Human body abdominal fat analysis method based on medical image

Abstract

The invention relates to the technical field of medical image processing, and discloses a human body abdominal fat analysis method based on a medical image, which comprises the following steps: acquiring multi-modal medical image data such as CT (Computed Tomography), MRI (Magnetic Resonance Imaging) and ultrasonic images, constructing a fusion database through multi-scale affine transformation alignment, de-noising by using a model based on a residual self-encoder, and enhancing a boundary in combination with Canny edge detection. A segmentation model is constructed based on an improved U-Net architecture, multi-modal features are fused, and a dynamic convolution kernel and a channel attention mechanism are used for optimization. A quantitative analysis model is established by adopting a multi-task joint learning framework, gradient conflicts are solved, and a lightweight sub-network is searched and generated. Carrying out uncertainty modeling on the segmentation model, and carrying out active learning annotation to optimize the performance. A multi-granularity feature fusion framework is designed, anatomical priori knowledge is embedded to construct an association graph, a structured analysis report is generated, abdominal fat can be accurately analyzed, and diagnosis and treatment of obesity-related diseases are assisted.

Description

Technical Field

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

[0002] With the improvement of people's living standards and the change of lifestyle, the incidence of obesity and its related diseases has been increasing year by year. Abdominal fat accumulation is one of the important risk factors for chronic diseases such as cardiovascular diseases and diabetes. Therefore, accurately analyzing human abdominal fat is crucial for the early prevention and intervention of diseases. As a non-invasive detection method, medical image technology plays a key role in abdominal fat analysis. However, the existing abdominal fat analysis methods based on medical images have many limitations.

[0003] In terms of medical image data processing, multi-modal 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 and affecting the accuracy of subsequent analysis results. Moreover, medical images are easily interfered by noise and artifacts during the acquisition, transmission, and storage processes. These noise and artifacts will blur the anatomical structure boundaries, reduce the clarity and readability of the images, making it difficult for doctors to accurately identify fat tissues and other organs. Currently, the existing denoising methods often lose some useful image details while removing noise and cannot effectively repair damaged anatomical structures.

[0004] From the perspective of the fat analysis model, the accuracy of the existing abdominal fat semantic segmentation models needs to be improved. When dealing with complex medical images, the traditional convolutional neural network structure is difficult to simultaneously consider local details and global context information. Its fixed convolutional kernel parameters cannot be adaptively adjusted according to the image content, resulting in inaccurate segmentation of the fat region boundaries. Moreover, the lack of an effective feature selection and weight allocation mechanism makes the model easily interfered by irrelevant features and affects the segmentation effect.

[0005] In terms of fat quantification analysis, the existing methods usually perform tasks such as fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction independently, without fully considering the internal connections between tasks and unable to achieve collaborative optimization of multiple tasks. There are imaging differences in medical images collected by different devices, and the existing methods are difficult to effectively adapt to these differences, resulting in poor generalization ability of the model on data from different devices.

[0006] In addition, in terms of model performance optimization and result analysis, existing methods lack effective means to evaluate and handle the uncertainty of the segmentation model, and are unable to accurately judge the reliability of the analysis results. Moreover, when generating a fat analysis report, only simple data descriptions are often provided, lacking in-depth analysis of the topological relationship between fat distribution and organs, making it difficult to meet the needs of clinical diagnosis and medical research. In summary, it is urgent to develop a more accurate, efficient and comprehensive method for analyzing human abdominal fat based on medical images. Summary of the Invention

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

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

[0009] Step 1: Obtain multi-modal medical image data, including CT, MRI and ultrasound images, align heterogeneous modal data through multi-scale affine transformation, and construct a multi-modal fusion database of abdominal fat; for image noise or artifacts, use a denoising model based on a residual autoencoder for repair, and combine the Canny edge detection algorithm to enhance the anatomical structure boundary;

[0010] Step 2: Construct an abdominal fat semantic segmentation model based on an improved U-Net architecture, use image superpixels as basic units, fuse multi-modal features as input channels, and fuse local details and global context information through cross-scale skip connections; replace traditional convolutional layers with dynamic convolutional kernels, and combine channel attention mechanisms to optimize the weight distribution of feature channels, and output a pixel-level fat region mask;

[0011] Step 3: Establish 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, and dynamically adjust task weights based on a gradient conflict resolution algorithm; automatically generate a lightweight sub-network through neural architecture search, and use domain adaptation technology to adapt to cross-device imaging differences;

[0012] Step 4: Perform uncertainty modeling on the segmentation model based on Gaussian process regression, screen high-confidence regions and construct a probability distribution map; perform active learning annotation on low-confidence regions through asynchronous distributed sampling, and use an entropy minimization strategy to iteratively optimize the generalization performance of the model;

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

[0014] Preferably, 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 deformations; 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 input image content; the channel attention mechanism adopts a squeeze-and-excitation structure, and the channel weight vector is generated through global average pooling and a fully connected layer.

