A metabolic syndrome image segmentation method and system based on a deep learning algorithm

Abstract

The application discloses a kind of metabolic syndrome image segmentation method and system based on deep learning algorithm, comprising: obtaining the abdominal CT scan image to be segmented and pre-processing;The image after pre-processing is input to the image segmentation model trained, and the segmentation mask of subcutaneous fat is output, the image segmentation of subcutaneous fat is completed;Image segmentation model is the ENDNet model improved based on U-net3+, specifically comprising: the structure of multilayer encoder and multilayer decoder is successively connected, Enhanced-Inception module is introduced in each layer encoder, multi-scale feature information is extracted by multilayer encoder respectively, and the encoding feature containing rich hierarchical information is obtained;The shallow features of encoder and the deep features of decoder are fused by jump connection, the fusion features are processed by decoder, and accurate pixel-level segmentation result is generated.The application can enhance the multi-scale feature extraction capability of model, fully adapt to the multi-scale characteristics of subcutaneous fat, so as to improve the segmentation precision.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to an image segmentation method and system for metabolic syndrome based on deep learning algorithms. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Metabolic syndrome, a group of metabolic disorders centered on abdominal obesity, has seen a significant increase in prevalence as global obesity rates rise, and has become a major public health problem threatening human health.

[0004] Traditional detection methods mostly rely on indicators such as waist circumference and blood pressure, which are subject to regional / ethnic differences and cannot accurately reflect the distribution of abdominal fat, thus having significant limitations. With the development of medical imaging technology, computed tomography (CT) scans of the abdomen can clearly display abdominal fat; subcutaneous fat can be segmented through image segmentation technology, and quantitative analysis of subcutaneous fat can provide auxiliary decision-making basis for the diagnosis of metabolic syndrome.

[0005] However, traditional image segmentation methods (such as thresholding and region growing) rely on manual feature and parameter adjustments, are sensitive to noise, and have difficulty handling issues such as overlapping tissue gray levels and blurred boundaries in abdominal CT scans. This results in low segmentation accuracy and a tendency for oversegmentation or undersegmentation, requiring extensive manual intervention.

[0006] Furthermore, subcutaneous fat exhibits significant multi-scale characteristics in medical images. Because the grayscale values ​​or signal intensities of subcutaneous fat, skin, and adjacent muscle tissue are sometimes very close, the boundaries are blurred and difficult to distinguish accurately. Existing deep learning models often suffer from limited feature extraction capabilities, insufficient capture of multi-scale contextual information, low accuracy, complex model structure, large number of parameters, and low inference efficiency when faced with subcutaneous fat image segmentation. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a method and system for segmenting metabolic syndrome images based on deep learning algorithms. It employs an ENDNet model improved from U-net3+, introducing an Enhanced-Inception module in the encoder to enhance multi-scale feature extraction capabilities, and incorporating a NAM attention mechanism in skip connections to strengthen key features. The decoder uses depthwise separable convolutions for lightweight implementation, thus resolving the problems of insufficient accuracy and large parameter count in subcutaneous fat segmentation using U-net3+, thereby improving segmentation accuracy.

[0008] In some implementations, the following technical solutions are adopted: A method for image segmentation of metabolic syndrome based on deep learning algorithms, comprising: Acquire abdominal CT scan images to be segmented and perform preprocessing; The preprocessed image is input into the trained image segmentation model, which outputs a segmentation mask for subcutaneous fat to complete the image segmentation of subcutaneous fat. The image segmentation model is an ENDNet model improved based on U-net3+, specifically including: a multi-layer encoder and a multi-layer decoder structure connected in series. An Enhanced-Inception module is introduced in each encoder layer. Multi-scale feature information is extracted through the multi-layer encoder to obtain encoded features containing rich hierarchical information. The shallow features of the encoder and the deep features of the decoder are fused through skip connections. The decoder processes the fused features to generate accurate pixel-level segmentation results.

[0009] As a further solution, the Enhanced-Inception module expands the number of channels of the original input by a set factor through 1×1 convolution to enhance feature representation, and then compresses it back to the original dimension using 1×1 convolution. At the same time, residual connections are introduced to add the original input and the compressed features as the output of the encoder, thereby effectively capturing multi-scale local details and global context.

