A human abdominal fat analysis method based on medical images

Abstract

The present invention relates to the technical field of medical image processing and discloses a method for analyzing abdominal fat in the human body based on medical images. Multimodal medical image data such as CT, MRI, and ultrasound images are acquired, and a fusion database is constructed through multi-scale affine transformation alignment. A model based on a residual autoencoder is used for denoising and Canny edge detection is combined to enhance the boundaries. A segmentation model is constructed based on an improved U-Net architecture, multimodal features are integrated, and optimization is performed using a dynamic convolution kernel and a channel attention mechanism. A quantitative analysis model is established using a multi-task joint learning framework to resolve gradient conflicts and search and generate lightweight sub-networks. The uncertainty of the segmentation model is modeled, and active learning annotation is performed to optimize performance. A multi-granularity feature fusion framework is designed, anatomical prior knowledge is embedded to construct an association graph, and a structured analysis report is generated. The framework can accurately analyze abdominal fat and assist in the diagnosis and treatment of obesity-related diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a human abdominal fat analysis method based on medical images. Background Art

[0002] With improvements in living standards and changes in lifestyles, the incidence of obesity and related diseases is increasing year by year. Abdominal fat accumulation is a significant risk factor for chronic diseases such as cardiovascular disease and diabetes. Therefore, accurate analysis of abdominal fat is crucial for early disease prevention and intervention. Medical imaging technology, as a non-invasive detection method, plays a key role in abdominal fat analysis. However, existing medical image-based abdominal fat analysis methods have numerous limitations.

[0003] In terms of medical image data processing, multimodal medical images (such as CT, MRI, and ultrasound images) have their own advantages. However, due to different imaging principles, the data is heterogeneous and difficult to directly fuse and analyze. Traditional image alignment methods cannot fully consider multi-scale features, and the alignment accuracy is limited, resulting in poor quality of the fused image, affecting the accuracy of subsequent analysis results. Moreover, medical images are easily interfered with by noise and artifacts during acquisition, transmission, and storage. These noise and artifacts blur the boundaries of anatomical structures, reduce the clarity and readability of the image, and make it difficult for doctors to accurately identify fat tissue and other organs. Current denoising methods often lose some useful image details while removing noise, and cannot effectively repair damaged anatomical structures.

[0004] From the perspective of fat analysis models, the accuracy of existing abdominal fat semantic segmentation models needs to be improved. Traditional convolutional neural network architectures struggle to simultaneously account for local details and global context when processing complex medical images. Their fixed convolution kernel parameters cannot adaptively adjust based on image content, resulting in inaccurate segmentation of fat region boundaries. Furthermore, the lack of effective feature selection and weight allocation mechanisms makes the model susceptible to interference from irrelevant features, affecting segmentation performance.

[0005] Existing methods for fat quantification often independently perform tasks such as fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction, failing to fully consider the inherent connections between these tasks and failing to achieve multi-task collaborative optimization. Medical images collected by different devices exhibit imaging differences, and existing methods struggle to effectively adapt to these variations, resulting in poor model generalization across data from different devices.

[0006] Furthermore, existing methods lack effective methods for evaluating and addressing segmentation model uncertainty and analysis results, making it difficult to accurately assess the reliability of analysis results. Furthermore, when generating fat analysis reports, they often only provide a simple data description and lack in-depth analysis of the relationship between fat distribution and organ topology, making them inadequate for clinical diagnosis and medical research. Therefore, the development of a more accurate, efficient, and comprehensive medical image-based method for abdominal fat analysis is urgently needed. Summary of the Invention

[0007] The purpose of the present invention is to provide a human abdominal fat analysis method based on medical images to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for analyzing abdominal fat of a human body based on medical images, the method comprising:

[0009] Step 1: Acquire multimodal medical image data, including CT, MRI, and ultrasound images. Align the heterogeneous modal data through multi-scale affine transformation to construct a multimodal fusion database of abdominal fat. For image noise or artifacts, a denoising model based on a residual autoencoder is used to repair them, and the Canny edge detection algorithm is combined to enhance anatomical structure boundaries.

[0010] Step 2: Build an abdominal fat semantic segmentation model based on an improved U-Net architecture. This model uses image superpixels as the basic unit, fuses multimodal features as input channels, and integrates local details with global contextual information through cross-scale skip connections. Dynamic convolution kernels are used to replace traditional convolutional layers, and a channel-wise attention mechanism is combined to optimize feature channel weight allocation, outputting pixel-level fat region masks.