[0016] Preferably, in step 3, the gradient conflict resolution algorithm uses the projected gradient descent method to project the multi-task gradient onto 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 continuous relaxation techniques.

[0017] Preferably, in step 4, Gaussian process regression uses a spectral mixture kernel function to model spatial correlation, and active learning annotation preferentially selects samples near the decision boundary based on the boundary sampling strategy; the entropy minimization strategy estimates the prediction uncertainty through Monte Carlo Dropout.

[0018] Preferably, in step 5, the anatomical prior knowledge embedding uses a knowledge graph representation, and the association graph construction introduces a graph attention mechanism to model the energy transfer relationship between organs; the contrast learning loss uses a hard example mining strategy to enhance the diversity of negative samples.

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

[0020] Preferably, the hypernetwork uses a lightweight Transformer structure to generate dynamic convolution parameters, and the input image is block-encoded into a sequence and then captures long-range dependencies through the self-attention mechanism.

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

[0022] Preferably, the spectral mixture kernel function is quickly calculated through Fourier feature mapping, and the Monte Carlo Dropout sampling uses the adaptive Markov chain Monte Carlo method to accelerate convergence.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] The present invention aligns heterogeneous modality data such as CT, MRI, and ultrasound images through multi-scale affine transformation, and combines the methods of bilinear interpolation and thin plate spline, which can effectively improve the alignment accuracy. The alignment error is evaluated online using the normalized mutual information index, ensuring the accuracy of data fusion. At the same time, based on the denoising model of the residual autoencoder, the encoder uses dilated convolution to expand the receptive field, and the decoder introduces deformable convolution to adapt to anatomical deformations. While removing noise and artifacts, it maximally retains image details and enhances the anatomical structure boundaries, providing a high-quality data basis for subsequent precise analysis.

[0025] The abdominal fat semantic segmentation model constructed based on the improved U-Net architecture takes image superpixels as the basic unit, fuses multi-modal features as input channels, and realizes the effective fusion of local details and global context information through cross-scale skip connections. The dynamic convolution kernel is generated by the super network, which can adaptively adjust the convolution parameters according to the content of the input image. Combining with the channel attention mechanism using the squeeze-excitation structure to optimize the feature channel weight allocation, it greatly improves the segmentation accuracy of the fat area, and the output pixel-level fat area mask is more accurate.

[0026] The fat quantification analysis model established using the multi-task joint learning framework defines the calculation of fat volume, visceral-subcutaneous fat classification, and prediction of fat metabolism correlation as parallel tasks. The gradient conflict resolution algorithm based on the projected gradient descent method dynamically adjusts the task weights, realizing the collaborative optimization of multiple tasks. The neural architecture search is based on the differentiable architecture parameterization method, automatically generating a lightweight sub-network, and combining with the domain adaptation technology to adapt to the imaging differences across devices, improving the computational efficiency and generalization ability of the model, and being able to more accurately quantify and analyze abdominal fat.

[0027] Based on Gaussian process regression, uncertainty modeling is performed on the segmentation model. The spectral mixture kernel function is used to model the spatial correlation, and the prediction uncertainty is estimated through Monte Carlo Dropout. The high-confidence regions are screened to construct a probability distribution map, and active learning annotation is performed on the low-confidence regions through asynchronous distributed sampling. Based on the boundary sampling strategy, samples near the decision boundary are preferentially selected, and the entropy minimization strategy is used to iteratively optimize the generalization performance of the model, improving the reliability and stability of the model.

