A medical image super-resolution reconstruction method and device based on anatomical semantic guidance, electronic equipment and storage medium

Abstract

The present application relates to a kind of medical image super-resolution reconstruction methods based on anatomical semantic guidance, comprising: acquisition low-resolution MRI image, to extract semantic information based on constructing anatomical semantic extractor;Import semantic information to double modulation parameter generator, to map the semantic information as modulation parameter;Import the modulation parameter to semantic denoising engine for semantic guided denoising and obtain final residual;Final residual fusion, to obtain super-resolution image.The device includes acquisition, first import, second import and output module.Also include electronic equipment and storage medium.A kind of medical image super-resolution reconstruction methods based on anatomical semantic guidance, device, electronic equipment and storage medium designed using the present application, can solve the problem that existing learning type super-resolution method is prone to boundary blur or pseudo detail in medical scene, the stability and explainability of the details of reconstruction are insufficient, lack of consistent constraint with anatomical structure.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer vision technology, specifically to a method, apparatus, electronic device, and storage medium for medical image super-resolution reconstruction based on anatomical semantic guidance. Background Technology

[0002] Magnetic resonance imaging (MRI) has significant applications in imaging the nervous system and soft tissues. However, constraints such as scan time, signal-to-noise ratio, hardware performance, and clinical procedures often necessitate trade-offs between imaging resolution and speed, resulting in low-resolution or thick-slice scans being prevalent in clinical and research settings. To improve image detail representation and subsequent quantitative analysis accuracy, medical image super-resolution reconstruction techniques typically utilize image priors and learned models to recover higher-resolution structural information without incurring additional scanning costs.

[0003] Existing learning-based super-resolution methods still face several common challenges in medical scenarios: First, subtle anatomical boundaries and tissue partitions often differ from natural image textures, and relying solely on pixel errors or perceptual losses can easily lead to blurred boundaries or false details; Second, the stability and interpretability of reconstructed details are insufficient under domain shifts between different subjects, devices, and sequences; Third, although generative reconstruction models can improve visual appeal and high-frequency components, the lack of constraints consistent with anatomical structures may introduce artifacts inconsistent with the actual tissue distribution, thereby affecting the reliability of diagnostic or quantitative tasks.

[0004] Therefore, existing technologies have shortcomings and need to be improved and developed. Summary of the Invention

[0005] The present invention provides a method, apparatus, electronic device, and storage medium for medical image super-resolution reconstruction based on anatomical semantic guidance, which addresses several common challenges faced by existing learning-based super-resolution methods in medical scenarios: First, subtle anatomical boundaries and tissue partitions often differ from natural image textures, and relying solely on pixel errors or perceptual losses can easily lead to blurred boundaries or pseudo-details; Second, the stability and interpretability of reconstructed details are insufficient under domain shifts between different subjects, different devices, and different sequences; Third, although generative reconstruction models can improve visual appeal and high-frequency components, if constraints consistent with anatomical structures are lacking, artifacts inconsistent with the actual tissue distribution may be introduced, thereby affecting the reliability of diagnostic or quantitative tasks.

[0006] In a first aspect, this invention provides a method for super-resolution reconstruction of medical images based on anatomical semantic guidance, comprising:

[0007] Acquiring low-resolution MRI images To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ;

[0008] The semantic information is imported into the Dual Modulation Parameter Generator (DMPG) to map the semantic information into modulation parameters, which include at least the modulation parameters for Spatial Adaptive Modulation (FiLM) and Global Channel Modulation (AdaLN).

[0009] The modulation parameters are imported into the semantic denoising engine (SDE), and using the UNet / ResShift denoising backbone as the main body, spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ;

[0010] Will Upsampling With the final residual Fusion to obtain super-resolution images : .

[0011] Furthermore, the acquisition of low-resolution MRI images To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ,include:

[0012] Using a U-Net-based medical segmentation network from Predicting multi-channel soft segmentation probability plot : ,in, This indicates that the medical segmentation network is used to output a soft segmentation probability map. , Indicates the number of channels. express Image height, express Image width, Indicates all shapes as The set of real three-dimensional tensors;

[0013] soft segmentation probability map Input to a lightweight convolutional encoder to obtain local spatial semantic features : ,in, This indicates that a lightweight convolutional encoder is used to output local spatial semantic features. , express The number of output channels; Indicates all shapes as The set of real three-dimensional tensors;

[0014] At the global semantic level, using segmentation probability graphs As spatial weights, they influence local spatial semantic features. Performing weighted global pooling to extract the prototype vectors for each anatomical category, then for the ... The class, whose global semantic token is defined as: ,in, For category number and , Represents spatial index and , Represents a soft-segmented probability graph In spatial location The first The probability value of the class. Representing the spatial location of the feature map The feature vector at that location, To prevent division by zero of extremely small constants, This is used to project the extracted prototype vectors of each anatomical category onto a unified dimension to form a global semantic token for subsequent modulation.