[0011] Step 3: A multi-task joint learning framework was used to establish a fat quantification analysis model. This model defined fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction as parallel tasks. Task weights were dynamically adjusted using a gradient conflict resolution algorithm. A lightweight subnetwork was automatically generated through neural architecture search, and domain adaptation techniques were used to accommodate cross-device imaging differences.

[0012] Step 4: Model the uncertainty of the segmentation model based on Gaussian process regression, screen high-confidence areas and construct a probability distribution map; actively learn and annotate low-confidence areas through asynchronous distributed sampling, and iteratively optimize the model generalization performance using an entropy minimization strategy;

[0013] Step 5: Design a multi-granularity feature fusion framework, embed anatomical prior knowledge into the graph convolutional network, and construct a topological association map of fat distribution and organs; use an adaptive pooling layer to extract multi-scale spatial features, and constrain feature discriminability through contrastive learning loss to generate a structured fat analysis report.

[0014] Preferably, the encoder of the residual autoencoder in step 1 uses dilated convolution to expand the receptive field, and the decoder introduces deformable convolution to adapt to anatomical deformation; the denoised data is shared among multiple medical institutions through a federated learning framework, and homomorphic encryption technology is used to protect patient privacy.

[0015] Preferably, in step 2, the dynamic convolution kernel is generated by a hypernetwork, and the convolution parameters are adaptively adjusted according to the content of the input image; the channel attention mechanism adopts a squeeze-excitation structure, and generates a channel weight vector through global average pooling and a fully connected layer.

[0016] Preferably, the gradient conflict resolution algorithm in step 3 adopts the projected gradient descent method to project the multi-task gradients to the orthogonal direction of the shared subspace; the neural architecture search is based on the differentiable architecture parameterization method, and the network structure and weights are jointly optimized through the continuous relaxation technology.

[0017] Preferably, in step 4, the Gaussian process regression adopts a spectral mixing kernel function to model spatial correlation, and the active learning annotation prioritizes samples near the decision boundary based on the boundary sampling strategy; the entropy minimization strategy predicts uncertainty through Monte Carlo Dropout estimation.

[0018] Preferably, in step 5, the anatomical prior knowledge is embedded using a knowledge graph representation, and the association graph is constructed by introducing a graph attention mechanism to model the energy transfer relationship between organs; the contrastive learning loss adopts a difficult example mining strategy to enhance the diversity of negative samples.

[0019] Preferably, the federated learning framework designs a gradient sparsification mechanism, where each node uploads the Top-K significant gradients to the aggregation server, and the server generates a global model update by injecting differential privacy noise.

[0020] Preferably, the super network uses a lightweight Transformer structure to generate dynamic convolution parameters, and after the input image is divided into blocks and encoded into sequences, the long-range dependencies are captured through a self-attention mechanism.

[0021] Preferably, in step 1, the multi-scale affine transformation adopts a method combining bilinear interpolation and thin plate spline, and the alignment error is evaluated online by the normalized mutual information index; the Canny edge detection parameters are dynamically optimized by the particle swarm algorithm.

[0022] Preferably, the spectral mixing kernel function is quickly calculated through Fourier eigenmapping, and the Monte Carlo Dropout sampling adopts an adaptive Markov chain Monte Carlo method to accelerate convergence.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] This method aligns heterogeneous modal data such as CT, MRI, and ultrasound images through multi-scale affine transformation, combining bilinear interpolation with thin-plate spline methods to effectively improve alignment accuracy. Alignment error is evaluated online using the normalized mutual information metric, ensuring the accuracy of data fusion. Furthermore, based on a denoising model based on a residual autoencoder, the encoder uses dilated convolution to expand the receptive field, while the decoder introduces deformable convolution to adapt to anatomical deformations. While removing noise and artifacts, this method maximizes the preservation of image details and enhances anatomical structure boundaries, providing a high-quality data foundation for subsequent precise analysis.

[0025] This abdominal fat semantic segmentation model, built on an improved U-Net architecture, uses image superpixels as basic units, integrates multimodal features as input channels, and effectively blends local details with global contextual information through cross-scale skip connections. Dynamic convolution kernels, generated through a supernetwork, adaptively adjust convolution parameters based on the input image content. Combined with a channel-wise attention mechanism employing a squeeze-excitation architecture to optimize feature channel weight allocation, this significantly improves fat region segmentation accuracy and produces more accurate pixel-level fat region masks.

