresponds to S4, and S23 corresponds to S5. Figure 3 This is a schematic diagram of the structure of the multi-source streaming acquisition and adaptive preprocessing system according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure gradient coding engine according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the spectrum cooperative modulation engine according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the local frequency domain consistency calibration and space-frequency coupling reconstruction engine according to an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the effects of medical image reconstruction in an embodiment of the present invention.

[0030] Figure labeling: 10. Multi-source streaming acquisition and adaptive preprocessing system; 101. Data interface unit; 102. Multi-source spatiotemporal alignment unit; 103. Intensity adaptive normalization unit; 104. Initial upsampling unit; 105. Frequency domain mapping unit; 106. Streaming rearrangement control unit; 20. Structural gradient coding engine; 201. Gradient parsing subunit; 202. Edge response extraction subunit; 203. Structural constraint generation subunit; 30. Spectrum collaborative modulation engine; 301. Frequency domain mapping subunit; 302. Frequency band decoupling subunit; 303. Spectrum response analysis subunit; 304. Adaptive spectrum weighting unit; Local frequency domain consistency calibration engine; 401, frequency domain spectrum partitioning subunit; 402, local frequency domain error calculation subunit; 403, frequency domain closed-loop calibration subunit; 50, space-frequency coupling reconstruction engine; On-chip cache array; 70. Programmable dataflow computing unit. Detailed Implementation

[0031] The following is in conjunction with the appendix Figure 1-7 This application will be described in further detail.

[0032] This application discloses a medical image reconstruction method and apparatus based on structural constraints and spectral co-modulation.

[0033] Reference Figure 1 This application provides a medical image reconstruction device based on structural constraints and spectrum co-modulation, including a multi-source streaming acquisition and adaptive preprocessing system 10, a structural gradient coding engine 20, a spectrum co-modulation engine 30, a local frequency domain consistency calibration engine 40, a space-frequency coupling reconstruction engine 50, an on-chip cache array 60, and a programmable data stream computing unit 70. The structural gradient coding engine 20, the spectrum co-modulation engine 30, the local frequency domain consistency calibration engine 40, and the space-frequency coupling reconstruction engine 50 can be used as functional modules scheduled and executed by the programmable data flow computing unit 70, or integrated into the programmable data flow computing unit 70 to realize pipelined transmission and parallel processing of image block data, structural feature data, and frequency domain spectral data. The target medical image and at least one reference medical image are input into the multi-source streaming acquisition and adaptive preprocessing system 10. The multi-source streaming acquisition and adaptive preprocessing system 10 is used to perform spatial alignment, intensity normalization, initial upsampling, frequency domain mapping and streaming rearrangement on the input image to obtain a structured input data stream and corresponding frequency domain spectral data. The on-chip cache array 60 is used to cache the target medical image, the reference medical image, structural feature data, frequency domain spectral data, and intermediate feature data during the reconstruction process. The programmable dataflow computing unit 70 is used to pipeline the various processing engines, enabling sequential transfer and parallel processing of image block data, structural feature data, and frequency domain spectral data between the processing engines. In one embodiment, the programmable dataflow computing unit 70 is one or more of a central processing unit, a graphics processing unit, a neural network processor, a digital signal processor, or a field-programmable gate array.

[0034] Reference Figure 3The multi-source streaming acquisition and adaptive preprocessing system 10 includes a data interface unit 101, a multi-source spatiotemporal alignment unit 102, an intensity adaptive normalization unit 103, an initial upsampling unit 104, a frequency domain mapping unit 105, and a streaming rearrangement control unit 106. The data interface unit 101 receives the target medical image and, if a reference medical image exists, receives the reference medical image. The multi-source spatiotemporal alignment unit 102 ensures that the target medical image and the reference medical image are in the same image coordinate system when a reference medical image exists. The intensity adaptive normalization unit 103 unifies the image grayscale distribution. The initial upsampling unit 104 obtains the initial representation of the target image. The frequency domain mapping unit 105 obtains frequency domain data. The streaming rearrangement control unit 106 performs block-level rearrangement of the image data and the frequency domain spectral data.