[0015] Collect all global semantic tokens to obtain global category semantics. : .

[0016] Furthermore, If the number of channels is 4, and the categories include background, white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF), then the categories are numbered 1, 2, 3, and 4 in the order from background, white matter (WM), gray matter (GM) to CSF.

[0017] Furthermore, the semantic information is imported into the Dual Modulation Parameter Generator (DMPG) to map the semantic information into modulation parameters, which at least include modulation parameters for Spatial Adaptive Modulation (FiLM) and Global Channel Modulation (AdaLN), including:

[0018] Import the semantic information into the dual modulation parameter generator DMPG;

[0019] For the Each scale level will be used to create a soft segmentation probability map. Downsampling to the corresponding resolution and through the spatial parameter generator Generate FiLM modulation parameters: ,in, The UNet / ResShift denoising backbone is represented at different resolution scales. Each scale level and , The total number of scale levels in the UNet / ResShift denoising backbone; They represent the first The number of channels, height, and width corresponding to each scale level; Indicates all shapes as The set of real three-dimensional tensors; Indicates the first The spatial parameter generator used at each scale has a structural feature consisting of 3×3 convolutions and 1×1 convolutions, with the input being... The output is split to obtain a spatially adaptive scaling factor. Translation coefficients that adapt to space ;

[0020] global category semantics Concatenate to form a global context vector : and through the channel parameter generator Generate AdaLN modulation parameters: ,in, This indicates that the concatenation operator is used to concatenate multiple tokens; Indicates the first The channel parameter generator used at each scale has the following input: The output is split to obtain the parameters used for scaling. and parameters used for translation ,and Broadcasting by channel in the spatial dimension to act upon The feature map.

[0021] Furthermore, the modulation parameters are imported into the semantic denoising engine (SDE), and using the UNet / ResShift denoising backbone as the main body, spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ,include:

[0022] In the semantic denoising engine SDE, Injected into the spatial modulation module (DSM) to provide location-dependent local structure guidance, Injected into the Global Channel Modulation Module (GAC) to provide global semantic constraints, and repeatedly injected at multiple scale levels to achieve the effect of noise residuals Anatomical structure perception denoising prediction, and after High-frequency residuals are obtained after iterative denoising. ;

[0023] right Upsampling was performed to obtain the reconstructed base. ;

[0024] At each time step Denoising network The ResShift / UNet denoising backbone is used as the main body, and a hybrid modulation module is used to replace the standard residual block. Spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected at the feature level.

[0025] Let the first Layer input features are Hybrid modulation is calculated in the order of global calibration first, followed by local constraints: Then apply FiLM spatial modulation: ,in This represents element-wise multiplication;

[0026] The output residual with semantic modulation is: ;

[0027] Under the constraints of the above multi-level modulation, the semantically guided single-step denoising process can be described as follows: Specifically, it means network. In soft segmentation probability map With global category semantics Under the combined conditions, from the noise state Gradually restore a clearer residual state. and after The final residual is obtained through iterative denoising. .

[0028] Furthermore, when obtaining super-resolution images After that, it also includes:

[0029] When the overall high-resolution slice size is large, Cut into several overlapping sub-blocks ,in This indicates the total number of overlapping sub-blocks. Indicates the first overlapping sub-blocks and ;

[0030] The final residual is obtained by performing the operation on several overlapping sub-blocks separately. Steps to obtain residual prediction ;

[0031] Residual prediction The results are fused with the corresponding upsampled basis to generate local super-resolution results: ;

[0032] The overlapping regions of several overlapping sub-blocks are normalized and weighted by a smooth weight window to reorganize them, thereby reducing memory usage, mitigating block boundary artifacts, and improving spatial consistency: Let the pixel set of the nth sub-block in the overall image coordinate system be... And define a smooth weight window for it. The entire reconstruction result is obtained by normalized weighted summation: Where Ω represents the pixel domain of the entire image. , To prevent division by zero of extremely small constants.

[0033] A second aspect of the present invention provides a medical image super-resolution reconstruction device based on anatomical semantic guidance, comprising:

[0034] The acquisition module is used to acquire low-resolution MRI images. To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ;

[0035] The first import module is used to import the semantic information into the dual modulation parameter generator DMPG to map the semantic information into modulation parameters, wherein the modulation parameters include at least the modulation parameters for spatial adaptive modulation FiLM and global channel modulation AdaLN.