[0026] A fat quantification analysis model, built using a multi-task joint learning framework, defines fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction as parallel tasks. A gradient conflict resolution algorithm based on projected gradient descent dynamically adjusts task weights, enabling collaborative optimization of multiple tasks. Neural architecture search, based on a differentiable architecture parameterization approach, automatically generates lightweight subnetworks. Combined with domain adaptation techniques to accommodate cross-device imaging differences, this improves the model's computational efficiency and generalization capabilities, enabling more accurate quantification of abdominal fat.

[0027] Uncertainty in the segmentation model is modeled using Gaussian process regression, spatial correlation is modeled using a spectral mixing kernel function, and uncertainty is predicted using Monte Carlo Dropout estimation. High-confidence regions are screened to construct a probability distribution map, and low-confidence regions are actively labeled using asynchronous distributed sampling. A boundary sampling strategy prioritizes samples near the decision boundary, and an entropy minimization strategy is used to iteratively optimize the model's generalization performance, improving its reliability and stability.

[0028] The designed multi-granularity feature fusion framework embeds anatomical prior knowledge into a graph convolutional network, uses a knowledge graph to represent anatomical prior knowledge, and introduces a graph attention mechanism to construct a topological association map of fat distribution and organs, enabling in-depth analysis of the relationship between fat and organs. Adaptive pooling layers are used to extract multi-scale spatial features, and a hard example mining strategy is employed through contrastive learning loss to enhance the diversity of negative samples. The resulting structured fat analysis report is rich and accurate, providing doctors with more valuable diagnostic information and facilitating the prevention, diagnosis, and treatment of obesity-related diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a diagram showing the working principle of the human abdominal fat analysis method of the present invention;

[0030] Figure 2 Schematic diagram of the working principle of multi-task learning and network optimization;

[0031] Figure 3 Workflow diagram for uncertainty modeling and active learning;

[0032] Figure 4 Workflow diagram for gradient processing in the federated learning framework. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] See also Figure 1-4 The present invention provides a method for analyzing abdominal fat of a human body based on medical images, and its overall implementation scheme is as follows:

[0035] Step 1: Acquire multimodal medical image data, including CT, MRI, and ultrasound images. Align the heterogeneous modal data using a multiscale affine transformation to construct a multimodal fusion database of abdominal fat. A denoising model based on a residual autoencoder is used to repair image noise or artifacts, and the Canny edge detection algorithm is combined to enhance anatomical structure boundaries.

[0036] Step 2: Build an abdominal fat semantic segmentation model based on a modified U-Net architecture. Using image superpixels as the basic unit, this model fuses multimodal features as input channels, integrating local details with global contextual information through cross-scale skip connections. Dynamic convolution kernels replace traditional convolutional layers, and a channel-wise attention mechanism is incorporated to optimize feature channel weight distribution, outputting pixel-level fat region masks.

[0037] Step 3: Build a fat quantification analysis model using a multi-task joint learning framework. Define fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction as parallel tasks. Dynamically adjust task weights using a gradient conflict resolution algorithm. Automatically generate lightweight subnetworks through neural architecture search, and employ domain adaptation techniques to accommodate cross-device imaging differences.

[0038] Step 4: Model the uncertainty of the segmentation model using Gaussian process regression, screen high-confidence regions, and construct a probability distribution map. Active learning is performed to annotate low-confidence regions through asynchronous distributed sampling, and an entropy minimization strategy is used to iteratively optimize the model's generalization performance.

[0039] Step 5: Design a multi-granularity feature fusion framework, embed anatomical prior knowledge into a graph convolutional network, and construct a map of fat distribution and organ topology. Adaptive pooling layers are used to extract multi-scale spatial features, and contrastive learning loss is used to constrain feature discriminability to generate a structured fat analysis report.

[0040] The implementation of the present invention will be further described below with reference to Examples 1 to 5.

[0041] Example 1:

[0042] This embodiment mainly optimizes the image denoising and data sharing links in step 1 to improve image quality and ensure data security. When denoising the acquired CT, MRI and ultrasound images, a denoising model based on residual autoencoder is adopted. The encoder of this model uses dilated convolution to expand the receptive field. Dilated convolution can increase the receptive field of the convolution kernel without increasing the parameters and computational complexity, so that the encoder can obtain a wider range of image information. In this way, the encoder can better capture the long-distance dependencies in the image and more accurately identify the noise features in the image.

