Method and device for segmenting pericardial adipose tissue based on MR images, equipment and medium

Abstract

The application discloses a pericardial fat tissue segmentation method and device based on an MR image, equipment and a medium. The method comprises the following steps: acquiring a target heart magnetic resonance image of a target object containing pericardial fat tissue; segmenting the pericardial fat tissue in the target heart magnetic resonance image based on a tissue segmentation model to obtain the target pericardial fat tissue of the target object; the tissue segmentation model comprises a three-branch cross-domain feature collaborative encoder, a double attention feature fusion module and a dynamic boundary perception decoder; the three-branch cross-domain feature collaborative encoder extracts features from the global, local and frequency domains based on a three-branch structure; the double attention feature fusion module is established based on a spatial attention mechanism and a parallel channel attention mechanism; and the dynamic boundary perception decoder is used to enhance the perception ability of the global and the boundary. The scheme can precisely and efficiently segment the pericardial fat tissue in the heart MR image by using the tissue segmentation model based on multi-feature collaboration.

Description

Technical Field

[0001] This invention relates to the field of cardiac medical image segmentation technology, and in particular to a method, apparatus, device and medium for segmenting pericardial adipose tissue based on MR images. Background Technology

[0002] Pericardial adipose tissue (PEAT) is an important pathological marker in the development of cardiovascular diseases, and its accurate segmentation is crucial for early diagnosis, risk assessment, and individualized treatment planning. In terms of imaging technology, magnetic resonance (MR) imaging, with its high soft tissue contrast, has become an important tool for PEAT segmentation.

[0003] In recent years, deep learning technology has made significant progress in the field of medical image segmentation. The fully supervised learning paradigm, by training models using labeled data, has demonstrated good performance in various medical image segmentation tasks. However, significant technical bottlenecks still exist when applying it to pericardial adipose tissue segmentation.

[0004] Existing fully supervised segmentation methods based on deep learning largely rely on convolutional neural networks to capture spatial features, making it difficult to effectively integrate global contextual information with local detailed features, resulting in insufficient ability to locate blurred PEAT boundaries. Furthermore, traditional network structures lack targeted optimization for the complex noise and motion artifacts in cardiac MR images, leading to poor model robustness. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for pericardial adipose tissue segmentation based on MR images. It can accurately and efficiently segment pericardial adipose tissue in cardiac MR images using a tissue segmentation model based on multi-feature collaboration, thereby improving the robustness and reliability of tissue segmentation.

[0006] According to one aspect of the present invention, a method for pericardial adipose tissue segmentation based on MR images is provided, the method comprising:

[0007] Acquire a target cardiac magnetic resonance image of the target object, wherein the target cardiac magnetic resonance image contains pericardial adipose tissue;

[0008] The pericardial adipose tissue in the target cardiac magnetic resonance image is segmented based on a pre-trained tissue segmentation model to obtain the target pericardial adipose tissue of the target object.

[0009] The tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual-attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual-attention feature fusion module is established based on spatial attention and parallel channel attention mechanisms. The dynamic boundary-aware decoder is used to enhance the perception of the global and boundary information.

[0010] According to another aspect of the present invention, a pericardial adipose tissue segmentation device based on MR images is provided, the device comprising:

[0011] The image acquisition module is used to acquire a target cardiac magnetic resonance image of the target object, wherein the target cardiac magnetic resonance image contains pericardial adipose tissue;

[0012] The image segmentation module is used to segment the pericardial adipose tissue in the target cardiac magnetic resonance image based on a pre-trained tissue segmentation model, so as to obtain the target pericardial adipose tissue of the target object.

[0013] The tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual-attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual-attention feature fusion module is established based on spatial attention and parallel channel attention mechanisms. The dynamic boundary-aware decoder is used to enhance the perception of the global and boundary information.

[0014] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0015] At least one processor; and,

[0016] A memory communicatively connected to the at least one processor; wherein,

[0017] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the pericardial fat tissue segmentation method based on MR images according to any embodiment of the present invention.

[0018] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the pericardial adipose tissue segmentation method based on MR images according to any embodiment of the present invention.