[0036] The second import module is used to import the modulation parameters into the semantic denoising engine (SDE). Using the UNet / ResShift denoising backbone as the main body, it injects spatial modulation parameters for FiLM and global modulation parameters for AdaLN during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ;

[0037] Output module, used to output Upsampling With the final residual Fusion to obtain super-resolution images : .

[0038] In a third aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the electronic device implements the aforementioned anatomical semantic-guided medical image super-resolution reconstruction method.

[0039] In a fourth aspect, the present invention provides a computer-readable storage medium for storing a computer program that, when run on a computer, causes the computer to perform the described anatomically semantically guided medical image super-resolution reconstruction method.

[0040] Beneficial effects:

[0041] As can be seen from the above technical solutions, the present invention provides a medical image super-resolution reconstruction method based on anatomical semantic guidance, which has the following beneficial effects:

[0042] 1. Improved structural consistency: Local soft segmentation probability provides spatial distribution conditions, and global category semantics provides organizational composition conditions. The two together constrain the generative denoising process, which helps to keep the reconstruction results consistent with the anatomical structure and reduce the risk of structural drift and artifacts.

[0043] 2. Balancing detail restoration with low-frequency stability: with The model primarily learns high-frequency details and boundary information, while low-frequency intensity and overall contrast are maintained by the base, thereby improving stability and controllability.

[0044] 3. Multi-scale collaborative constraints: Modulation parameters are generated and injected at different scale levels, so that the coarse scale is responsible for organizing the layout and the consistency of the large structure, while the fine scale is responsible for the restoration of boundary and texture details, which helps to reduce the detail distortion caused by insufficient single-scale constraints.

[0045] 4. Explainability and reproducibility of semantic conditions: The global semantic token is obtained by soft probability weighted pooling, which has a clear calculation path and is easy to reproduce and debug.

[0046] 5. Enhanced engineering deployment capabilities: Through overlapping block segmentation and weighted window reorganization, large-size slices can be processed under memory constraints, and block splicing artifacts can be reduced, improving the feasibility of implementation in clinical and engineering scenarios.

[0047] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other.

[0048] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0049] The accompanying drawings are not drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein:

[0050] Figure 1 This is a flowchart illustrating the overall process of a medical image super-resolution reconstruction method based on anatomical semantic guidance, as described in this application.

[0051] Figure 2 This is a flowchart of step S102 of a medical image super-resolution reconstruction method based on anatomical semantic guidance in an embodiment of this application.

[0052] Figure 3 This is a flowchart of step S104 of a medical image super-resolution reconstruction method based on anatomical semantic guidance in an embodiment of this application.

[0053] Figure 4 This is a flowchart of step S106 of a medical image super-resolution reconstruction method based on anatomical semantic guidance in an embodiment of this application.

[0054] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application.

[0055] Figure 6 This is an overall framework diagram of a medical image super-resolution reconstruction method based on anatomical semantic guidance, as described in an embodiment of this application.

[0056] Figure 7 This is a framework diagram of an anatomical semantic extractor for a medical image super-resolution reconstruction method based on anatomical semantic guidance, as described in an embodiment of this application.

[0057] Figure 8 This is a schematic diagram illustrating the dual modulation parameter generation and injection of a medical image super-resolution reconstruction method based on anatomical semantic guidance in an embodiment of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art to which this invention pertains.

[0059] The terms "first," "second," and similar words used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, unless the context clearly indicates otherwise, the singular forms of "an," "a," or "the," etc., do not indicate a quantity limitation, but rather indicate the presence of at least one. Terms such as "comprising" or "including" mean that the element or object preceding "comprising" encompasses the features, integrals, steps, operations, elements, and / or components listed following "comprising" or "including," and do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0060] Existing learning-based super-resolution methods still face several common challenges in medical scenarios: First, subtle anatomical boundaries and tissue partitions often differ from natural image textures, and relying solely on pixel errors or perceptual losses can easily lead to blurred boundaries or false details; Second, the stability and interpretability of reconstructed details are insufficient under domain shifts between different subjects, devices, and sequences; Third, although generative reconstruction models can improve visual appeal and high-frequency components, the lack of constraints consistent with anatomical structures may introduce artifacts inconsistent with the actual tissue distribution, thereby affecting the reliability of diagnostic or quantitative tasks.

[0061] Therefore, embodiments of the present invention provide a method for super-resolution reconstruction of medical images based on anatomical semantic guidance, referring to... Figure 1 and Figure 6 ,in Figure 6 This example illustrates how to obtain prior anatomical information from monkey brain MRI, including:

[0062] Step S102: Acquire low-resolution MRI images To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. .