[0043] The decoder incorporates deformable convolution to adapt to anatomical deformation. The sampling point positions of deformable convolution can be adaptively adjusted based on the image content, which is crucial for the highly variable anatomical structures found in abdominal medical images. Traditional convolution uses a fixed, regular grid for sampling points. However, in abdominal images, the shapes and positions of organs and adipose tissue vary significantly. Deformable convolution can dynamically adjust the sampling point positions based on the local characteristics of the image, resulting in a more accurate restoration of anatomical structures.

[0044] The denoised data is shared among multiple medical institutions through the federated learning framework. In order to improve the efficiency of data sharing and protect patient privacy, the federated learning framework has designed a gradient sparsification mechanism. In the federated learning process, each medical institution acts as a node and uploads the Top-K significant gradients to the aggregation server. Specifically, after calculating the gradients of the local model, each node sorts the gradients and selects the K gradients with the largest absolute values to upload. Assume that the local gradient vector is The Top-K significant gradients selected after sorting are expressed as in is the i-th gradient value after sorting.

[0045] The aggregation server generates global model updates through differential privacy noise injection. Differential privacy is a technology for protecting data privacy. It adds noise to query results or calculation results, making it difficult for attackers to infer individual data from the output results. In this embodiment, after receiving the Top-K significant gradients uploaded by each node, the aggregation server adds noise ∈ that satisfies a specific distribution (such as Laplace distribution). Assume that the aggregated gradient is (N is the number of nodes), the global model update gradient after adding noise is This can not only ensure the training effect of the model, but also effectively protect patient privacy.

[0046] Example 2:

[0047] This example focuses on the key components of the improved U-Net architecture in step 2—the dynamic convolution kernels and channel attention mechanism—to improve the model's accuracy in fat region segmentation. When building the abdominal fat semantic segmentation model, the dynamic convolution kernels are generated using a hypernetwork, which adaptively adjusts convolution parameters based on the input image content. The hypernetwork uses a lightweight Transformer architecture to generate dynamic convolution parameters.

[0048] First, the input image is divided into blocks and encoded into a sequence. Assuming the input image size is H×W×C (H is the height, W is the width, and C is the number of channels), it is divided into image blocks of size h×w×C, where h and w are the dimensions of the image block. Each image block is flattened and projected into a low-dimensional vector to form a sequence x = [x1, x2, ..., x n ], n is the number of image blocks. Then, the long-range dependencies are captured by the self-attention mechanism. The self-attention mechanism calculates the association weight between each vector and other vectors, and the formula is:

[0049]

[0050] Among them, Q, K, and V are query matrix, key matrix, and value matrix, respectively. They are obtained by linear transformation of the input sequence. k is the dimension of the bond matrix. In this way, the hypernetwork is able to generate dynamic convolution parameters based on the global information of the image, so that the convolution kernel can better adapt to the characteristics of the fat area at different locations and scales.

[0051] The channel attention mechanism adopts a squeeze-excitation structure. The feature map of each channel is compressed into a scalar through global average pooling to obtain channel statistics. Assume that the input feature map is F∈R H×W×C , the global average pooling operation can be expressed as:

[0052]

[0053] Among them, zc is the global average pooling result of the cth channel. Then, the channel weight vector is generated through the fully connected layer. The fully connected layer consists of two fully connected layers, with the ReLU activation function used in the middle and the Sigmoid activation function used at the end to generate the channel weight ω c ,ω c ∈[0,1]. Finally, the channels of the original feature map are weighted by channel weights to highlight the features of important channels, suppress the features of unimportant channels, and optimize the feature channel weight distribution, thereby improving the segmentation accuracy of the fat area.

[0054] Example 3:

[0055] This example optimizes the multi-task joint learning framework and neural architecture search in step 3 to improve the performance and adaptability of the fat quantification analysis model. When establishing the fat quantification analysis model, a multi-task joint learning framework is used to define fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction as parallel tasks. Because gradient conflicts may exist between different tasks, a gradient conflict resolution algorithm is used to dynamically adjust task weights. This gradient conflict resolution algorithm uses projected gradient descent to project multi-task gradients onto orthogonal directions in a shared subspace.

[0056] Assume that the gradient vectors of multiple tasks are (m is the number of tasks), and the shared subspace is S. First, calculate the projection of all task gradients on the shared subspace The projected gradient vector is then projected to a direction orthogonal to the shared subspace to obtain the adjusted gradient vector In this way, the gradient conflicts between multiple tasks can be effectively alleviated, so that the model can achieve better training results on different tasks.

