Forearm blood vessel image segmentation method and device based on HAMU-Net deep learning network

Abstract

The invention provides a forearm blood vessel image segmentation method and device based on an HAMU-Net deep learning network, and relates to the technical field of medical image processing. The method comprises the following steps: performing forearm blood vessel image acquisition through a near-infrared acquisition device to obtain an original image; based on a Gaussian filter, according to a preset background gray threshold, preprocessing the original image to obtain a processed image; based on the Reinforcement-CLAHE technology, according to the processed image, a residual error convolution automatic encoder is used for carrying out multi-stage image enhancement processing, and an enhanced image is obtained; the method comprises the following steps of: constructing an HAMU-Net blood vessel segmentation model on the basis of a Hybrid Action module, a Triplet Action module, a Spatial Action module and a U-Net network structure; and inputting the enhanced image into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image. The near-infrared forearm blood vessel image segmentation method is an efficient and accurate near-infrared forearm blood vessel image segmentation method based on the mixed attention module.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a forearm blood vessel image segmentation method and device based on a HAMU-Net deep learning network. Background Art

[0002] In recent years, near-infrared optical imaging technology has gradually become an important non-invasive imaging tool in the biomedical field and is widely used in the visualization and monitoring of subcutaneous vascular structures. However, the visibility of subcutaneous vascular structures is often limited by factors such as biological tissue absorption, scattering, and light attenuation, and as a result, it cannot be directly put into practical application. Therefore, an effective image enhancement and image segmentation method is highly needed to improve the visualization quality of subcutaneous veins and provide strong support for venipuncture operations in auxiliary clinical medicine.

[0003] Medical image segmentation is a core task in the field of medical image processing. At present, the commonly used medical image segmentation methods are mainly divided into traditional segmentation methods and deep learning segmentation methods based on convolutional neural networks. Traditional segmentation methods can be divided into threshold-based, edge-based and region-based. Threshold-based methods, such as full threshold segmentation and multi-threshold segmentation, segment images based on the threshold of pixel grayscale values. They are simple and fast to calculate, suitable for processing images with high contrast, but sensitive to complex backgrounds and noise; edge-based methods, such as edge detection and active contour models, aim to detect edges or contours in images, but have poor segmentation effects for images with complex textures and large illumination changes; region-based methods, such as region growing and region segmentation, divide image pixels into several adjacent regions or segmentation units. The segmentation results are continuous and have smooth edges, but are sensitive to the selection of seed points. In recent years, the rise of deep learning technology has led to a revolutionary change in the field of medical image segmentation. Deep learning methods, especially convolutional neural networks and semantic segmentation networks, achieve a high level of understanding of image semantic information by learning image features from large-scale data. These deep learning methods can use the learned features for pixel-level classification and segmentation, thereby achieving more accurate and automated medical image segmentation. However, compared with traditional methods, deep learning often requires a large amount of manually annotated training data to train the model, and medical images are more difficult to obtain than other types of images. At the same time, the annotation process usually requires the annotator to have certain professional knowledge and takes a lot of time.

[0004] U-Net, as a variant of Fully Convolutional Networks (FCN), is the most commonly used and simplest segmentation model. This model is efficient and easy to build, and supports centralized training with small data and small samples. The original intention of U-Net design was to solve the problem of medical image segmentation, especially the segmentation task at the cellular level. Each submodule of U-Net is replaced by a residual module, and a residual U-Net (Residual U-Net, Res-Unet) is designed and applied to retinal image segmentation. In order to enhance relevant features and suppress irrelevant features, the Attention U-Net proposed by the attention mechanism is introduced, and the attention module is added before the features at each resolution of the encoder are concatenated with the corresponding features in the decoder. The output features of the encoder are readjusted and the Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) is proposed. This method combines residual connections and circular convolutions to replace the original submodules in U-Net. This method ensures the depth of the network while reducing the impact of gradient disappearance. The proposed fully automatic segmentation model based on spatial attention and residual recurrent convolution (Spatial Attentionand Residual Recurrental Convolution Neural Network based on U-Net, AttR2U-Net), this model integrates advanced mechanisms such as spatial attention, residual connection, recurrent convolution and normalization based on the classic U-Net to improve the segmentation performance; the U-Net network based on residual multi-scale fusion and attention skip layer (ResidualAttention U-Net, RAU-Net) adds an enhanced attention module to fuse multi-level features to solve the mirror reflection problem in cataract surgical instrument segmentation.

[0005] The U-Net series of models have achieved great advantages in the field of medical image segmentation, but most of the current research is on the segmentation of human organs or retinal blood vessels, and almost no results have been applied to near-infrared forearm blood vessel images. In addition, the performance of existing models is highly dependent on the size of the data set and the proportion of the effective area in the image. The segmentation effect of forearm blood vessel images with a small number of images, complex contours, and less dense distribution than fundus blood vessels is not ideal.