[0063] Step S104: Import the semantic information into the Dual Modulation Parameter Generator (DMPG) to map the semantic information into modulation parameters, which include at least modulation parameters for Spatial Adaptive Modulation (FiLM) and Global Channel Modulation (AdaLN).

[0064] Step S106: Import the modulation parameters into the semantic denoising engine SDE, and using the UNet / ResShift denoising backbone as the main body, inject spatial modulation parameters for FiLM and global modulation parameters for AdaLN during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. .

[0065] Step S108: ... Upsampling With the final residual Fusion to obtain super-resolution images : .

[0066] First, two types of semantic information are extracted from the input low-resolution image: a soft segmentation probability map representing the distribution of local anatomical structures and a global category semantic representing the overall tissue composition. Then, the semantic information is mapped to two types of modulation parameters, which are used for spatial location-related modulation and channel-related modulation, respectively. The modulation parameters are then injected into the inversion iteration process of the diffusion denoising backbone, so that the denoising prediction outputs the final residual under semantic constraints. Finally, the reconstructed base obtained by upsampling is summed and fused with the residual to obtain the super-resolution output image.

[0067] By establishing a closed-loop conditional generation framework encompassing anatomical semantics, multi-granularity modulation parameters, and diffusion denoising residual prediction, this framework provides spatial distribution information of local structures using soft segmentation probability maps, enabling the model to obtain differentiated structural constraints at different locations. Simultaneously, it provides global conditions at the tissue composition level using global category semantics, ensuring that channel representations maintain tissue semantic consistency across the entire map. Furthermore, it employs residual form fusion with upsampling basis mapping, focusing the generative model's predictions on high-frequency compensation rather than overall intensity reconstruction, thereby improving controllability and stability. To address the issue of insufficient anatomical structural consistency in generative reconstruction, a joint semantic modulation of local and global elements is introduced at the feature level of diffusion iteration, with detail compensation output as a residual.

[0068] By introducing joint conditions of local spatial semantics and global category semantics during the denoising and inversion process, the consistency of reconstruction of anatomical boundaries and tissue partitions can be improved, and the risks of false details and structural drift can be reduced. The use of residual compensation and fusion with the upsampled substrate helps to maintain the stability of low-frequency intensity and overall contrast, while enhancing the ability to recover high-frequency details, thereby improving the overall performance of the reconstructed image in terms of visual quality and structural credibility.

[0069] In some embodiments, the acquisition of low-resolution MRI images To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. , refer to Figure 2 and Figure 7 ,include:

[0070] Step S1021: Utilize a U-Net-based medical segmentation network from... Predicting multi-channel soft segmentation probability plot : ,in, This indicates that the medical segmentation network is used to output a soft segmentation probability map. , Indicates the number of channels. express Image height, express Image width, Indicates all shapes as The set of real three-dimensional tensors.

[0071] Step S1022: Calculate the soft segmentation probability map Input to a lightweight convolutional encoder to obtain local spatial semantic features : ,in, This indicates that a lightweight convolutional encoder is used to output local spatial semantic features. , express The number of output channels; Indicates all shapes as The set of real three-dimensional tensors.

[0072] Step S1023: At the global semantic level, use the segmentation probability graph As spatial weights, they influence local spatial semantic features. Performing weighted global pooling to extract the prototype vectors for each anatomical category, then for the ... The class, whose global semantic token is defined as: ,in, For category number and , Represents spatial index and , Represents a soft-segmented probability graph In spatial location The first The probability value of the class. Representing the spatial location of the feature map The feature vector at that location, To prevent division by zero of extremely small constants, This is used to project the extracted prototype vectors of each anatomical category onto a unified dimension to form a global semantic token for subsequent modulation.

[0073] Step S1024: Collect all global semantic tokens to obtain global category semantics. : .

[0074] A segmentation network is used to output a multi-channel soft segmentation probability map for low-resolution images; the probability map is input into a lightweight convolutional encoder to obtain local spatial semantic features; the probability map is used as spatial weights to perform weighted global pooling on the local features to obtain the prototype vector corresponding to each anatomical category, and then projected through a multilayer perceptron to form a global semantic token; all category tokens are collected to form a global category semantic set.

[0075] Unlike using only hard segmentation labels or global average pooling, this method employs soft probability as weights for weighted pooling. This allows each category of token to more accurately aggregate the feature contributions of the spatially occupied region and reduces the contamination of category semantics by uncertain boundary regions. Simultaneously, projection aligns different category prototypes to a unified semantic space, facilitating stable use of the subsequent modulation parameter generator in multi-scale networks. This mechanism ensures that "global semantics" originates from spatial distribution information and possesses a computable and reproducible formation path.