[0010] As a further solution, a NAM attention mechanism is added to the skip connection. The NAM attention mechanism uses a batch normalized scaling factor to generate channel weights and combines the L1 norm to calculate spatial weights, so as to fuse the shallow features of the encoder and the deep features of the decoder.

[0011] As a further solution, the decoder processes the fused features through depthwise separable convolution, which is divided into depthwise convolution and pointwise convolution. The depthwise convolution performs spatial convolution on each channel independently, and the pointwise convolution fuses channel information through 1×1 convolution.

[0012] As a further solution, the ENDNet model based on U-net3+ improvement includes a five-layer encoder and a four-layer decoder structure, where the output of each encoder layer serves as the input of the next encoder layer; the input of each decoder layer includes: the output of the decoders preceding that decoder layer, the output of the last encoder layer, the output of its corresponding encoder, and the outputs of all encoders preceding it.

[0013] As a further step, the preprocessing of abdominal CT scan images is as follows: Set the CT value range and center point of subcutaneous fat, and linearly transform the CT values ​​within this range to an 8-bit grayscale range [0, 255].

[0014] As a further approach, when training the image segmentation model, images are selected from abdominal CT scan slices of a predetermined number of patients with metabolic syndrome. The specific process includes: Exclude images containing motion artifacts, metal artifacts, or lesions; Adjust the CT value range and center point at different locations to enhance tissue contrast; Extract the region of interest, compress the CT value range of the region of interest to [0, 255] grayscale values, and save it; Professionals mark the subcutaneous fat boundaries in the saved images to create segmentation labels; The saved images and their corresponding segmentation labels serve as the final training dataset.

[0015] In other embodiments, the following technical solutions are adopted: A metabolic syndrome image segmentation system based on deep learning algorithms, comprising: The data acquisition module is configured to acquire and preprocess abdominal CT scan images to be segmented. The image segmentation module is configured to input the preprocessed image into the trained image segmentation model and output a segmentation mask for subcutaneous fat to complete the image segmentation of subcutaneous fat. The image segmentation model is an ENDNet model improved based on U-net3+, specifically including: a multi-layer encoder and a multi-layer decoder structure connected in series. An Enhanced-Inception module is introduced in each encoder layer. Multi-scale feature information is extracted through the multi-layer encoder to obtain encoded features containing rich hierarchical information. The shallow features of the encoder and the deep features of the decoder are fused through skip connections. The decoder processes the fused features to generate accurate pixel-level segmentation results.

[0016] As a further solution, the ENDNet model based on U-net3+ improvement includes a five-layer encoder and a four-layer decoder structure, where the output of each encoder layer serves as the input of the next encoder layer; the input of each decoder layer includes: the output of the decoders preceding that decoder layer, the output of the last encoder layer, the output of its corresponding encoder, and the outputs of all encoders preceding it.

[0017] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, the processor being used to implement instructions; the memory being used to store multiple instructions adapted to be loaded and executed by the processor for the above-described image segmentation method for metabolic syndrome based on a deep learning algorithm.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention uses an ENDNet model based on U-net3+ to segment images of metabolic syndrome. The ENDNet model is an improved version of U-net3+, which introduces skip connections between the encoding and decoding layers. Each decoder integrates feature maps from different scales of the encoder, providing multi-scale information. This allows the model to accurately segment fat regions, whether they are extremely thin or extremely thick. It can explicitly model the uncertainties in the segmentation problem, yielding the most reasonable and precise segmentation results, especially when dealing with subcutaneous fat, which has blurred boundaries and varied shapes. Furthermore, it allows for precise quantitative analysis of subcutaneous fat, providing auxiliary decision-making support for the diagnosis of metabolic syndrome.

[0019] (2) The present invention introduces an Enhanced-Inception module in the encoder based on the U-net3+ improved ENDNet model. By using convolution kernels of different sizes in parallel, it can capture features of different ranges at the same network level, enhance the multi-scale feature extraction capability of the model, fully adapt to the multi-scale characteristics of subcutaneous fat, and thus improve the segmentation accuracy.

[0020] (3) The present invention is based on the improved ENDNet model of U-net3+. The NAM attention mechanism is added to the skip connection to strengthen the key features. Without introducing additional parameters and computational overhead, the important and unimportant features are accurately distinguished by weight sparsification, which helps to reduce feature redundancy and overfitting.