[0035] Reference Figure 4 The structural gradient coding engine 20 includes a gradient parsing subunit 201, an edge response extraction subunit 202, and a structural constraint generation subunit 203. The gradient parsing subunit 201 performs gradient parsing on the low-resolution observation data of the target medical image or the initial representation of the target image obtained through initial upsampling to obtain the gradient response of the target image. The edge response extraction subunit 202 extracts edge information and tissue structure responses from the target image based on the gradient response. The structural constraint generation subunit 203 generates structural condition information based on the gradient response, edge information, and tissue structure responses, forming a structural prior to constrain the reconstruction process of the target medical image. In the presence of a reference medical image, the reference medical image can be aligned and used as an auxiliary reference for structural condition enhancement.

[0036] Reference Figure 5 The spectrum-coordinated modulation engine 30 includes a frequency domain mapping subunit 301, a frequency band decoupling subunit 302, a spectrum response analysis subunit 303, and an adaptive spectrum weighting unit 304. The frequency domain mapping subunit 301 maps spatial domain intermediate features to the frequency domain during reconstruction; the frequency band decoupling subunit 302 separates or partitions the low-frequency structural components and high-frequency detail components in the frequency domain features; the spectrum response analysis subunit 303 analyzes the contribution of different frequency components to the target image reconstruction result; and the adaptive spectrum weighting unit 304 generates learnable spectrum modulation weights and performs weighted modulation on the frequency domain features. The modulated frequency domain features are then fused with the spatial domain intermediate features after inverse frequency domain mapping to form the spectrum-coordinated modulation result.

[0037] Reference Figure 6The local frequency domain consistency calibration engine 40 includes a frequency domain spectrum partitioning subunit 401, a local frequency domain error calculation subunit 402, and a frequency domain closed-loop calibration subunit 403. During the training phase, the frequency domain spectrum partitioning subunit 401 performs local spectrum partitioning on the frequency domain representations of the reconstructed image and the corresponding real high-resolution supervision image, respectively, to obtain multiple corresponding local spectral regions. The local frequency domain error calculation subunit 402 calculates the frequency domain consistency error between the reconstructed image and the real high-resolution supervision image within each corresponding local spectral region. The frequency domain closed-loop calibration subunit 403 generates a frequency domain calibration constraint signal based on the frequency domain consistency error and feeds the frequency domain calibration constraint signal back to the space-frequency coupled reconstruction engine 50 or the programmable data flow calculation unit 70 for updating model parameters during the training phase. During the actual reconstruction phase, the real high-resolution supervision image is not used as input; the frequency domain consistency constraints corresponding to the local frequency domain consistency calibration engine 40 are reflected through the trained reconstruction model parameters.

[0038] It should be noted that the claims and the invention description use S1-S5 to summarize the main steps of the medical image reconstruction method of this application; Figure 2 and the following specific embodiments use S10, S21, S22, S23 and S24 to identify the process nodes in a specific embodiment. These are not different technical solutions. S10 corresponds to S1, S21 to S2, S22 to S3, S24 to S4, and S23 to S5. S23 / S5 are local frequency domain consistency calibration steps in the training phase; no real high-resolution medical image is input in the actual reconstruction phase.

[0039] Based on the above device, referring to Figure 2 This application also provides a medical image reconstruction method, including the following steps: S1, the target medical image is input into the multi-source streaming acquisition and adaptive preprocessing system 10; if a reference medical image exists, at least one reference medical image is input as an auxiliary input into the multi-source streaming acquisition and adaptive preprocessing system 10, and image input buffering, spatial alignment, intensity normalization, initial upsampling, frequency domain mapping and streaming rearrangement processing are completed in the programmable data stream computing unit 70 to obtain a structured input data stream and the corresponding frequency domain spectral data.

[0040] Specifically, the target medical image and the reference medical image are received through the data interface unit 101 and loaded into the on-chip cache array 60; the target medical image and the reference medical image are spatially registered through the multi-source spatiotemporal alignment unit 102; the registered image is normalized through the intensity adaptive normalization unit 103; the target medical image is upsampled through the initial upsampling unit 104 to obtain the initial representation of the target image; the initial representation of the target image and the reference medical image are frequency domain transformed through the frequency domain mapping unit 105, and the spectrum is divided and the data stream is rearranged through the streaming rearrangement control unit 106.