[0076] By using soft probability weighted prototype extraction, we can obtain organizational-level semantic representations more robustly under low-resolution conditions, enhancing cross-sample consistency. At the same time, it provides structured and interpretable semantic conditions for subsequent global channel modulation, which helps to reduce class confusion and intensity drift in generative reconstruction, thereby improving organizational consistency and reconstruction reliability.

[0077] In some embodiments, If the number of channels is 4, and the categories include background, white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF), then the categories are numbered 1, 2, 3, and 4 in the order from background, white matter (WM), gray matter (GM) to CSF.

[0078] The number of segmentation channels is limited to four categories, with clearly defined category sets and numbering orders, corresponding to background, white matter, gray matter, and cerebrospinal fluid, respectively. The category system and channel order of semantic conditions are standardized and constrained to ensure that the semantic channel alignment during training and inference phases is fixed and verifiable, avoiding semantic condition mismatches caused by inconsistent category encoding.

[0079] In some embodiments, the semantic information is imported into a dual modulation parameter generator (DMPG) to map the semantic information into modulation parameters, which at least include modulation parameters for spatial adaptive modulation (FiLM) and global channel modulation (AdaLN), as referenced. Figure 3 and Figure 8 ,include:

[0080] Step S1041: Import the semantic information into the dual modulation parameter generator DMPG.

[0081] Step S1042: For the first Each scale level will be used to create a soft segmentation probability map. Downsampling to the corresponding resolution and through the spatial parameter generator Generate FiLM modulation parameters: ,in, The UNet / ResShift denoising backbone is represented at different resolution scales. Each scale level and , The total number of scale levels in the UNet / ResShift denoising backbone; They represent the first The number of channels, height, and width corresponding to each scale level; Indicates all shapes as The set of real three-dimensional tensors; Indicates the first The spatial parameter generator used at each scale has a structural feature consisting of 3×3 convolutions and 1×1 convolutions, with the input being... The output is split to obtain a spatially adaptive scaling factor. Translation coefficients that adapt to space .

[0082] Step S1043: Assign global category semantics Concatenate to form a global context vector : and through the channel parameter generator Generate AdaLN modulation parameters: ,in, This indicates that the concatenation operator is used to concatenate multiple tokens; Indicates the first The channel parameter generator used at each scale has the following input: The output is split to obtain the parameters used for scaling. and parameters used for translation ,and Broadcasting by channel in a spatial dimension to act upon The feature map.

[0083] For each scale level of the denoising backbone, the soft segmentation probability map is downsampled to the corresponding resolution and input into the spatial parameter generator, which outputs spatially adaptive scaling and translation parameters for spatial modulation. At the same time, the global category semantic tokens are concatenated into a global context vector, which is input into the channel parameter generator, which outputs channel-dimensional scaling and translation parameters for global channel modulation, and explains that the channel parameters are broadcast by channel in the spatial dimension.

[0084] On the one hand, spatial modulation parameters with the same resolution as the feature map at different scale levels are generated, ensuring that local structural constraints remain effective at the feature level from coarse to fine. On the other hand, a unified global context is formed by concatenating global semantic tokens, and then channel modulation parameters are generated layer by layer, enabling the global semantics to have adaptable channel calibration capabilities at different scales. This design explicitly transforms local and global semantics into modulation variables usable by the network through a computable mapping.

[0085] Multi-scale spatial modulation can constrain structural morphology at different resolution stages, reducing boundary drift caused by insufficient constraints at a single scale; global channel modulation can enhance tissue-level semantic consistency and stabilize feature distribution, helping to reduce artifacts and tissue misreconstruction, thereby improving reconstruction quality and semantic consistency.

[0086] In some embodiments, the modulation parameters are imported into the semantic denoising engine (SDE), and a UNet / ResShift denoising backbone is used as the main body. During the diffusion inversion process, spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected to perform semantically guided denoising and obtain the final residual. , refer to Figure 4 ,include:

[0087] Step S1061: In the semantic denoising engine SDE, Injected into the spatial modulation module (DSM) to provide location-dependent local structure guidance, Injected into the Global Channel Modulation Module (GAC) to provide global semantic constraints, and repeatedly injected at multiple scale levels to achieve the effect of noise residuals Anatomical structure perception denoising prediction, and after High-frequency residuals are obtained after iterative denoising. .

[0088] Step S1062: For Upsampling was performed to obtain the reconstructed base. .

[0089] Step S1063: At each time step Denoising network The ResShift / UNet denoising backbone is used as the main body, and a hybrid modulation module is used to replace the standard residual block. Spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected at the feature level.