[0028] The designed multi-granularity feature fusion framework embeds anatomical prior knowledge into the graph convolutional network, represents the anatomical prior knowledge using a knowledge graph, and introduces a graph attention mechanism to construct a graph of the association between fat distribution and organ topology, which can deeply analyze the relationship between fat and organs. The adaptive pooling layer is used to extract multi-scale spatial features, and the hard example mining strategy is adopted through the contrast learning loss to enhance the diversity of negative samples. The generated structured fat analysis report is rich and accurate in content, providing more valuable diagnostic information for doctors and contributing to the prevention, diagnosis, and treatment of obesity-related diseases. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0031] Figure 3 This is the working flowchart of uncertainty modeling and active learning;

[0032] Figure 4 This is the working flowchart of gradient processing in the federated learning framework. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0034] Please refer to Figures 1-4 , the present invention provides a human abdominal fat analysis method based on medical images, and its overall implementation scheme is as follows:

[0035] Step 1: Obtain multi-modal medical image data, including CT, MRI, and ultrasound images. Align heterogeneous modal data through multi-scale affine transformation to construct an abdominal fat multi-modal fusion database. For image noise or artifacts, use a denoising model based on a residual autoencoder for repair, and combine the Canny edge detection algorithm to enhance the anatomical structure boundaries.

[0036] Step 2: Construct an abdominal fat semantic segmentation model based on an improved U-Net architecture. Use image superpixels as basic units, fuse multi-modal features as input channels, and fuse local details and global context information through cross-scale skip connections. Replace the traditional convolutional layer with a dynamic convolutional kernel, and combine a channel attention mechanism to optimize the feature channel weight distribution, and output a pixel-level fat region mask.

[0037] Step 3: Establish 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, and dynamically adjust the task weights based on a gradient conflict resolution algorithm. Automatically generate a lightweight sub-network through neural architecture search, and use domain adaptation technology to adapt to cross-device imaging differences.

[0038] Step 4: Perform uncertainty modeling on the segmentation model based on Gaussian process regression, screen high-confidence regions, and construct a probability distribution map. Actively learn and annotate low-confidence regions through asynchronous distributed sampling, and adopt the entropy minimization strategy to iteratively optimize the generalization performance of the model.

[0039] Step 5: Design a multi-granularity feature fusion framework, embed anatomical prior knowledge into the graph convolutional network, and construct a graph of the association between fat distribution and organ topology. Use an adaptive pooling layer to extract multi-scale spatial features, and constrain the discriminability of features through contrastive learning loss to generate a structured fat analysis report.

[0040] The following further illustrates the implementation of the present invention in combination with Embodiments 1 to 5.

[0041] Embodiment 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 a residual autoencoder is used. The encoder of this model uses dilated convolutions to expand the receptive field. Dilated convolutions can increase the receptive field of the convolutional kernel without increasing the number of parameters and computational complexity, enabling the encoder to obtain more extensive image information. In this way, the encoder can better capture the long-range dependencies in the image and more accurately identify the noise features in the image.

[0043] The decoder introduces deformable convolutions to adapt to anatomical deformations. The sampling point positions of deformable convolutions can be adaptively adjusted according to the image content, which is very important for the variable-shaped anatomical structures in abdominal medical images. In traditional convolutions, the sampling point positions are fixed regular grids, but in abdominal images, the shapes and positions of organs and adipose tissues vary greatly. Deformable convolutions can dynamically adjust the sampling point positions according to the local features of the image, thus more accurately restoring the anatomical structure.

[0044] The denoised data is shared among multiple medical institutions through the federated learning framework. To improve data sharing efficiency and protect patient privacy, the federated learning framework designs a gradient sparsification mechanism. During the federated learning process, each medical institution acts as a node and uploads the Top-K significant gradients to the aggregation server. Specifically, after each node calculates the gradients of the local model, it sorts the gradients and selects the K gradients with the largest absolute values for uploading. Suppose the local gradient vector is The Top-K significant gradients selected after sorting are represented as where 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 the Laplace distribution). Assuming 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 the privacy of patients.

[0046] Embodiment 2:

[0047] This embodiment focuses on the key components in the improved U-Net architecture in step 2, namely the dynamic convolution kernel and the channel attention mechanism, to improve the segmentation accuracy of the model for the fat area. When constructing the abdominal fat semantic segmentation model, the dynamic convolution kernel is generated by a hypernetwork and can adaptively adjust the convolution parameters according to the content of the input image. The hypernetwork uses a lightweight Transformer structure to generate dynamic convolution parameters.

[0048] First, the input image is block-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 sizes of the image blocks. 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 through the self-attention mechanism. The self-attention mechanism calculates the correlation weights between each vector and other vectors, and the formula is:

[0049]

[0050] where Q, K, and V are the query matrix, key matrix, and value matrix respectively, which are obtained by linear transformation of the input sequence, and d k is the dimension of the key matrix. In this way, the hypernetwork can generate dynamic convolution parameters according to the global information of the image, enabling the convolution kernel to better adapt to the characteristics of the fat area at different positions and scales.