[0021] (4) The decoder of the present invention uses depth-separable convolution instead of traditional convolution, which can improve segmentation accuracy while reducing the number of parameters and realize lightweight design of model structure.

[0022] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] Figure 1 This is a flowchart of the metabolic syndrome image segmentation method based on deep learning algorithm in an embodiment of the present invention; Figure 2 This is a schematic diagram of the ENDNet model structure based on the U-net3+ improvement in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Enhanced-Inception module structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of a metabolic syndrome image segmentation system based on a deep learning algorithm in an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Terminology Explanation: U-net3+: U-Net3+ is a medical image segmentation architecture that is a further upgrade and innovation of the classic U-Net and U-Net++. It fully inherits the core skeleton of the classic U-Net's U-shaped encoder-decoder structure and can achieve more complete and efficient full-scale feature fusion through full-scale skip connections.

[0027] ENDNet Model: ENDNet is a deep learning image segmentation model improved based on the U-net3+ architecture. This model introduces an Enhanced-Inception module in the encoder, which enhances the ability to extract multi-scale contextual features through a multi-branch parallel convolutional structure to better capture feature information of targets of different sizes. It embeds a NAM attention mechanism in skip connections to perform channel and spatial dual-dimensional adaptive calibration on cross-layer fused features, effectively suppressing background redundancy interference. The decoder part of the model uses depthwise separable convolution to replace the traditional two-layer convolution, which significantly reduces the number of model parameters and computational complexity while improving the semantic consistency of feature fusion, ultimately achieving higher-precision pixel-level segmentation results.

[0028] Enhanced-Inception module: Enhanced Inception refers to a more powerful and efficient modern convolutional module that has undergone a series of optimizations and improvements based on the classic Inception module.

[0029] In this invention, the Enhanced-Inception module expands the number of channels of the original input by a set factor through 1×1 convolution to enhance feature representation, and then compresses it back to the original dimension using 1×1 convolution. While keeping the receptive field unchanged, it significantly reduces parameters and computational cost. At the same time, residual connections are introduced to add the original input and the compressed features as the output of the encoder, thereby effectively capturing multi-scale local details and global context.

[0030] NAM stands for Normalization-based Attention Module; it allows attention weights to be directly derived using the normalization of weights without introducing new learnable parameters.

[0031] Example 1 In one or more embodiments, a method for image segmentation of metabolic syndrome based on a deep learning algorithm is disclosed, combining... Figure 1 Specifically, it includes the following process: S101: Acquire the abdominal CT scan image to be segmented and perform preprocessing.

[0032] In this embodiment, the patient to be tested first undergoes an abdominal CT scan to obtain abdominal CT scan images.

[0033] The abdominal CT scan images were then preprocessed, specifically as follows: The CT value range (window width) and center point (window level) of subcutaneous fat are set, and the CT values ​​within this range are linearly transformed to an 8-bit grayscale range [0, 255]. This maps the massive amount of grayscale data acquired by CT to a limited grayscale level that the human eye can effectively distinguish, thereby highlighting specific tissues of interest.

[0034] S102: Input the preprocessed image into the trained image segmentation model, output the subcutaneous fat segmentation mask, and complete the image segmentation of subcutaneous fat; The image segmentation model is an ENDNet model based on U-net3+, which specifically includes a multi-layer encoder and a multi-layer decoder structure connected in series. An Enhanced-Inception module is introduced in each encoder layer to extract multi-scale feature information through the multi-layer encoders, resulting in encoded features containing rich hierarchical information. The shallow features of the encoder and the deep features of the decoder are fused through skip connections, and the fused features are processed by the decoder to generate accurate pixel-level segmentation results.

[0035] As a specific implementation method, the structure of the image segmentation model is as follows: Figure 2 As shown, Figure 2In the diagram, the blue boxes represent the original Unet3+ convolutional layers. The image segmentation model specifically includes a five-layer encoder and a four-layer decoder structure connected sequentially. The output of each encoder layer serves as the input to the next encoder layer. The input to each decoder layer includes: the output of the decoders preceding it, the output of the last encoder layer, and the outputs of its corresponding encoder and all preceding encoders. Specifically, the preprocessed image is input to the first encoder layer, which is then connected sequentially to the second, third, fourth, and fifth encoder layers. The output of the previous encoder layer serves as the input to the next encoder layer. The output of the fifth encoder layer is upsampled and then input to the first, second, third, and fourth decoder layers respectively.