[0041] In one specific implementation, let the target medical image be... Reference medical images are The normalization operator corresponding to the intensity adaptive normalization unit 103 is: The upsampling operator corresponding to the initial upsampling unit 104 is: The spatial registration operator corresponding to the multi-source spatiotemporal alignment unit 102 is: The target image is initially represented as:

[0042] The reference medical image, after spatial registration and intensity normalization, is represented as follows:

[0043] in, This represents the initial representation of the target medical image after normalization and initial upsampling. This represents a reference medical image that is in the same image coordinate system and has a uniform grayscale distribution as the target medical image.

[0044] Furthermore, the streaming rearrangement control unit 106 performs block-level rearrangement of the initial representation of the target image and the reference medical image representation to obtain a structured input data stream:

[0045] The frequency domain mapping unit 105 performs frequency domain transformation on the initial representation of the target image and the representation of the reference medical image, and the streaming rearrangement control unit 106 performs spectral segmentation to obtain frequency domain spectral data:

[0046] in, Represents a structured input data stream. Represents frequency domain spectral data, This represents the structured data stream generation operator. Represents the frequency domain mapping operator, This represents the frequency domain spectral partitioning and data stream rearrangement operator.

[0047] S2, the structured input data stream and frequency domain spectral data are respectively input into the structure gradient coding engine 20, the spectrum co-modulation engine 30, the local frequency domain consistency calibration engine 40 and the space-frequency coupling reconstruction engine 50 to generate structure constraint results, spectrum co-modulation results and local frequency domain consistency calibration results respectively. The space-frequency coupling reconstruction engine 50 fuses the above results and outputs the final reconstructed image.

[0048] Specifically, the structural gradient coding engine 20 performs gradient parsing and edge response extraction on the low-resolution observation data or initial representation of the target medical image to generate structural priors and construct structural constraint terms.

[0049] In one specific implementation, the gradient parsing subunit 201 provides an initial representation of the target image. Gradient analysis is performed to obtain the gradient responses in the horizontal and vertical directions:

[0050] Edge response extraction subunit 202 extracts edge information and tissue structure response in the target image based on the gradient response:

[0051] in, and Let represent the gradient operators in the horizontal and vertical directions, respectively. This represents a constant used to avoid numerical instability.

[0052] In the presence of a reference medical image, the reference medical image is aligned to obtain the reference structure response. The structural constraint generation subunit 203 will generate the edge response of the target image. Response of reference structure The fusion process generates structural condition information:

[0053] in, This represents the structural prior generated from the structural conditions of the target medical image itself and enhanced by auxiliary structural information from a reference medical image. This indicates a convolution operation or a local weighted operation. , and Indicates learnable parameters, This represents a nonlinear mapping function.

[0054] Based on the aforementioned structural priors, structural constraint terms are constructed as follows:

[0055] in, Represents structural constraint terms. Indicates the image to be reconstructed. This represents the gradient of the image to be reconstructed. This represents the structural prior generated from the structural conditions of the target medical image itself. In the presence of a reference medical image, the structural prior can be further enhanced by incorporating auxiliary structural information from the reference medical image.

[0056] The intermediate features are frequency-domain mapped and frequency band decoupled by the spectrum cooperative modulation engine 30. Adaptive spectrum weights are generated according to the response relationship of different spectrum regions, and the frequency domain features are modulated.

[0057] In one specific implementation, let the spatial domain intermediate features of the input spectrum cooperative modulation engine 30 be... The frequency domain mapping subunit 301 maps the intermediate features of the spatial domain to the frequency domain:

[0058] Frequency band decoupling subunit 302 utilizes the first The mask function corresponding to each spectral region Response partitioning of frequency domain features:

[0059] in, Indicates the number of spectrum regions. This indicates element-wise multiplication.

[0060] Spectral response analysis subunit 303 calculates the first Response description for each spectral region:

[0061] in, Represents frequency domain coordinates, Indicates the first The number of sampling points within each spectrum region.

[0062] The adaptive spectrum weighting unit 304 generates adaptive spectrum weights based on the response descriptor:

[0063] in, Indicates the first Adaptive spectral weights for each spectral region This represents the corresponding learnable response coefficient.