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

[0052]

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

[0054] Example 3:

[0055] In this example, the multi-task joint learning framework and neural architecture search in step 3 are optimized 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 adopted, and fat volume calculation, visceral-subcutaneous fat classification, and fat metabolism correlation prediction are defined as parallel tasks. Since there may be gradient conflicts between different tasks, a gradient conflict resolution algorithm is used to dynamically adjust the task weights. The gradient conflict resolution algorithm adopts the projected gradient descent method to project the multi-task gradients onto the orthogonal direction of the shared subspace.

[0056] Suppose the gradient vectors of multiple tasks are respectively (m is the number of tasks), and the shared subspace is S. First, calculate the projection of all task gradients on the shared subspace Then project the projected gradient vector onto the direction orthogonal to the shared subspace to obtain the adjusted gradient vector In this way, the gradient conflict between multiple tasks can be effectively alleviated, enabling the model to obtain better training effects on different tasks.

[0057] Neural architecture search is based on the differentiable architecture parameterization method, and jointly optimizes the network structure and weights through continuous relaxation techniques. The differentiable architecture parameterization method represents the network architecture as differentiable parameters, enabling the use of optimization algorithms such as gradient descent to search for the optimal network structure. During the search process, the discrete architecture search space is transformed into a continuous space through continuous relaxation techniques. For example, discrete operations such as the selection of network layers and the size of convolutional kernels are represented as continuous parameters. Suppose the network architecture parameter is α, and by simultaneously differentiating the objective function (such as the loss function) with respect to α and the network weight θ, that is and Use the gradient descent method to update α and θ, thereby realizing the joint optimization of the network structure and weights, automatically generating a lightweight sub-network, and improving the computational efficiency and generalization ability of the model.

[0058] Example 4:

[0059] This embodiment focuses on optimizing the uncertainty modeling and active learning annotation of the segmentation model in step 4 to improve the generalization performance of the model. When performing uncertainty modeling on the segmentation model based on Gaussian process regression, the Gaussian process regression uses a spectral mixture kernel function to model spatial correlation. The spectral mixture kernel function enables fast computation through Fourier feature mapping.

[0060] Assume the input data is \(x\in\mathbb{R}\) d (where \(d\) is the data dimension), the spectral mixture kernel function \(k(x,x'\)) ′ can be expressed as:

[0061]

[0062] where \(S(\omega)\) is the power spectral density function. Through Fourier feature mapping, the calculation of the kernel function in the high-dimensional space is converted into the inner product calculation in the low-dimensional space, greatly improving the computational efficiency. When constructing the probability distribution map, the probability that each pixel belongs to the fat area is calculated through Gaussian process regression, and the high-confidence regions are screened out.

[0063] For the low-confidence regions, active learning annotation is performed through asynchronous distributed sampling. Active learning annotation preferentially selects samples near the decision boundary based on the boundary sampling strategy. The decision boundary is the region where the classification result of the model is uncertain. Annotating in these regions can obtain more information, which helps the model better learn the boundary between different categories.

[0064] The entropy minimization strategy is used to iteratively optimize the generalization performance of the model. The entropy minimization strategy estimates the prediction uncertainty through Monte Carlo Dropout. The Monte Carlo Dropout sampling uses an adaptive Markov chain Monte Carlo method to accelerate convergence. During model inference, the Dropout layer is randomly turned on multiple times to obtain multiple prediction results. Assume \(T\) samplings are performed, and the prediction results are \(y_1,y_2,\cdots,y_T\). T Calculate the entropy \(H(y)\) of the prediction results:

[0065]

[0066] where \(C\) is the number of categories, and \(p(y = i)\) is the probability of being predicted as the \(i\)-th category. By selecting samples with larger entropy for annotation, 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 embodiment mainly optimizes the multi-granularity feature fusion framework and the construction of the fat distribution and organ topology association map in step 5 to generate a more valuable structured fat analysis report. When designing the multi-granularity feature fusion framework, anatomical prior knowledge is embedded into the graph convolutional network. The anatomical prior knowledge is represented by a knowledge graph, which contains information such as abdominal organs, adipose tissues, and the relationships between them. By mapping the nodes and edges in these knowledge graphs into the graph convolutional network, prior information is provided for the model.