[0036] The outputs of the first, second, and third encoders are downsampled and then connected to the output of the fourth encoder via a skip connection to serve as the input to the fourth decoder. The outputs of the first and second encoders are downsampled and then connected to the output of the third encoder via a skip connection to serve as the input to the third decoder. The output of the first encoder is downsampled and then connected to the output of the second encoder via a skip connection to serve as the input to the second decoder. The output of the first encoder is connected via a skip connection to serve as the input to the first decoder. The output of the fourth decoder is upsampled (e.g., by bilinear interpolation or transposed convolution) and then used as the input to the third and second decoders respectively. The output of the third decoder is upsampled (e.g., by bilinear interpolation or transposed convolution) and then used as the input to the second and first decoders respectively. The output of the second decoder is upsampled (e.g., by bilinear interpolation or transposed convolution) and then used as the input to the first decoder.

[0037] In this embodiment, each encoder layer downsamples through convolution or pooling to progressively extract features, increase the number of channels, and reduce spatial size. Each decoder layer is a feature fusion center, fusing features from its corresponding encoder and all previous encoders, as well as features from all previous decoders. These features include high-resolution, detailed shallow features (to help accurately locate subcutaneous fat boundaries), features at the same scale (to provide key information at the current scale), and low-resolution, semantically rich deep features (to provide context). This makes each decoder node an information hub, aggregating comprehensive information from local to global perspectives. This results in higher segmentation accuracy, clearer boundaries, and stronger model robustness. Whether dealing with extremely thin or extremely thick subcutaneous fat regions, it can provide corresponding features for accurate segmentation. This allows for explicit modeling of uncertainties in the segmentation problem, yielding the most reasonable and accurate segmentation results, especially when dealing with subcutaneous fat, which has blurred boundaries and varied shapes, demonstrating significant advantages.

[0038] In this embodiment, each encoder layer uses an Enhanced-Inception module instead of traditional convolution, capturing multi-scale features through multi-branch parallel convolution; the structure of the Enhanced-Inception module is as follows: Figure 3 As shown, the input channel number is first expanded to 4 times through 1×1 convolution, BN, and GeLu in the channel expansion module, providing rich features for multi-branch operations. Then, multi-scale feature parallel extraction is achieved through multi-branch depthwise separable convolution of InceptionDWConv2d. Subsequently, the dynamic feature calibration module generates attention weights through average pooling, convolution, activation, and sigmoid, adaptively calibrating the features in terms of channels and space, and multiplying them element-wise with the original features. Next, the channel compression module restores the number of channels through 1×1 convolution and BN. Finally, the residual connection is added to the input features to achieve multi-scale feature extraction, redundancy suppression, and feature enhancement, thereby improving the model's feature representation ability. This effectively captures multi-scale local details and global context, enhances the model's multi-scale feature extraction ability, fully adapts to the multi-scale characteristics of subcutaneous fat, and improves segmentation accuracy.

[0039] In this embodiment, a NAM attention mechanism is added to the skip connections. The NAM attention mechanism provides two sub-modules: a channel attention sub-module and a spatial attention sub-module. The channel attention sub-module performs global average pooling on each channel c of the image, utilizing a batch-normalized scaling factor. The original importance score for this channel is generated, and then normalized using a modified Sigmoid function: ; Where k is a hyperparameter used to control the flatness of the weight distribution; It is a vector L1 norm (all) The sum of ( ) is a regularization term that penalizes larger values. This value promotes weight sparsity, meaning that the weights of a few important channels are close to 1, while the weights of most unimportant channels are suppressed to close to 0. The resulting channel weights... Multiplying each channel by its corresponding original channel c can highlight important channels and suppress secondary channels.

[0040] The spatial attention submodule calculates each spatial location ( i , j By iterating through the L1 norm of all channels, a saliency map of H x W x 1 is obtained. Here, H is the image height and W is the image width. Similarly, a batch normalized scaling factor is used. Generate the original importance scores for this space, and then normalize them using the sigmoid function with regularization: ; in, yes( i , j The significance value of the position. It is a vector L1 norm (all) The resulting spatial weights are multiplied by the corresponding original spatial locations to highlight important areas and suppress background or noisy areas.