[0057] Neural architecture search is based on the differentiable architecture parameterization method, which jointly optimizes the network structure and weights through continuous relaxation technology. The differentiable architecture parameterization method represents the network architecture as differentiable parameters, so that optimization algorithms such as gradient descent can be used to search for the optimal network structure. During the search process, the discrete architecture search space is converted into a continuous space through continuous relaxation technology. For example, discrete operations such as the selection of network layers and the size of convolution kernels are represented as continuous parameters. Assuming that the network architecture parameter is α, by simultaneously derivatizing the objective function (such as the loss function) with respect to α and the network weight θ, that is, and The gradient descent method is used to update α and θ, thereby achieving joint optimization of network structure and weights, automatically generating lightweight sub-networks, and improving the computational efficiency and generalization ability of the model.

[0058] Example 4:

[0059] This example focuses on optimizing the uncertainty modeling and active learning annotation of the segmentation model in step 4 to improve the model's generalization performance. When modeling the uncertainty of the segmentation model based on Gaussian process regression, the Gaussian process regression uses a spectral mixing kernel function to model spatial correlation. The spectral mixing kernel function achieves fast computation through Fourier eigenmaps.

[0060] Assume that the input data is x∈R d (d is the data dimension), the spectral mixing kernel function k(x,x ′ ) can be expressed as:

[0061]

[0062] Here, S(ω) is the power spectral density function. Fourier eigenmapping transforms kernel function calculations in high-dimensional space into inner product calculations in low-dimensional space, significantly improving computational efficiency. When constructing the probability distribution map, Gaussian process regression is used to calculate the probability of each pixel belonging to the fat region, screening for high-confidence areas.

[0063] Active learning annotation is performed for low-confidence regions through asynchronous distributed sampling. Active learning annotation prioritizes samples near the decision boundary based on a boundary sampling strategy. Decision boundaries are areas where the model's classification results are uncertain. Annotating these areas can provide more information, helping the model better learn the boundaries between different categories.

[0064] The entropy minimization strategy is used to iteratively optimize the model's generalization performance. The entropy minimization strategy uses Monte Carlo Dropout to estimate prediction uncertainty. Monte Carlo Dropout sampling uses an adaptive Markov chain Monte Carlo method to accelerate convergence. During model inference, the Dropout layer is randomly enabled multiple times to obtain multiple prediction results. Assume that T samplings are performed and the prediction results are y1,y2,…,y T . Calculate the entropy H(y) of the prediction result:

[0065]

[0066] Where C is the number of categories and p(y=i) is the probability of predicting the i-th category. By selecting samples with large entropy for labeling, the prediction uncertainty of the model can be reduced and the generalization performance of the model can be iteratively optimized.

[0067] Example 5:

[0068] This example primarily optimizes the multi-granular feature fusion framework and the construction of a topological association graph between fat distribution and organs in step 5 to generate a more valuable structured fat analysis report. When designing the multi-granular feature fusion framework, anatomical prior knowledge is embedded in a graph convolutional network. This prior knowledge is represented using a knowledge graph that contains information about abdominal organs, adipose tissue, and the relationships between them. By mapping the nodes and edges in this knowledge graph to the graph convolutional network, prior information is provided to the model.

[0069] When constructing the fat distribution and organ topology association graph, the graph attention mechanism is introduced to model the energy transfer relationship between organs. The graph attention mechanism determines the importance of each node to other nodes by calculating the attention weights between nodes. Assume that the node in the graph is v i and v j , their eigenvectors are h i and h j , attention coefficient α ij The calculation formula is:

[0070]

[0071] Among them, W is the weight matrix, a is the attention vector, is node v i In this way, the topological relationship between fat distribution and organs can be captured more accurately.

[0072] Adaptive pooling layer is used to extract multi-scale spatial features. The adaptive pooling layer can automatically adjust the size of the pooling window according to the size of the input feature map to extract features of different scales. When generating structured fat analysis reports, contrastive learning loss uses a difficult example mining strategy to enhance the diversity of negative samples. The difficult example mining strategy improves the model's discriminative ability by selecting negative samples that are significantly different from positive samples for training. Assume that the positive sample is x + , the negative sample set is {x -}, during the training process, choose + Negative samples that are far away in the feature space By contrastive learning loss function L contrast Constraining feature discriminability enables the model to better distinguish different fat areas and organs, generating more accurate and structured fat analysis reports.