[0090] Step S1064: Let the first... Layer input features are Hybrid modulation is calculated in the order of global calibration first, followed by local constraints: Then apply FiLM spatial modulation: ,in This indicates element-wise multiplication.

[0091] Step S1065: Output the semantically modulated residual as follows: .

[0092] Step S1066: Under the constraints of the above multi-level modulation, the semantically guided single-step denoising process is described as follows: Specifically, it means network. In soft segmentation probability map With global category semantics Under the combined conditions, from the noise state Gradually restore a clearer residual state. and after The final residual is obtained through iterative denoising. .

[0093] Spatial modulation parameters are injected into the spatial modulation module to provide location-dependent local guidance, while global channel modulation parameters are injected into the global channel modulation module to provide global semantic constraints. This injection is repeated at multiple scale levels. Within each time step, a hybrid modulation sequence—global-to-local—is achieved by first performing normalization calibration of the global channel modulation and then spatial modulation. The residual structure, constructed from convolution and activation, outputs the semantically modulated residual prediction. The final residual is obtained after one iteration.

[0094] First, global channel modulation is used to perform semantic consistency calibration on the feature distribution, which can be regarded as a global constraint on organizational semantics. Then, spatial modulation is used to apply local structural guidance at the pixel location level, which can be regarded as a local constraint on boundaries and morphology. This sequence enables the two types of semantic conditions to form synergy rather than mutual interference in the same residual block, and through multi-scale repeated injection, the semantic constraints are propagated along the network depth and scale.

[0095] Sequential global calibration and local constraints can improve the control of semantic conditions on the denoising trajectory and reduce the randomness of generative denoising in details that does not conform to the anatomical structure. Multi-scale repeated injection is conducive to ensuring tissue layout in coarse layers and restoring boundaries and textures in fine layers, thereby improving structural consistency, boundary clarity and detail credibility, and reducing the probability of artifacts.

[0096] In some embodiments, when obtaining a super-resolution image After that, it also includes:

[0097] When the overall high-resolution slice size is large, Cut into several overlapping sub-blocks ,in This indicates the total number of overlapping sub-blocks. Indicates the first overlapping sub-blocks and .

[0098] The final residual is obtained by performing the operation on several overlapping sub-blocks separately. Steps to obtain residual prediction .

[0099] Residual prediction The results are fused with the corresponding upsampled basis to generate local super-resolution results: .

[0100] The overlapping regions of several overlapping sub-blocks are normalized and weighted by a smooth weight window to reorganize them, thereby reducing memory usage, mitigating block boundary artifacts, and improving spatial consistency: Let the pixel set of the nth sub-block in the overall image coordinate system be... And define a smooth weight window for it. The entire reconstruction result is obtained by normalized weighted summation: Where Ω represents the pixel domain of the entire image. , To prevent division by zero of extremely small constants.

[0101] The low-resolution image is cropped into multiple sub-blocks with overlapping regions. Residual prediction is performed on each sub-block and fused with the corresponding upsampled basis to obtain a local super-resolution result. The overlapping regions are normalized and weighted by a smooth weight window to obtain the whole super-resolution result, which is used to reduce memory usage and alleviate block boundary artifacts.

[0102] Diffusion denoising iterations typically have high computational costs, and directly processing large images can be limited by video memory. Overlapping blocks can break down the computation into manageable scales. By normalizing and recombining the weighted smoothing windows, a continuous transition can be formed in the overlapping areas, avoiding the blocky boundary artifacts introduced by simple stitching.

[0103] Block-based inference can handle large slices under hardware resource constraints, improving the usability and deployability of the method; normalized reorganization using smooth weight windows can reduce stitching marks in overlapping areas, improve spatial consistency and visual continuity, thereby improving the overall reconstruction quality.

[0104] Another embodiment of the present invention provides a medical image super-resolution reconstruction device based on anatomical semantic guidance, comprising:

[0105] The acquisition module is used to acquire low-resolution MRI images. To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ;

[0106] The first import module is used to import the semantic information into the dual modulation parameter generator DMPG to map the semantic information into modulation parameters, wherein the modulation parameters include at least the modulation parameters for spatial adaptive modulation FiLM and global channel modulation AdaLN.

[0107] The second import module is used to import the modulation parameters into the semantic denoising engine (SDE). Using the UNet / ResShift denoising backbone as the main body, it injects spatial modulation parameters for FiLM and global modulation parameters for AdaLN during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ;

[0108] Output module, used to convert Upsampling With the final residual Fusion to obtain super-resolution images : .

[0109] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0110] Based on the same inventive concept as the above method embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it enables the electronic device to implement the control method described in the above embodiments.