[0064] The spectrum cooperative modulation engine 30 performs weighted modulation of the frequency domain features according to the adaptive spectrum weights:

[0065] The modulated frequency domain features are fused with intermediate features in the spatial domain after inverse frequency domain mapping to form a spectrum-co-modulation result:

[0066] in, This indicates the result of spectrum-coordinated modulation. This represents the space-frequency fusion operator.

[0067] During the training phase, the local frequency domain consistency error between the reconstructed image and the corresponding real high-resolution medical image is calculated using the local frequency domain consistency calibration engine 40, with the following constraints:

[0068] in:

[0069] in, This represents the local frequency domain consistency constraint term. Indicates the first Masking functions for each spectral region, Indicates the first The weighting coefficients corresponding to each spectral region The frequency domain representation of the reconstructed image. This represents the frequency domain representation of the target medical image corresponding to the real high-resolution medical image. This represents a constant used to avoid a denominator of zero. In the actual reconstruction phase, the real high-resolution medical image is not used as input; the local frequency domain consistency constraint is reflected through the trained model parameters.

[0070] A joint optimization model is constructed by fusing the structural constraint results, spectral cooperative modulation results, and local frequency domain consistency calibration results from the space-frequency coupling reconstruction engine 50.

[0071] in, This represents the joint optimization objective function. Indicates image reconstruction error. Represents structural constraint terms. Represents the spectrum modulation term. This represents the local frequency domain consistency constraint term. , and These represent the weight coefficients of the corresponding constraint terms.

[0072] The image reconstruction error is:

[0073] in, This represents the imaging degradation operator. Indicates the image to be reconstructed. This represents the input target medical image observation data. This refers to real, high-resolution medical images used during the training or experimental evaluation phase. This represents the observation consistency weighting coefficient.

[0074] The spectrum modulation term is:

[0075] in, This indicates the first generation generated by the spectrum cooperative modulation engine 30. Adaptive spectral weights for each spectral region Indicates the first Masking functions for each spectral region, This represents the initial representation of the target image.

[0076] During the training phase, the space-frequency coupled reconstruction engine 50 updates the reconstruction model parameters according to the joint optimization model; during the actual reconstruction phase, the real high-resolution medical image is not used as input, and the space-frequency coupled reconstruction engine 50 outputs the final reconstructed image according to the trained model parameters.

[0077] In a preferred embodiment, the medical image is a multi-contrast magnetic resonance image, the target medical image is a target contrast low-resolution magnetic resonance image, the reference medical image is a reference contrast magnetic resonance image, and the frequency domain spectral data is K-space spectral data.

[0078] It should be noted that during the training or experimental evaluation phase, the real high-resolution medical image corresponding to the target medical image can be used as the supervision image or evaluation benchmark image; during the actual reconstruction or inference phase, the real high-resolution medical image is not used as input.

[0079] Reference Figure 7 , Figure 7 (a) is the target medical image. Figure 7 (b) is an optional reference medical image. Figure 7 (c) Real high-resolution medical images used during the training or evaluation phase. Figure 7 (d) is the reconstructed image obtained by the method of the present invention. Compared with the input target medical image, the reconstruction result obtained by the present invention is improved in terms of tissue boundary clarity, local structural continuity and high-frequency detail, and is closer to the real high-resolution medical image.

[0080] This application reduces repeated reading and writing of intermediate data through an on-chip cache array 60, realizes pipelined processing of image block data, structural feature data and frequency domain spectral slice data through a programmable data flow computing unit 70, and achieves high-quality reconstruction of target medical images by combining structural gradient coding, spectral co-modulation, local frequency domain consistency calibration and space-frequency coupling reconstruction.