[0019] The technical solution of this invention first acquires a target cardiac magnetic resonance image, which contains pericardial adipose tissue. Then, based on a pre-trained tissue segmentation model, the pericardial adipose tissue in the target cardiac magnetic resonance image is segmented to obtain the target pericardial adipose tissue of the target object. The tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual-attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual-attention feature fusion module is established based on spatial attention and parallel channel attention mechanisms. The dynamic boundary-aware decoder enhances the perception of the global and boundary information. This technical solution, through parallel feature extraction by the three-branch cross-domain feature co-encoder, efficient feature fusion through the dual-attention feature fusion module, and improved boundary clarity and completeness of the segmentation results by the dynamic boundary-aware decoder, enables accurate and efficient segmentation of pericardial adipose tissue in cardiac MR images using a multi-feature collaborative tissue segmentation model, improving the robustness and reliability of pericardial adipose tissue segmentation.

[0020] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for segmenting pericardial adipose tissue based on MR images according to Embodiment 1 of the present invention;

[0023] Figure 2 This is an overall structural diagram of a tissue segmentation model provided in Embodiment 1 of the present invention;

[0024] Figure 3 This is a structural diagram of a Mamba branch provided in Embodiment 1 of the present invention;

[0025] Figure 4 This is a structural diagram of a dual attention feature fusion module provided in Embodiment 1 of the present invention;

[0026] Figure 5 This is a structural diagram of a dynamic boundary-aware decoder provided in Embodiment 1 of the present invention;

[0027] Figure 6 This is a flowchart of a method for segmenting pericardial adipose tissue based on MR images according to Embodiment 2 of the present invention;

[0028] Figure 7 This is a schematic diagram of a pericardial adipose tissue segmentation device based on MR images provided in Embodiment 3 of the present invention;

[0029] Figure 8 This is a schematic diagram of the structure of an electronic device that implements a pericardial adipose tissue segmentation method based on MR images according to an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," "target," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] Example 1

[0033] Figure 1 This is a flowchart of a method for segmenting pericardial adipose tissue based on MR images, provided in Embodiment 1 of the present invention. This embodiment is applicable to the precise and efficient segmentation of pericardial adipose tissue in cardiac MR images. The method can be executed by a pericardial adipose tissue segmentation device based on MR images, which can be implemented in hardware and / or software. This device can be configured in an electronic device with data processing capabilities. Figure 1 As shown, the method includes:

[0034] S110, acquire the target cardiac magnetic resonance image of the target object, which contains pericardial adipose tissue.

[0035] In this context, a cardiac magnetic resonance image can refer to an image generated by imaging the heart using magnetic resonance imaging (MRI) technology. A target cardiac magnetic resonance image can refer to a cardiac magnetic resonance image that requires segmentation of pericardial fat tissue. The target object can refer to the object to which the target cardiac magnetic resonance image belongs, i.e., the object from which the target cardiac magnetic resonance image originates.

[0036] In this embodiment, it is first necessary to obtain one or more cardiac magnetic resonance images of the target object as target cardiac magnetic resonance images, and it is necessary to ensure that the target cardiac magnetic resonance images contain pericardial fat tissue so that the corresponding pericardial fat tissue can be identified from the target cardiac magnetic resonance images in the future.

[0037] S120, based on a pre-trained tissue segmentation model, segments the pericardial adipose tissue in the target cardiac magnetic resonance image to obtain the target pericardial adipose tissue of the target object; wherein, the tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual attention feature fusion module is established based on spatial attention mechanism and parallel channel attention mechanism. The dynamic boundary-aware decoder is used to enhance the perception ability of the global and boundary.

[0038] Here, the tissue segmentation model can refer to a deep learning model capable of identifying and segmenting pericardial adipose tissue from cardiac magnetic resonance images. The target pericardial adipose tissue can refer to the pericardial adipose tissue region identified from a target cardiac magnetic resonance image using the tissue segmentation model.

[0039] The three-branch cross-domain feature coencoder employs a three-branch structure, with each branch responsible for extracting global, local, and frequency domain features, respectively. By integrating multi-scale information, it can more effectively capture the diverse features and complex structures of pericardial adipose tissue. Optionally, the three-branch structure includes a Mamba branch, a CNN branch, and a frequency domain branch. The Mamba branch is used for global feature extraction, the CNN branch for local feature extraction, and the frequency domain branch for frequency feature extraction. Through the collaborative work of these three branches, the model can comprehensively consider global and local information at different scales and spatial levels, effectively overcoming the shortcomings of existing methods in handling irregular boundaries, multi-scale features, and regions with similar gray values, thereby more accurately identifying the boundaries and morphology of pericardial adipose tissue.

[0040] Among them, CNN (Convolutional Neural Network) is a convolutional neural network; Mamba stands for State Space Model with Selective Scan Mechanism, which is often referred to as Mamba or Mamba Architecture in the field of deep learning. It is a new type of sequence modeling architecture based on the State Space Model (SSM) and introducing a selective scanning mechanism.