[0069] When constructing the fat distribution and organ topology association map, a 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 nodes in the graph are v i and v j , and their feature vectors are h i and h j , respectively. The calculation formula for the attention coefficient α ij is as follows:

[0070]

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

[0072] An 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 at different scales. When generating the structured fat analysis report, the contrastive learning loss adopts a hard example mining strategy to enhance the diversity of negative samples. The hard example mining strategy improves the discriminative ability of the model by selecting negative samples with large differences from the positive samples for training. Assume that the positive sample is x + , and the set of negative samples is {x -}. During the training process, negative samples that are far away from x + in the feature space are selected The discriminability of features is constrained by the contrastive learning loss function L contrast , enabling the model to better distinguish different fat regions and organs and generate a more accurate and structured fat analysis report.

[0073] It should be noted that, in this text, relational terms such as first and second are only used 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 "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

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

Claims

1. A method for analyzing human abdominal fat based on medical images, characterized in that, It includes the following steps: Step 1: Obtain multi-modal medical image data, including CT, MRI and ultrasound images. Align heterogeneous modal data through multi-scale affine transformation to construct an abdominal fat multi-modal fusion database. For image noise or artifacts, use a denoising model based on a residual autoencoder for repair, and combine with the Canny edge detection algorithm to enhance the anatomical structure boundaries; Step 2: Build an abdominal fat semantic segmentation model based on an improved U-Net architecture. Take image superpixels as basic units, fuse multi-modal features as input channels, and fuse local details and global context information through cross-scale skip connections. Replace traditional convolutional layers with dynamic convolutional kernels, and combine with a channel attention mechanism to optimize the feature channel weight allocation, and output a pixel-level fat region mask; Step 3: Establish 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, and dynamically adjust task weights based on a gradient conflict resolution algorithm. Automatically generate a lightweight sub-network through neural architecture search, and use domain adaptation technology to adapt to cross-device imaging differences; Step 4: Perform uncertainty modeling on the segmentation model based on Gaussian process regression, screen high-confidence regions and construct a probability distribution map. Actively learn and annotate low-confidence regions through asynchronous distributed sampling, and use an entropy minimization strategy to iteratively optimize the model generalization performance; Step 5: Design a multi-granularity feature fusion framework, embed anatomical prior knowledge into a graph convolutional network, and construct a graph of the association between fat distribution and organ topology. Use an adaptive pooling layer to extract multi-scale spatial features, and generate a structured fat analysis report by constraining the feature discriminability through contrastive learning loss; 2. The method for analyzing human abdominal fat based on medical images according to claim 1, wherein In the encoder of the residual autoencoder in Step 1, dilated convolution is used to expand the receptive field, and deformable convolution is introduced in the decoder to adapt to anatomical deformations; 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 human abdominal fat based on medical images according to claim 1, wherein In Step 2, the dynamic convolutional kernel is generated by a hypernetwork, and the convolutional parameters are adaptively adjusted according to the input image content; the channel attention mechanism adopts a squeeze-excitation structure to generate a channel weight vector through global average pooling and a fully connected layer.

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

5. The method for analyzing human abdominal fat based on medical images according to claim 1, wherein In Step 4, Gaussian process regression uses a spectral mixture kernel function to model spatial correlation, and active learning annotation preferentially selects samples near the decision boundary based on a boundary sampling strategy; the entropy minimization strategy estimates prediction uncertainty through Monte Carlo Dropout.

6. The method for analyzing human abdominal fat based on medical images according to claim 1, wherein In Step 5, the anatomical prior knowledge embedding uses a knowledge graph representation, and a graph attention mechanism is introduced in the graph construction to model the energy transfer relationship between organs; the contrastive learning loss uses a hard example mining strategy to enhance the diversity of negative samples.

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

8. The method for analyzing human abdominal fat based on medical images according to claim 3, wherein The hypernetwork adopts a lightweight Transformer structure to generate dynamic convolution parameters, and the input image is block-encoded into a sequence and then captures long-range dependencies through the self-attention mechanism.

9. The method for analyzing human abdominal fat 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 through the normalized mutual information metric; the Canny edge detection parameters are dynamically optimized by the particle swarm algorithm.

10. The method for analyzing human abdominal fat based on medical images according to claim 5, wherein The spectral mixture kernel function is realized by Fourier feature mapping for fast calculation, and the Monte Carlo Dropout sampling adopts the adaptive Markov chain Monte Carlo method to accelerate convergence.

Images