[0111] In one embodiment, the electronic device may be a server, and in this embodiment, the structure of the electronic device may be as follows: Figure 5 As shown, it includes a memory, a communication module, and one or more processors.

[0112] Memory is used to store computer programs executed by the processor. Memory can be mainly divided into a program storage area and a data storage area. The program storage area can store the operating system and programs required to run instant messaging functions, etc.; the data storage area can store various instant messaging information and operation instruction sets, etc.

[0113] Memory can be volatile memory, such as random access memory (RAM); memory can also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory can be any other medium capable of carrying or storing a desired computer program having the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory can be a combination of the above-mentioned types of memory.

[0114] A processor may include one or more central processing units (CPUs) or digital processing units, etc. The processor is used to implement the aforementioned audio data processing methods when it invokes computer programs stored in memory.

[0115] The communication module is used to communicate with terminal devices and other servers.

[0116] This application embodiment does not limit the specific connection medium between the above-described memory, communication module, and processor. This application embodiment... Figure 5 The memory and processor are connected via a bus, and the bus is in... Figure 5 The connections between other components are illustrated with arrows and are for illustrative purposes only, not as limiting information. Buses can be categorized as address buses, data buses, control buses, etc. For ease of description, Figure 5 The text uses only one arrow to describe it, but does not indicate that there is only one bus or one type of bus.

[0117] Based on the same inventive concept as the above-described method embodiments, embodiments of the present invention also provide a computer-readable storage medium for storing a computer program. When the computer program is run on a computer, it enables an electronic device to implement the control methods described in the above embodiments. The computer-readable storage medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0118] Based on the same inventive concept as the above-described method embodiments, embodiments of the present invention also provide a computer program product. The computer program product includes a computer program that, when run on an electronic device, causes the electronic device to perform the steps of the control methods described above according to various exemplary embodiments of this application. The program product may take the form of any combination of one or more readable media. These computer program commands can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the commands executed by the processor of the computer or other programmable data processing device generate a process for implementing... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for super-resolution reconstruction of medical images based on anatomical semantic guidance, characterized in that, include: Acquiring low-resolution MRI images To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ; The semantic information is imported into the Dual Modulation Parameter Generator (DMPG) to map the semantic information into modulation parameters, which include at least the modulation parameters for Spatial Adaptive Modulation (FiLM) and Global Channel Modulation (AdaLN). The modulation parameters are imported into the semantic denoising engine (SDE), and using the UNet / ResShift denoising backbone as the main body, spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ; Will Upsampling With the final residual Fusion to obtain super-resolution images : .

2. The method for super-resolution reconstruction of medical images based on anatomical semantic guidance according to claim 1, characterized in that, The acquisition of low-resolution MRI images To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ,include: Using a U-Net-based medical segmentation network from Predicting multi-channel soft segmentation probability plot : ,in, This indicates that the medical segmentation network is used to output a soft segmentation probability map. , Indicates the number of channels. express Image height, express Image width, Indicates all shapes as The set of real three-dimensional tensors; soft segmentation probability map Input to a lightweight convolutional encoder to obtain local spatial semantic features : ,in, This indicates that a lightweight convolutional encoder is used to output local spatial semantic features. , express The number of output channels; Indicates all shapes as The set of real three-dimensional tensors; At the global semantic level, using segmentation probability graphs As spatial weights, they influence local spatial semantic features. Performing weighted global pooling to extract the prototype vectors for each anatomical category, then for the ... The class, whose global semantic token is defined as: ,in, For category number and , Represents spatial index and , Represents a soft-segmented probability graph In spatial location The first The probability value of the class. Representing the spatial location of the feature map The feature vector at that location, To prevent division by zero of extremely small constants, This is used to project the extracted prototype vectors of each anatomical category onto a unified dimension to form a global semantic token for subsequent modulation. Collect all global semantic tokens to obtain global category semantics. : .

3. The method for super-resolution reconstruction of medical images based on anatomical semantic guidance according to claim 2, characterized in that, If the number of channels is 4, and the categories include background, white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF), then the categories are numbered 1, 2, 3, and 4 in the order from background, white matter (WM), gray matter (GM) to CSF.