[0041] In this embodiment, optionally, the Mamba branch includes a Mamba module, a scanning module, and a residual connection structure. The Mamba module is used to handle global spatial dependencies, the scanning module is used to extract local features, and the residual connection structure is used to fuse the outputs of the Mamba module and the scanning module.

[0042] Figure 2 This is an overall structural diagram of a tissue segmentation model provided in Embodiment 1 of the present invention. Figure 2 As shown, the Mamba branch employs a dynamic scanning Mamba mode, integrating lightweight convolutional scanning operations into the Mamba module, enhancing its ability to model spatial features while maintaining high computational efficiency. Specifically, the Mamba branch adopts a two-branch structure, with core components including: a Mamba module for handling global spatial dependencies, a scanning module for lightweight local feature extraction, and a residual connection structure that fuses the outputs of the two branch modules.

[0043] Figure 3 This is a structural diagram of a Mamba branch provided in Embodiment 1 of the present invention. Figure 3 As shown, the input image is fed into both the Mamba module and the scanning module. The Mamba module transposes the image tensor to adapt to the model's processing paradigm, normalizes it, and then feeds it into the Mamba model for feature extraction. Finally, it reverses the tensor to restore the output to the original image dimensions. The scanning module uses depthwise separable convolution to extract more detailed local features and applies the GELU activation function to enhance non-linear expressive power. Finally, the features extracted by the two modules in the Mamba branch are fused to fully utilize the advantages of both, thereby improving the overall performance and accuracy of the model.

[0044] In this embodiment, optionally, the CNN branch includes two convolutional layers, one deformable convolutional layer, and one max pooling layer, and the frequency domain branch includes a frequency domain filter.

[0045] like Figure 2As shown, the deformable CNN branch is a sequential single-branch structure. It utilizes deformable convolutions to achieve adaptive receptive field adjustment, enhancing the feature extraction capability for irregular fat regions. Specifically, it first undergoes a 3×3 convolution operation through a regular convolutional layer, maintaining the same number of channels for both input and output, with padding set to 1 to ensure spatial alignment between the output and input. Next, the input passes through a deformable convolutional layer, as shown in the following formula: ,in, It is the output feature map at the location The value at that location, These are the convolution kernel weights. It is the input feature map at the location The value at that location, It is a learnable offset. This refers to the bias term. Deformable convolution is an extension of the standard convolution operation, allowing the kernel position to be dynamically adjusted based on the input feature map. Specifically, it is implemented by first calculating the offset through an additional offset convolution. The number of output channels for the offset is set to twice the square of the convolution kernel size. This way, the positional information of each convolution kernel is passed through the offset, guiding the weighted computation of the convolution operation in the image. This dynamic adjustment of the offset allows the convolution kernels to adaptively adjust their position, thus better capturing deformations and irregular structures in the image. The output feature map is then generated through standard convolutional layers and non-linearly transformed using the GELU (Gaussian Error Linear Unit) activation function. Finally, max pooling is used to halve the spatial size to reduce computation, and the number of channels is doubled through convolutional layers to continue extracting high-level features from the image, thereby enhancing the model's expressive power.

[0046] like Figure 2 As shown, the frequency domain branch includes convolutional layers and learnable frequency domain filters. For example, the frequency domain filter can be a Fourier filter. The frequency domain branch introduces a learnable Fourier filter, which helps the model better focus on important features of pericardial adipose tissue by enhancing specific frequency components in the image. Furthermore, the Fourier filter can effectively reduce noise and artifact interference without compromising important image structures, thereby improving the segmentation quality of pericardial adipose tissue.

[0047] Specifically, taking a Fourier filter as an example, the processing flow of a learnable frequency domain filter is as follows: First, perform a Fast Fourier Transform on the input image to convert the two-dimensional image from the time domain to the frequency domain. Then, resize the learnable filter to match the frequency components of the input image. Multiply the resized filter by the corresponding frequency components of the input image in the frequency domain to perform frequency domain filtering. Perform an Inverse Fourier Transform on the filtered frequency domain data to restore it to the time domain, obtaining the processed image.