[0073] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0074] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing abdominal fat in a human body based on medical images, characterized in that: The following steps are involved: Step 1: Acquire multimodal medical image data, including CT, MRI, and ultrasound images. Align the heterogeneous modal data through multi-scale affine transformation to construct a multimodal fusion database of abdominal fat. For image noise or artifacts, a denoising model based on a residual autoencoder is used to repair them, and the Canny edge detection algorithm is combined to enhance anatomical structure boundaries. Step 2: Build an abdominal fat semantic segmentation model based on an improved U-Net architecture. This model uses image superpixels as the basic unit, fuses multimodal features as input channels, and integrates local details with global contextual information through cross-scale skip connections. Dynamic convolution kernels are used to replace traditional convolutional layers, and a channel-wise attention mechanism is combined to optimize feature channel weight allocation, outputting pixel-level fat region masks. Step 3: A multi-task joint learning framework was used to establish a fat quantification analysis model. This model defined fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction as parallel tasks. Task weights were dynamically adjusted using a gradient conflict resolution algorithm. A lightweight subnetwork was automatically generated through neural architecture search, and domain adaptation techniques were used to accommodate cross-device imaging differences. Step 4: Model the uncertainty of the segmentation model based on Gaussian process regression, screen high-confidence areas and construct a probability distribution map; actively learn and annotate low-confidence areas through asynchronous distributed sampling, and iteratively optimize the model generalization performance using an entropy minimization strategy; Step 5: Design a multi-granularity feature fusion framework, embed anatomical prior knowledge into the graph convolutional network, and construct a topological association map of fat distribution and organs; use an adaptive pooling layer to extract multi-scale spatial features, and constrain feature discriminability through contrastive learning loss to generate a structured fat analysis report.

2. The method for analyzing abdominal fat of a human body based on medical images according to claim 1, characterized in that: In step 1, the encoder of the residual autoencoder uses dilated convolution to expand the receptive field, and the decoder introduces deformable convolution to adapt to anatomical deformation; The denoised data is shared among multiple medical institutions through a federated learning framework, and homomorphic encryption technology is used to protect patient privacy.

3. The method for analyzing abdominal fat of a human body based on medical images according to claim 1, wherein: In step 2, the dynamic convolution kernel is generated by a hypernetwork, and the convolution parameters are adaptively adjusted according to the input image content; the channel attention mechanism adopts a squeeze-excitation structure, and generates a channel weight vector through global average pooling and a fully connected layer.

4. The method for analyzing abdominal fat of a human body based on medical images according to claim 1, wherein: In step 3, the gradient conflict resolution algorithm adopts the projected gradient descent method to project the multi-task gradients to the orthogonal direction of the shared subspace; the neural architecture search is based on the differentiable architecture parameterization method, and jointly optimizes the network structure and weights through the continuous relaxation technology.

5. The method for analyzing abdominal fat of a human body based on medical images according to claim 1, wherein: In step 4, Gaussian process regression uses a spectral mixing kernel function to model spatial correlation, and active learning annotation prioritizes samples near the decision boundary based on a boundary sampling strategy; the entropy minimization strategy uses Monte Carlo Dropout to estimate prediction uncertainty.

6. The method for analyzing abdominal fat of a human body based on medical images according to claim 1, characterized in that: In step 5, the anatomical prior knowledge is embedded using a knowledge graph representation, and the association graph is constructed by introducing a graph attention mechanism to model the energy transfer relationship between organs; the contrastive learning loss adopts a difficult example mining strategy to enhance the diversity of negative samples.

7. The method for analyzing abdominal fat of a human body based on medical images according to claim 2, wherein: The federated learning framework designs a gradient sparsification mechanism. Each node uploads the Top-K significant gradients to the aggregation server, and the server generates a global model update through differential privacy noise injection.

8. The method for analyzing abdominal fat of a human body based on medical images according to claim 3, characterized in that: The hypernetwork adopts a lightweight Transformer structure to generate dynamic convolution parameters, and the input image is divided into blocks and encoded into sequences, and the long-range dependencies are captured through the self-attention mechanism.

9. The method for analyzing abdominal fat of a human body based on medical images according to claim 1, wherein: In step 1, the multi-scale affine transformation adopts a method combining bilinear interpolation and thin plate spline, and the alignment error is evaluated online by the normalized mutual information index; the Canny edge detection parameters are dynamically optimized by the particle swarm algorithm.

10. The method for analyzing abdominal fat of a human body based on medical images according to claim 5, characterized in that: The spectral mixing kernel function is quickly calculated through Fourier feature mapping, and the Monte Carlo Dropout sampling adopts the adaptive Markov chain Monte Carlo method to accelerate convergence.

Images