[0081] The implementation principle of a medical image reconstruction method based on structural constraints and spectral co-modulation in this application is as follows: First, a target medical image is acquired, and a reference medical image is acquired simultaneously if one exists. Through preprocessing operations such as spatial alignment, intensity normalization, initial upsampling, frequency domain mapping, and streaming rearrangement, a unified structured input data stream and frequency domain spectral data are constructed to provide a standardized data foundation for the subsequent reconstruction process. The target medical image is then structurally gradient encoded, and the horizontal and vertical gradient responses of the image are obtained through gradient analysis. Edge response and tissue structure response information are further extracted to form a structural prior reflecting tissue boundaries, anatomical structures, and local texture features. If a reference medical image exists, the structural response information in the reference medical image is further fused to enhance the structural prior, thereby providing reliable structural constraints for the subsequent reconstruction process. The spatial domain intermediate features are mapped to the frequency domain, and multiple spectral region mask functions are used to decouple the frequency domain features by frequency band, extracting low-frequency structural information and high-frequency detail information respectively. Adaptive spectral weights are generated based on the response descriptors of each spectral region, and different frequency components are differentially weighted and modulated to give more attention to important frequency regions, thereby enhancing the recovery ability of tissue boundaries, texture details and local structural features. The modulated frequency domain features are then returned to the spatial domain through inverse frequency domain mapping and fused with the spatial domain features to form a spectral co-modulation result. By employing a space-frequency coupling reconstruction mechanism, structural constraint results and spectral co-modulation results are jointly fused. This fully utilizes structural information in the spatial domain and spectral distribution information in the frequency domain to achieve high-quality reconstruction of the target medical image. During the model training phase, the reconstructed medical image is mapped to the frequency domain along with the corresponding real high-resolution medical image and divided into multiple local spectral regions. Frequency domain consistency error is calculated in each local spectral region to generate a local frequency domain consistency constraint signal. Simultaneously, a joint optimization objective function is constructed by combining the image reconstruction error, structural constraint term, spectral modulation constraint term, and local frequency domain consistency constraint term. The model parameters are iteratively updated to ensure that the model simultaneously meets the requirements of spatial structural consistency and spectral distribution consistency. By fully utilizing the structural and spectral information of the target medical image itself, and introducing additional anatomical structural auxiliary information when a reference medical image is available, the spatial and frequency domain information can be synergistically optimized. This effectively improves the structural continuity, tissue boundary clarity, high-frequency detail recovery capability, and spectral stability during the medical image reconstruction process, reduces problems such as edge blurring, structural breakage, detail loss, and spectral distortion, and enhances the realism, accuracy, and clinical application value of the reconstructed image.

[0082] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A medical image reconstruction method based on structural constraints and spectral co-modulation, characterized in that, Includes the following steps: S1. Acquire the target medical image, and if a reference medical image exists, acquire at least one reference medical image. Perform spatial alignment, intensity normalization, initial upsampling, frequency domain mapping, and streaming rearrangement on the target medical image and the reference medical image to obtain a structured input data stream and the corresponding frequency domain spectral data. S2. Perform structural gradient encoding on the structured input data stream, extract the gradient response, edge response and tissue structure response of the target medical image, and generate structural constraint results; S3. Perform frequency domain mapping, frequency band decoupling, and spectral response analysis on the structured input data stream and / or frequency domain spectral data to generate adaptive spectral weights, and perform collaborative modulation on different frequency components to obtain spectral collaborative modulation results; S4. Fuse the structural constraint results and the spectral co-modulation results, perform space-frequency coupling reconstruction, and output the reconstructed medical image; S1-S4 represent the image reconstruction process in the actual reconstruction stage, and S5 represents the model parameter update steps in the training stage. S5. During the training phase, the reconstructed medical image is mapped to the frequency domain along with the corresponding real high-resolution medical image. Frequency domain consistency error is calculated in multiple local spectral regions, and the reconstruction model parameters are updated based on the frequency domain consistency error.

2. The method according to claim 1, characterized in that: In step S1, the target medical image is initially upsampled to obtain the initial representation of the target image Xt0=U(N(It)); Where It is the target medical image, N(·) is the normalization operator, and U(·) is the upsampling operator; In step S2, the horizontal and vertical gradient responses of the target medical image are obtained through gradient analysis; And extract edge response and tissue structure response based on the horizontal gradient response and vertical gradient response; When a reference medical image is present, the edge response of the target medical image is fused with the structural response of the reference medical image to generate a structural prior: Cs = σ(Wt × Et + Wr × Er + bs); in: Cs represents structural prior; Et represents the structural response of the target medical image; Er represents the structural response to the reference medical image; Wt, Wr, and bs are learnable parameters; σ(·) is a nonlinear mapping function.