[0048] The specific process of Fourier transform and filtering is as follows: Assuming the input image... It is a multi-channel two-dimensional image, in which, , and These represent the input images respectively. The number of channels, height, and width are determined. A frequency domain representation is obtained after a two-dimensional Fourier transform. See formula .in, It is a frequency domain index. This refers to the pixel values ​​of the input image in the spatial domain. For a learnable Fourier filter... This method aims to enhance the features of an input image through frequency domain filtering. The filter is adjusted through interpolation to fit the frequency domain size of the input image. The interpolated filter and the input frequency domain representation are then compared. The frequency domain filtering result is obtained by performing element-wise multiplication, as shown in the following formula: .in, This is the result of frequency domain filtering. This represents element-wise multiplication. The filter is resized to have the same dimensions as the frequency domain representation of the input image. Then, use the inverse Fourier transform to return to the time domain, see the formula. .in, This represents the pixels of the filtered temporal image. Ultimately, it has a structure similar to the input image in the temporal domain, but specific frequency components are further enhanced through frequency domain operations, improving the model's ability to perceive details.

[0049] The three-branch cross-domain feature co-encoder proposed in this invention overcomes the limitations of existing single-domain modeling by innovatively employing a parallel feature extraction structure combining Mamba, CNN, and frequency domain features. Unlike existing methods that rely on a single feature extraction approach, this encoder, through the collaborative extraction of global, local, and frequency domain features, can more comprehensively capture information from pericardial adipose tissue at different scales and spatial levels. This effectively addresses the challenges of existing methods in handling irregular boundaries, multi-scale features, and regions with similar grayscale values, significantly improving the accuracy and comprehensiveness of feature extraction.

[0050] In this embodiment, a dual-attention feature fusion module is used to independently evaluate the importance of features in each branch of the three-branch structure through a parallel channel attention mechanism, dynamically adjusting the contribution of each branch to ensure that features in key channels receive higher attention. Simultaneously, a spatial attention mechanism is used to generate pixel-level weight maps, enhancing the model's ability to capture key regions and optimizing feature fusion. Furthermore, residual connections effectively alleviate the gradient vanishing problem, ensuring stable gradient propagation and thus improving the network's convergence speed and overall performance. The dual-attention feature fusion module differs fundamentally from existing feature fusion methods, effectively overcoming the shortcomings of existing fusion methods such as insufficient attention to key features and gradient vanishing, achieving efficient feature fusion and significantly improving network convergence speed and overall performance.

[0051] Specifically, such as Figure 4 As shown, each branch in the three-branch structure calculates the channel attention weight and spatial attention weight, as illustrated in the following formula:

[0052] ;

[0053] .

[0054] in, and These represent the channel attention weights and spatial attention weights, respectively. As input features, It is the Sigmoid activation function. Indicates average pooling. Indicates splicing Indicates adoption Convolution operations are performed using convolution kernels of varying sizes. Specifically, channel attention mechanisms can be used to process each input feature independently, calculating the importance weight of each channel. Spatial attention mechanisms can be used to process concatenated multi-source features, calculating the importance weight of each spatial location. Applying channel attention weights and spatial attention weights to the input features respectively yields weighted features. See the following formula: .

[0055] To preserve information from the original features, the dual-attention feature fusion module fuses the weighted features with the concatenated original features through residual connections, as shown in the following formula:

[0056] .

[0057] in, This represents the output feature of the dual attention feature fusion module.

[0058] In this embodiment, an innovation is proposed based on existing decoders to enhance the global awareness and boundary awareness capabilities of the decoder. Optionally, the dynamic boundary awareness decoder includes multiple identical decoding modules. The decoding module includes a dynamic scanning Mamba module and a boundary refinement module. The dynamic scanning Mamba module includes a Mamba module and a scanning module. The boundary refinement module includes an dilation layer, an erosion layer, and an attention layer.

[0059] Figure 5 This is a structural diagram of a dynamic boundary-aware decoder provided in Embodiment 1 of the present invention. Figure 5 As shown, the dynamic boundary-aware decoder introduces a dynamic scanning Mamba module on top of the traditional upsampling convolution, enhancing the decoder's global perception capability. Furthermore, by combining a boundary refinement module with dilation, erosion, and attention mechanisms, the decoder's boundary-awareness capability is effectively improved. This design addresses the shortcomings of traditional decoders in boundary processing, enabling the decoder to handle pericardial adipose tissue boundaries more accurately, significantly improving the boundary clarity and integrity of the segmentation results, and effectively enhancing the overall segmentation accuracy and reliability.

[0060] Specifically, such as Figure 5 As shown, the dynamic boundary-aware decoder consists of four identical stacked decoding modules. Each iteration processes the input features and progressively refines the image segmentation results. In each decoding module, the input features are first upsampled through a transposed convolutional layer, restoring the low-resolution feature map to a higher spatial resolution to enhance image details. The upsampled feature map then passes through the same Mamba and scanning modules as the encoder, enabling fine-grained utilization of encoder features when handling irregular regions and complex boundaries, thus improving the model's boundary optimization capabilities.