[0041] By incorporating skip connections with the NAM attention mechanism, important and unimportant features can be accurately distinguished through weight sparsity without introducing additional parameters and computational overhead. This helps reduce feature redundancy and overfitting, enabling efficient fusion of shallow features from the encoder and deep features from the decoder.

[0042] In this embodiment, each decoder processes the fused features through depthwise separable convolution, which can reduce the number of parameters while improving segmentation accuracy, thus achieving a lightweight design of the model structure.

[0043] This embodiment uses depthwise separable convolution instead of traditional 3×3 convolution. Depthwise separable convolution is divided into depthwise convolution and pointwise convolution. Depthwise convolution performs spatial convolution on each channel independently, while pointwise convolution fuses channel information through 1×1 convolution. Compared with the U-net3+ network, the number of parameters can be reduced by 26.32%.

[0044] This embodiment is based on the ENDNet model improved by U-net3+. The specific processing procedure for the input image is as follows: Feature extraction is performed through the encoder. The Enhanced-Inception module of each encoder layer adopts a multi-branch parallel convolutional structure. First, the number of input channels is increased by a factor of 4 through 1×1 convolution to enhance feature expression. Then, 1×1 convolution is used to compress it back to the original dimension. At the same time, residual connections are introduced to add the original input and the compressed features, thereby effectively capturing multi-scale local details and global context, avoiding the loss of deep network features, and obtaining encoded features containing rich hierarchical information. Subsequently, the skip connection part fuses the shallow features of the encoder and the deep features of the decoder through the NAM attention mechanism. This mechanism uses the scaling factor of the batch normalization layer to generate channel weights and combines the cross-channel L1 norm to calculate the spatial saliency distribution, realizing feature weighting in both channel and spatial dimensions, suppressing irrelevant background and redundant information, highlighting key features such as edges and textures related to the target, and obtaining optimized fused features. Finally, the decoder processes the fused features through depthwise separable convolution, which is decomposed into depthwise convolution (spatial convolution is performed independently on each channel) and pointwise convolution (channel information is fused through 1×1 convolution). This refines the features while reducing the number of parameters and computational cost, gradually restoring the spatial resolution of the image. Finally, it generates accurate pixel-level segmentation results through nonlinear transformation, achieving precise segmentation of the subcutaneous fat region.

[0045] Currently, large, publicly available labeled datasets are scarce in research on abdominal fat CT image segmentation related to metabolic syndrome. This restricts the development of image segmentation algorithms for the auxiliary diagnosis of metabolic syndrome. In this embodiment, images were selected from abdominal CT scan slices of 209 subjects and preprocessed to construct a training dataset for the image segmentation model. The specific process is as follows: (1) Exclude images that have motion artifacts, metal artifacts, lesions, etc. that affect segmentation; (2) Contrast adjustment: Set the corresponding window width and window level for different tissues (e.g., subcutaneous fat window width 350HU, window level -80HU). (3) Extract the region of interest (ROI), linearly transform the CT value range of the ROI (e.g., [-1000, +1000]) to the 8-bit grayscale range [0,255], and save it as a JPEG format; (4) Professionals mark the subcutaneous fat boundaries on the saved images to form segmentation labels; (5) The saved images and their corresponding segmentation labels are used as the final training dataset.

[0046] In addition, a CT image dataset containing Gaussian (μ=0, σ²=0.02) and Poisson noise was constructed, and denoising verification was performed using Wiener filtering, guided filtering, median filtering, and other methods. It can be concluded that: for Gaussian noise, median filtering is preferred (SSIM up to 0.439), and for Poisson noise, Wiener filtering is preferred (SSIM up to 0.962).

[0047] The image segmentation model was trained using the aforementioned training dataset. The segmentation performance was then evaluated using metrics such as Dice (F1 score), IoU (Intersection over Union), Recall, and Precision. Table 1 compares the image segmentation results of the subcutaneous fat dataset on existing U-net, Att-Unet (Attention U-Net), Swin-Unet (U-Net reconstructed using Transformer), U-net2+, U-net3+, Trans-Unet (fusion of Transformer and classic U-Net architecture), and the model in this embodiment (an ENDNet model improved based on U-net3+). It can be seen that the model in this embodiment achieves an average Dice of 98.04% in the subcutaneous fat segmentation task, an improvement of 0.76% compared to the U-net3+ network, while reducing the number of parameters by 26.32%, significantly improving segmentation accuracy and reducing the number of parameters.