4. The method for super-resolution reconstruction of medical images based on anatomical semantic guidance according to claim 2, characterized in that, The semantic information is imported into the Dual Modulation Parameter Generator (DMPG) to map the semantic information into modulation parameters. These modulation parameters include at least those used for Spatial Adaptive Modulation (FiLM) and Global Channel Modulation (AdaLN), including: Import the semantic information into the dual modulation parameter generator DMPG; For the first Each scale level will be used to create a soft segmentation probability map. Downsampling to the corresponding resolution and through the spatial parameter generator Generate FiLM modulation parameters: ,in, The UNet / ResShift denoising backbone is represented at different resolution scales. Each scale level and , The total number of scale levels in the UNet / ResShift denoising backbone; They represent the first The number of channels, height, and width corresponding to each scale level; Indicates all shapes as The set of real three-dimensional tensors; Indicates the first The spatial parameter generator used at each scale has a structural feature consisting of 3×3 convolutions and 1×1 convolutions, with the input being... The output is split to obtain a spatially adaptive scaling factor. Translation coefficients that adapt to space ; global category semantics Concatenate to form a global context vector : and through the channel parameter generator Generate AdaLN modulation parameters: ,in, This indicates that the concatenation operator is used to concatenate multiple tokens; Indicates the first The channel parameter generator used at each scale has the following input: The output is split to obtain the parameters used for scaling. and parameters used for translation ,and Broadcasting by channel in the spatial dimension to act upon The feature map.

5. The method for super-resolution reconstruction of medical images based on anatomical semantic guidance according to claim 4, characterized in that, The modulation parameters are imported into the semantic denoising engine (SDE), and using the UNet / ResShift denoising backbone as the main body, spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ,include: In the semantic denoising engine SDE, Injected into the spatial modulation module (DSM) to provide location-dependent local structure guidance, Injected into the Global Channel Modulation Module (GAC) to provide global semantic constraints, and repeatedly injected at multiple scale levels to achieve the effect of noise residuals Anatomical structure perception denoising prediction, and after High-frequency residuals are obtained after iterative denoising. ; right Upsampling was performed to obtain the reconstructed base. ; At each time step Denoising network The ResShift / UNet denoising backbone is used as the main body, and a hybrid modulation module is used to replace the standard residual block. Spatial modulation parameters for FiLM and global modulation parameters for AdaLN are injected at the feature level. Let the first Layer input features are Hybrid modulation is calculated in the order of global calibration first, followed by local constraints: Then apply FiLM spatial modulation: ,in This represents element-wise multiplication; The output residual with semantic modulation is: ; Under the constraints of the above multi-level modulation, the semantically guided single-step denoising process can be described as follows: Specifically, it means network. In soft segmentation probability map With global category semantics Under the combined conditions, from the noise state Gradually restore a clearer residual state. and after The final residual is obtained through iterative denoising. .

6. The method for super-resolution reconstruction of medical images based on anatomical semantic guidance according to claim 5, characterized in that, When obtaining super-resolution images After that, it also includes: When the overall high-resolution slice size is large, Cut into several overlapping sub-blocks ,in This indicates the total number of overlapping sub-blocks. Indicates the first overlapping sub-blocks and ; The final residual is obtained by performing the operation on several overlapping sub-blocks separately. Steps to obtain residual prediction ; Residual prediction The results are fused with the corresponding upsampled basis to generate local super-resolution results: ; The overlapping regions of several overlapping sub-blocks are normalized and weighted by a smooth weight window to reorganize them, thereby reducing memory usage, mitigating block boundary artifacts, and improving spatial consistency: Let the pixel set of the nth sub-block in the overall image coordinate system be... And define a smooth weight window for it. The entire reconstruction result is obtained by normalized weighted summation: Where Ω represents the pixel domain of the entire image. , To prevent division by zero of extremely small constants.

7. A medical image super-resolution reconstruction device based on anatomical semantic guidance, characterized in that, include: The acquisition module is used to acquire low-resolution MRI images. To build an anatomical semantic extractor ASE from Semantic information is extracted from the data, and the semantic information includes at least a soft segmentation probability map for characterizing the distribution of local anatomical structures. and global category semantics used to characterize the organizational composition of the entire image. ; The first import module is used to import the semantic information into the dual modulation parameter generator DMPG to map the semantic information into modulation parameters, wherein the modulation parameters include at least the modulation parameters for spatial adaptive modulation FiLM and global channel modulation AdaLN. The second import module is used to import the modulation parameters into the semantic denoising engine (SDE). Using the UNet / ResShift denoising backbone as the main body, it injects spatial modulation parameters for FiLM and global modulation parameters for AdaLN during the diffusion inversion process to perform semantically guided denoising and obtain the final residual. ; Output module, used to convert Upsampling With the final residual Fusion to obtain super-resolution images : .

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it enables the electronic device to implement the anatomical semantic-guided medical image super-resolution reconstruction method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that, when run on a computer, causes the computer to perform the anatomically semantically guided medical image super-resolution reconstruction method as described in any one of claims 1 to 7.

Images