[0061] Meanwhile, the dynamic boundary-aware decoder further enhances the boundary accuracy of image features through a boundary refinement module. This module consists of a dilation layer, an erosion layer, and an attention layer, all working together on the image's boundary regions. The dilation layer expands the object boundaries in the feature map, helping to capture the object's contours and external structure, enhancing boundary visibility. The erosion layer further refines the boundaries by shrinking the boundaries in the feature map, removing noise. When the outputs of the dilation and erosion layers are subtracted, the result highlights the boundary differences of the object, emphasizing the boundary between the object and the background, while weakening minor noise and irrelevant information in the background. This operation helps eliminate irrelevant small details, making the boundaries clearer and more accurate. The attention layer generates an attention map, precisely controlling the boundary refinement process, allowing the network to selectively focus on important boundary regions, refine these areas, and suppress irrelevant parts, thereby improving the quality of the segmentation results. The attention layer employs an attention mechanism, constructed based on convolution, ReLU (Rectified Linear Unit) functions, and the Sigmoid function (i.e., an S-shaped curve function). In each decoding module, residual connections are used to preserve the original features, avoid feature loss and prevent gradient vanishing, and ensure the effective transmission of information across multiple layers of the network.

[0062] In this embodiment, the tissue segmentation model needs to be pre-trained so that it can be used to segment pericardial adipose tissue in the target cardiac magnetic resonance image. Optionally, the training process of the tissue segmentation model includes: acquiring multiple candidate cardiac magnetic resonance images containing pericardial adipose tissue, determining training data based on the candidate cardiac magnetic resonance images; building an initial segmentation model using a semi-supervised training framework based on cross-teaching; and performing fully supervised training on the initial segmentation model based on a hybrid loss function and the training data to obtain the tissue segmentation model.

[0063] Specifically, when training the tissue segmentation model, multiple candidate cardiac magnetic resonance images containing pericardial adipose tissue are first acquired, and the pericardial adipose tissue in each candidate cardiac magnetic resonance image is labeled accordingly to construct the training data. The candidate cardiac magnetic resonance images can refer to the cardiac magnetic resonance images used in model training. Alternatively, existing public datasets can be directly used as training data. For example, the MRPEAT dataset can be used as training data. This dataset contains short-axis cardiac magnetic resonance images of three groups of patients, with each patient having 8 to 10 consecutive dynamic image views at end-diastole and end-systole, and each image is accompanied by corresponding annotations.

[0064] Then, a semi-supervised training framework based on cross-teaching is built as the initial segmentation model. Pre-labeled pericardial adipose tissue is used as the supervision target. A hybrid loss function is used to train and optimize the initial segmentation model based on the training data using a fully supervised learning approach. Finally, a tissue segmentation model for segmenting pericardial adipose tissue in cardiac magnetic resonance images is obtained. The hybrid loss function can be a loss function composed of at least two types of losses. Optionally, the hybrid loss function includes Dice loss, edge loss, and uncertainty loss. Dice loss, edge loss, and uncertainty loss correspond to different optimization objectives, as detailed in the following formula:

[0065] ;

[0066] ;

[0067] .

[0068] in, , and These represent Dice loss, marginal loss, and uncertainty loss, respectively. This is the probability graph output by the model. A binary image representing the true labels. As a smoothing factor, It is the Sigmoid activation function. This represents the total number of pixels involved in the calculation.

[0069] It should be noted that Dice loss addresses the class imbalance problem and is suitable for scenarios with significant differences between the foreground and background. It emphasizes the overlap between the predicted and ground truth regions, improving segmentation coverage. Edge loss amplifies edge differences through a high-slope sigmoid function and uses mean squared error to align predicted and ground truth edges, improving boundary accuracy. Uncertainty loss minimizes the cross-entropy between the predicted probability and the ground truth label, providing a global optimization signal and enhancing pixel classification confidence.

[0070] The hybrid loss function is shown in the following formula: .in, It is a mixed loss function; , and These represent the weight coefficients for Dice loss, edge loss, and uncertainty loss, respectively, used to control the contribution of each part of the loss. Hybrid loss functions, by combining loss functions with different properties, can simultaneously optimize region overlap, edge alignment, and pixel classification, thereby leveraging the advantages of each loss function.