[0048] Table 1. Comparison of image segmentation results of the subcutaneous fat dataset on different models.

[0049] This embodiment of the image segmentation model introduces skip connections between the encoding and decoding layers. Each decoder fuses feature maps from different scales of the encoder, possessing multi-scale information. This allows the model to accurately segment fat regions, whether extremely thin or extremely thick, providing corresponding features. It exhibits significant advantages, especially when dealing with subcutaneous fat, which has blurred boundaries and varied shapes. Furthermore, it enables precise quantitative analysis of subcutaneous fat, providing auxiliary decision-making support for the diagnosis of metabolic syndrome.

[0050] Example 2 In one or more embodiments, a metabolic syndrome image segmentation system based on a deep learning algorithm is disclosed, combined with... Figure 4 ,include: The data acquisition module is configured to acquire and preprocess abdominal CT scan images to be segmented. The image segmentation module is configured to input the preprocessed image into the trained image segmentation model and output a segmentation mask for subcutaneous fat to complete the image segmentation of subcutaneous fat. The image segmentation model is an ENDNet model improved based on U-net3+, specifically including: a multi-layer encoder and a multi-layer decoder structure connected in series. An Enhanced-Inception module is introduced in each encoder layer. Multi-scale feature information is extracted through the multi-layer encoder to obtain encoded features containing rich hierarchical information. The shallow features of the encoder and the deep features of the decoder are fused through skip connections. The decoder processes the fused features to generate accurate pixel-level segmentation results.

[0051] The ENDNet model based on U-net3+ includes a five-layer encoder and a four-layer decoder structure. The output of each encoder layer serves as the input of the next encoder layer. The input of each decoder layer includes the output of the decoders preceding it, the output of the last encoder layer, the output of its corresponding encoder, and the outputs of all encoders preceding it.

[0052] The Enhanced-Inception module uses 1×1 convolutions to expand the number of channels in the original input by a set factor to enhance feature representation, and then uses 1×1 convolutions to compress it back to the original dimension. At the same time, residual connections are introduced to add the original input and the compressed features as the output of the encoder, thereby effectively capturing multi-scale local details and global context.

[0053] A NAM attention mechanism is added to the skip connections. The NAM attention mechanism uses a batch normalized scaling factor to generate channel weights and combines the L1 norm to calculate spatial weights, so as to fuse the shallow features of the encoder and the deep features of the decoder.

[0054] The decoder processes the fused features through depthwise separable convolution, which is divided into depthwise convolution and pointwise convolution. The depthwise convolution performs spatial convolution on each channel independently, and the pointwise convolution fuses channel information through 1×1 convolution.

[0055] The specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.

[0056] Example 3 In one or more embodiments, a terminal device is disclosed, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the processor to perform the metabolic syndrome image segmentation method based on a deep learning algorithm as described in Embodiment 1.

[0057] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0058] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0059] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0060] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for image segmentation of metabolic syndrome based on deep learning algorithms, characterized in that, include: Acquire abdominal CT scan images to be segmented and perform preprocessing; The preprocessed image is input into the trained image segmentation model, which outputs a segmentation mask for subcutaneous fat to complete the image segmentation of subcutaneous fat. The image segmentation model is an ENDNet model based on U-net3+, specifically including: a multi-layer encoder and a multi-layer decoder structure connected in series. An Enhanced-Inception module is introduced in each encoder layer. Multi-scale feature information is extracted through the multi-layer encoder to obtain encoded features containing rich hierarchical information. The shallow features of the encoder and the deep features of the decoder are fused through skip connections. The decoder processes the fused features to generate accurate pixel-level segmentation results. The Enhanced-Inception module expands the input channels to a set factor through 1×1 convolution, BN, and GeLu to enhance feature representation; multi-scale feature extraction is achieved through multi-branch depthwise separable convolution of InceptionDWConv2d; attention weights are generated by the dynamic feature calibration module using average pooling, convolution, activation, and sigmoid to adaptively calibrate the features in terms of channels and space and multiply them element-wise with the original features; the number of channels is restored through 1×1 convolution and BN of the channel compression module, and residual connections are introduced to add the original input and the compressed features as the output of the encoder, thereby effectively capturing multi-scale local details and global context; The skip connection incorporates a NAM attention mechanism, which uses a batch normalized scaling factor to generate channel weights and combines the L1 norm to calculate spatial weights, thereby fusing the shallow features of the encoder and the deep features of the decoder. By incorporating skip connections with the NAM attention mechanism, important and unimportant features can be accurately distinguished through weight sparsity without introducing additional parameters and computational overhead. This helps reduce feature redundancy and overfitting, enabling efficient fusion of shallow features from the encoder and deep features from the decoder. The decoder processes the fused features through depthwise separable convolution, which is divided into depthwise convolution and pointwise convolution. The depthwise convolution performs spatial convolution on each channel independently, and the pointwise convolution fuses channel information through 1×1 convolution. The ENDNet model based on U-net3+ improvement includes a five-layer encoder and a four-layer decoder structure. The output of each encoder layer serves as the input of the next encoder layer. The input of each decoder layer includes the output of the decoders preceding it, the output of the last encoder layer, the output of its corresponding encoder, and the outputs of all encoders preceding it.