3. The method according to claim 1, characterized in that: Step S3 includes: mapping intermediate features in the spatial domain to the frequency domain; Frequency domain features are decoupled by using multiple spectral region masking functions; Generate corresponding adaptive spectral weights based on the response descriptors of each spectral region; The frequency domain features are weighted and modulated using the adaptive spectral weights. The modulated frequency domain features are inversely mapped to the spatial domain and fused with the spatial domain features; The adaptive spectrum weights are obtained in the following manner: ωb=exp(ρbqb) / Σexp(ρjqj); Where: ωb represents the spectral weight corresponding to the b-th spectral region; qb represents the response descriptor for the b-th spectral region; ρb represents the learnable response coefficient.

4. The method according to claim 1, characterized in that: In step S5, the reconstructed medical image and the real high-resolution medical image are mapped to the frequency domain respectively and divided into multiple local spectral regions. The frequency domain consistency error is calculated in each local spectral region to generate a local frequency domain consistency constraint signal. The local frequency domain consistency constraint signal is obtained by the following function: Llfc=Λb‖Mb⊙KX-Mb⊙KHR‖ 2 2 / (‖Mb⊙KHR‖ 2 2+e); Where: KX is the frequency domain representation of the reconstructed image; KHR is the frequency domain representation of true high-resolution medical images; Mb is a local spectral region masking function; λb is the corresponding weight coefficient; ε is the stability constant.

5. The method according to claim 1, characterized in that: During the training phase, the space-frequency coupling reconstruction in step S4 adopts a joint optimization objective function: Ltotal=Lrec+β1Lstr+β2Lfcm+β3Llfc; in: Lrec represents the image reconstruction error; Lstr represents structural constraint terms; Lfcm is the spectrum modulation constraint term; Llfc is a local frequency domain consistency constraint term; β1, β2 and β3 are weighting coefficients.

6. The method according to any one of claims 1 to 5, characterized in that: The medical image is one or more of magnetic resonance imaging, computed tomography (CT) images, positron emission tomography (PET) images, or ultrasound images; preferably, it is a multi-contrast magnetic resonance imaging, and the frequency domain spectral data is K-space spectral data.

7. A medical image reconstruction device based on structural constraints and spectral co-modulation, characterized in that, include: Multi-source streaming acquisition and adaptive preprocessing system, structure gradient coding engine, spectrum cooperative modulation engine, local frequency domain consistency calibration engine, space-frequency coupling reconstruction engine, on-chip cache array, programmable data stream computing unit; Each module works in concert to perform the medical image reconstruction method as described in any one of claims 1 to 6.

8. The apparatus according to claim 7, characterized in that: The multi-source streaming acquisition and adaptive preprocessing system includes a data interface unit, a multi-source spatiotemporal alignment unit, an intensity adaptive normalization unit, an initial upsampling unit, a frequency domain mapping unit, and a streaming rearrangement control unit. The structural gradient coding engine includes a gradient parsing subunit, an edge response extraction subunit, and a structural constraint generation subunit; The spectrum collaborative modulation engine includes a frequency domain mapping subunit, a frequency band decoupling subunit, a spectrum response analysis subunit, and an adaptive spectrum weighting unit. The local frequency domain consistency calibration engine includes a frequency domain spectrum partitioning subunit, a local frequency domain error calculation subunit, and a frequency domain closed-loop calibration subunit. The on-chip cache array is used to cache the target medical image, the reference medical image, structural feature data, frequency domain spectral data, and intermediate feature data during the reconstruction process; The programmable data stream computing unit is used to complete image preprocessing, structure coding, spectrum modulation, frequency domain calibration, and image reconstruction calculations according to a pipeline mechanism. The programmable dataflow computing unit includes at least one of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), and a field-programmable gate array (FPGA).

9. An electronic device, characterized in that: Includes a processor, a memory, and a computer program stored in the memory and executable by the processor; When the processor executes the computer program, it implements the medical image reconstruction method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the medical image reconstruction method according to any one of claims 1 to 6.

Images