[0071] The proposed hybrid loss function integrates Dice loss, edge loss, and uncertainty loss, and sets optimization objectives for the volume overlap, boundary localization accuracy, and prediction confidence of the segmented regions, respectively. Through a multi-task joint learning mechanism, it effectively improves the overall segmentation performance of the model.

[0072] The technical solution of this invention first acquires a target cardiac magnetic resonance image, which contains pericardial adipose tissue. Then, based on a pre-trained tissue segmentation model, the pericardial adipose tissue in the target cardiac magnetic resonance image is segmented to obtain the target pericardial adipose tissue of the target object. The tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual-attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual-attention feature fusion module is established based on spatial attention and parallel channel attention mechanisms. The dynamic boundary-aware decoder enhances the perception of the global and boundary information. This technical solution, through parallel feature extraction by the three-branch cross-domain feature co-encoder, efficient feature fusion through the dual-attention feature fusion module, and improved boundary clarity and completeness of the segmentation results by the dynamic boundary-aware decoder, enables accurate and efficient segmentation of pericardial adipose tissue in cardiac MR images using a fully supervised tissue segmentation model based on multi-feature collaboration, thus improving the reliability and robustness of pericardial adipose tissue segmentation.

[0073] In this embodiment, optionally, after acquiring multiple candidate cardiac magnetic resonance images containing pericardial adipose tissue, the method further includes: preprocessing the candidate cardiac magnetic resonance images, the preprocessing including image size adjustment; correspondingly, determining training data based on the candidate cardiac magnetic resonance images includes: determining training data based on the preprocessed candidate cardiac magnetic resonance images.

[0074] Specifically, to meet the input requirements for model training, all original image data (i.e., candidate cardiac MRI images) used in training were standardized. Through image scaling and resizing techniques, the original images were uniformly adjusted to a standard size of 256×256, which not only effectively eliminated the differences in data resolution but also laid the foundation for efficient model training and stable performance.

[0075] Example 2

[0076] Figure 6This is a flowchart of a method for segmenting pericardial adipose tissue based on MR images, provided in Embodiment 2 of the present invention. This embodiment is an optimization based on the above embodiment. Specifically, the optimization includes: after acquiring the target cardiac magnetic resonance image of the target object, the method further includes: preprocessing the target cardiac magnetic resonance image, including image size adjustment; correspondingly, segmenting the pericardial adipose tissue in the target cardiac magnetic resonance image based on a pre-trained tissue segmentation model, including: segmenting the pericardial adipose tissue in the preprocessed target cardiac magnetic resonance image based on the pre-trained tissue segmentation model.

[0077] like Figure 6 As shown, the method in this embodiment specifically includes the following steps:

[0078] S210, acquire the target cardiac magnetic resonance image of the target object, which contains pericardial adipose tissue.

[0079] S220 performs preprocessing on the target cardiac magnetic resonance image, including image resizing.

[0080] Similarly, referring to the preprocessing operations in the model training process described above, the acquired target cardiac MRI image also needs to be standardized to meet the input requirements of the model. Image scaling and resizing techniques can adjust the target cardiac MRI image to a standard size of 256×256, which helps to further improve the prediction accuracy of the tissue segmentation model. Furthermore, if other preprocessing operations were used during model training, such as image resolution adjustment and image intensity normalization, corresponding preprocessing operations are required when using the model.

[0081] S230, based on a pre-trained tissue segmentation model, the pericardial adipose tissue in the pre-processed target cardiac magnetic resonance image is segmented to obtain the target pericardial adipose tissue of the target object.

[0082] The tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual-attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual-attention feature fusion module is built based on spatial attention and parallel channel attention mechanisms. The dynamic boundary-aware decoder is used to enhance the perception of the global and boundary information.

[0083] The technical solution of this invention, after acquiring the target cardiac magnetic resonance image, further preprocesses the target cardiac magnetic resonance image. This preprocessing includes image resizing and segmenting the pericardial adipose tissue in the preprocessed target cardiac magnetic resonance image based on a pre-trained tissue segmentation model. This technical solution, by adding image preprocessing operations, can meet the input requirements of the model, helping to further improve the prediction accuracy of the tissue segmentation model.