2. The image segmentation method for metabolic syndrome based on deep learning algorithm as described in claim 1, characterized in that, The preprocessing process for abdominal CT scan images is as follows: Set the CT value range and center point of subcutaneous fat, and linearly transform the CT values ​​within this range to an 8-bit grayscale range [0, 255].

3. The image segmentation method for metabolic syndrome based on deep learning algorithm as described in claim 1, characterized in that, When training the image segmentation model, images are selected from abdominal CT scan slices of a set number of patients with metabolic syndrome. The specific process includes: Exclude images containing motion artifacts, metal artifacts, or lesions; Adjust the CT value range and center point at different locations to enhance tissue contrast; Extract the region of interest, compress the CT value range of the region of interest to [0, 255] grayscale values, and save it; Professionals mark the subcutaneous fat boundaries in the saved images to create segmentation labels; The saved images and their corresponding segmentation labels serve as the final training dataset.

4. A metabolic syndrome image segmentation system based on a deep learning algorithm, characterized in that, include: The data acquisition module is configured to acquire and preprocess abdominal CT scan images to be segmented. The image segmentation module is configured to input the preprocessed image into the trained image segmentation model and output a segmentation mask for subcutaneous fat to complete the image segmentation of subcutaneous fat. The image segmentation model is an ENDNet model based on U-net3+, specifically including: a multi-layer encoder and a multi-layer decoder structure connected in series. An Enhanced-Inception module is introduced in each encoder layer. Multi-scale feature information is extracted through the multi-layer encoder to obtain encoded features containing rich hierarchical information. The shallow features of the encoder and the deep features of the decoder are fused through skip connections. The decoder processes the fused features to generate accurate pixel-level segmentation results. The Enhanced-Inception module expands the input channels to a set factor through 1×1 convolution, BN, and GeLu to enhance feature representation; multi-scale feature extraction is achieved through multi-branch depthwise separable convolution of InceptionDWConv2d; attention weights are generated by the dynamic feature calibration module using average pooling, convolution, activation, and sigmoid to adaptively calibrate the features in terms of channels and space and multiply them element-wise with the original features; the number of channels is restored through 1×1 convolution and BN of the channel compression module, and residual connections are introduced to add the original input and the compressed features as the output of the encoder, thereby effectively capturing multi-scale local details and global context; The skip connections incorporate a NAM attention mechanism, which uses a batch-normalized scaling factor to generate channel weights and combines the L1 norm to calculate spatial weights, thereby fusing shallow features of the encoder and deep features of the decoder. By incorporating the NAM attention mechanism into the skip connections, important and unimportant features can be accurately distinguished through weight sparsity without introducing additional parameters or computational overhead. This helps reduce feature redundancy and overfitting, enabling efficient fusion of shallow features of the encoder and deep features of the decoder. The decoder processes the fused features through depthwise separable convolution, which is divided into depthwise convolution and pointwise convolution. The depthwise convolution performs spatial convolution on each channel independently, and the pointwise convolution fuses channel information through 1×1 convolution. The ENDNet model based on U-net3+ improvement includes a five-layer encoder and a four-layer decoder structure. The output of each encoder layer serves as the input of the next encoder layer. The input of each decoder layer includes the output of the decoders preceding it, the output of the last encoder layer, the output of its corresponding encoder, and the outputs of all encoders preceding it.

5. A terminal device comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-3, the image segmentation method for metabolic syndrome based on a deep learning algorithm.

Images