[0084] Example 3

[0085] Figure 7 This is a schematic diagram of a pericardial adipose tissue segmentation device based on MR images provided in Embodiment 3 of the present invention. This device can execute the pericardial adipose tissue segmentation method based on MR images provided in any embodiment of the present invention, and possesses the corresponding functional modules and beneficial effects for executing the method. For example... Figure 7 As shown, the device includes:

[0086] Image acquisition module 310 is used to acquire a target cardiac magnetic resonance image of a target object, wherein the target cardiac magnetic resonance image contains pericardial adipose tissue;

[0087] Image segmentation module 320 is used to segment the pericardial adipose tissue in the target cardiac magnetic resonance image based on a pre-trained tissue segmentation model to obtain the target pericardial adipose tissue of the target object.

[0088] The tissue segmentation model includes a three-branch cross-domain feature co-encoder, a dual-attention feature fusion module, and a dynamic boundary-aware decoder. The three-branch cross-domain feature co-encoder extracts features from the global, local, and frequency domains based on a three-branch structure. The dual-attention feature fusion module is established based on spatial attention and parallel channel attention mechanisms. The dynamic boundary-aware decoder is used to enhance the perception of the global and boundary information.

[0089] Optionally, the three-branch structure includes a Mamba branch, a CNN branch, and a frequency domain branch. The Mamba branch is used for global feature extraction, the CNN branch is used for local feature extraction, and the frequency domain branch is used for frequency feature extraction.

[0090] Optionally, the Mamba branch includes a Mamba module, a scanning module, and a residual connection structure. The Mamba module is used to handle global spatial dependencies, the scanning module is used to extract local features, and the residual connection structure is used to fuse the outputs of the Mamba module and the scanning module.

[0091] Optionally, the CNN branch includes two convolutional layers, a deformable convolutional layer, and a max pooling layer, and the frequency domain branch includes a frequency domain filter.

[0092] Optionally, the dynamic boundary-aware decoder includes multiple identical decoding modules, each decoding module including a dynamic scanning Mamba module and a boundary refinement module. The dynamic scanning Mamba module includes a Mamba module and a scanning module, and the boundary refinement module includes an expansion layer, an erosion layer, and an attention layer.

[0093] Optionally, the apparatus further includes: a model training module, used for:

[0094] Acquire multiple candidate cardiac magnetic resonance images containing pericardial adipose tissue, and determine training data based on the candidate cardiac magnetic resonance images;

[0095] An initial segmentation model was built using a semi-supervised training framework based on cross-teaching.

[0096] The tissue segmentation model is obtained by performing fully supervised training on the initial segmentation model based on the hybrid loss function and the training data.

[0097] Optionally, the hybrid loss function includes Dice loss, edge loss, and uncertainty loss.

[0098] Optionally, the model training module is further configured to:

[0099] After acquiring multiple candidate cardiac magnetic resonance images containing pericardial adipose tissue, the candidate cardiac magnetic resonance images are preprocessed, including image size adjustment.

[0100] Accordingly, the model training module is used for:

[0101] Training data is determined based on preprocessed candidate cardiac magnetic resonance images.

[0102] The pericardial adipose tissue segmentation device based on MR images provided in this embodiment of the invention can execute the pericardial adipose tissue segmentation method based on MR images provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0103] Example 4

[0104] Figure 8A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0105] like Figure 8 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0106] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0107] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the pericardial adipose tissue segmentation method based on MR images.

[0108] In some embodiments, the pericardial adipose tissue segmentation method based on MR images can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the pericardial adipose tissue segmentation method based on MR images described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the pericardial adipose tissue segmentation method based on MR images by any other suitable means (e.g., by means of firmware).

[0109] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0110] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0111] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0112] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0113] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0114] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through a communication network. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0115] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0116] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method of pericardial adipose tissue segmentation based on MR images, characterized by, The method comprises: obtaining a target cardiac magnetic resonance image of a target object, the target cardiac magnetic resonance image containing pericardial adipose tissue; segmenting the pericardial adipose tissue in the target cardiac magnetic resonance image based on a pre-trained tissue segmentation model to obtain target pericardial adipose tissue of the target object; wherein the tissue segmentation model comprises a three-branch cross-domain feature collaborative encoder, a double attention feature fusion module and a dynamic boundary perception decoder, the three-branch cross-domain feature collaborative encoder extracts features from global, local and frequency domains based on a three-branch structure, the double attention feature fusion module is established based on a spatial attention mechanism and a parallel channel attention mechanism, and the dynamic boundary perception decoder is used to enhance the perception ability of the global and boundary; the three-branch structure comprises a Mamba branch, a CNN branch and a frequency domain branch, the Mamba branch is used for global feature extraction, the CNN branch is used for local feature extraction, and the frequency domain branch is used for frequency feature extraction; the double attention feature fusion module is specifically used for: calculating channel attention weights for each branch in the three-branch structure respectively, and calculating spatial attention weights for the three-branch structure, as shown in the following formula: ； ； wherein, and denote channel attention weight and spatial attention weight, respectively, is an input feature, is a Sigmoid activation function, denotes average pooling, denotes concatenation, denotes convolution operation with a convolution kernel of size size; applying the channel attention weights and the spatial attention weights to the input features respectively to obtain weighted features, as shown in the following formula: ； wherein, denotes the weighted features; fusing the weighted features with the spliced input features through a residual connection, as shown in the following formula: ； wherein, represents the output feature of the double attention feature fusion module.

2. The method of claim 1, wherein, the Mamba branch comprises a Mamba module, a scanning module and a residual connection structure, the Mamba module is used for processing global spatial dependence, the scanning module is used for local feature extraction, and the residual connection structure is used for fusing the outputs of the Mamba module and the scanning module.

3. The method of claim 1, wherein, the CNN branch comprises two convolution layers, one deformable convolution layer and one maximum pooling layer, and the frequency domain branch comprises a frequency domain filter.

4. The method of claim 1, wherein, the dynamic boundary perception decoder comprises a plurality of identical decoding modules, the decoding module comprises a dynamic scanning Mamba module and a boundary refinement module, the dynamic scanning Mamba module comprises a Mamba module and a scanning module, and the boundary refinement module comprises a dilation layer, an erosion layer and an attention layer.

5. The method according to any one of claims 1-4, characterized in that, The training process of the tissue segmentation model comprises: obtaining a plurality of candidate cardiac magnetic resonance images containing pericardial adipose tissue, and determining training data according to the candidate cardiac magnetic resonance images; building an initial segmentation model using a semi-supervised training framework based on cross-teaching; performing full-supervised training on the initial segmentation model according to a hybrid loss function and the training data to obtain a tissue segmentation model.

6. The method of claim 5, wherein, The hybrid loss function comprises a Dice loss, an edge loss and an uncertainty loss.

7. The method of claim 5, wherein, After obtaining a plurality of candidate cardiac magnetic resonance images containing pericardial adipose tissue, the method further comprises: preprocessing the candidate cardiac magnetic resonance images, the preprocessing comprising image size adjustment; correspondingly, determining training data according to the candidate cardiac magnetic resonance images comprises: determining training data according to the preprocessed candidate cardiac magnetic resonance images.

8. A pericardial adipose tissue segmentation apparatus based on MR images, characterized by, The device comprises: The image acquisition module is configured to acquire a target cardiac magnetic resonance (MR) image of a target object, the target cardiac MR image containing pericardial adipose tissue; The image segmentation module is configured to segment the pericardial adipose tissue in the target cardiac MR image based on a pre-trained tissue segmentation model to obtain target pericardial adipose tissue of the target object. The tissue segmentation model comprises a three-branch cross-domain feature collaborative encoder, a double-attention feature fusion module, and a dynamic boundary perception decoder. The three-branch cross-domain feature collaborative encoder extracts features from global, local, and frequency domains based on a three-branch structure. The double-attention feature fusion module is established based on a spatial attention mechanism and a parallel channel attention mechanism. The dynamic boundary perception decoder is used to enhance the perception ability of the global and boundary. The three-branch structure comprises a Mamba branch, a CNN branch, and a frequency domain branch. The Mamba branch is used for global feature extraction. The CNN branch is used for local feature extraction. The frequency domain branch is used for frequency feature extraction. The double-attention feature fusion module is specifically configured to: Calculate the channel attention weight for each branch in the three-branch structure, and calculate the spatial attention weight for the three-branch structure, as shown in the following formula: ； ； wherein, and denote channel attention weight and spatial attention weight, respectively, is an input feature, is a Sigmoid activation function, denotes average pooling, denotes concatenation, denotes convolution operation with a convolution kernel of size size; Apply the channel attention weight and the spatial attention weight to the input features to obtain weighted features, as shown in the following formula: ； wherein, denotes the weighted features; Fuse the weighted features with the spliced input features through a residual connection, as shown in the following formula: ； wherein, represents the output feature of the double attention feature fusion module.

9. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory connected to the at least one processor in communication; wherein The memory stores a computer program executable by the at least one processor. The computer program is executed by the at least one processor to enable the at least one processor to execute the pericardial adipose tissue segmentation method based on the MR image according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for enabling the processor to implement the pericardial adipose tissue segmentation method based on the MR image according to any one of claims 1-7 when executed.

Images