[0006] In the prior art, there is a lack of an efficient and accurate near-infrared forearm vascular image segmentation method based on a hybrid attention module. Summary of the invention

[0007] In order to solve the technical problems in the prior art that the performance of the existing model is highly dependent on the size of the data set and the proportion of the effective area in the image, and the segmentation effect of forearm blood vessel images with a small number of images, complex contours and sparse distribution is not ideal, the embodiment of the present invention provides a forearm blood vessel image segmentation method and device based on the HAMU-Net deep learning network. The technical solution is as follows:

[0008] On the one hand, a forearm vascular image segmentation method based on a HAMU-Net deep learning network is provided, the method is implemented by a forearm vascular image segmentation device, and the method includes:

[0009] The forearm blood vessel image is collected by a near-infrared acquisition device to obtain an original image;

[0010] Based on a Gaussian filter, the original image is preprocessed according to a preset background grayscale threshold to obtain a processed image;

[0011] Based on the Reinforcement-CLAHE technology, according to the processed image, a residual convolutional autoencoder is used to perform multi-stage image enhancement processing to obtain an enhanced image;

[0012] Construct the HAMU-Net blood vessel segmentation model based on the Hybrid Attention module, Triplet Attention module, Spatial attention module and U-Net network structure;

[0013] The enhanced image is input into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image.

[0014] On the other hand, a forearm vascular image segmentation device based on a HAMU-Net deep learning network is provided, and the device is applied to a forearm vascular image segmentation method based on a HAMU-Net deep learning network, and the device comprises:

[0015] An image acquisition module is used to acquire forearm blood vessel images through a near-infrared acquisition device to obtain an original image;

[0016] An image preprocessing module, used to preprocess the original image based on a Gaussian filter according to a preset background grayscale threshold to obtain a processed image;

[0017] An image enhancement module, for performing multi-stage image enhancement processing using a residual convolutional autoencoder based on the processed image to obtain an enhanced image based on the Reinforcement-CLAHE technology;

[0018] Model building module, used to build HAMU-Net blood vessel segmentation model based on Hybrid Attention module, Triplet Attention module, Spatial attention module and U-Net network structure;

[0019] The image segmentation module is used to input the enhanced image into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image.

[0020] On the other hand, a forearm vascular image segmentation device is provided, comprising: a processor; a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, any one of the forearm vascular image segmentation methods based on the HAMU-Net deep learning network is implemented.

[0021] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement any one of the forearm vascular image segmentation methods based on the HAMU-Net deep learning network.

[0022] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0023] The present invention proposes a forearm vascular image segmentation method based on the HAMU-Net deep learning network. Through multiple effective segmentation network architectures and modules, including encoder-decoder structure, U-Net backbone network, HybridAttention, Triplet Attention and Spatial Attention modules, as well as Deeplabv3 segmentation head and FCN auxiliary segmentation head, the HAMU-Net vascular segmentation network model has outstanding advantages such as multi-level feature extraction, multi-scale receptive field, and accurate segmentation, which improves the robustness and generalization ability of the model in the forearm vascular image segmentation task. The present invention is an efficient and accurate near-infrared forearm vascular image segmentation method based on a hybrid attention module. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1It is a flow chart of a forearm blood vessel image segmentation method based on a HAMU-Net deep learning network provided by an embodiment of the present invention;

[0026] Figure 2 is a schematic diagram of a target image enhancement process provided by an embodiment of the present invention;

[0027] Figure 3 1 is a schematic diagram of the structure of a HAMU-Net blood vessel segmentation network model that introduces a hybrid attention module provided in an embodiment of the present invention;

[0028] Figure 4 Schematic diagram of the structures of a Hybrid Attention module, a Triplet Attention module, and a Spatial attention module provided in an embodiment of the present invention;

[0029] Figure 5 It is a schematic diagram of the composition of the Deeplabv3 segmentation head and the FCN auxiliary segmentation head provided in an embodiment of the present invention;

[0030] Figure 6 It is a block diagram of a forearm blood vessel image segmentation device based on a HAMU-Net deep learning network provided by an embodiment of the present invention;

[0031] Figure 7 It is a structural schematic diagram of a forearm blood vessel image segmentation device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0033] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0034] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same. "of", "corresponding, relevant" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same.

[0035] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0036] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0037] The embodiment of the present invention provides a forearm vascular image segmentation method based on the HAMU-Net deep learning network, which can be implemented by a forearm vascular image segmentation device, which can be a terminal or a server. Figure 1 The flowchart of the forearm blood vessel image segmentation method based on the HAMU-Net deep learning network is shown. The processing flow of the method may include the following steps:

[0038] S1. Capture forearm blood vessel images through a near-infrared acquisition device to obtain an original image.

[0039] In a feasible implementation manner, the present invention uses a near-infrared acquisition device to obtain an initial raw image, and sets the image size in batches to 960×720 pixels to facilitate subsequent preprocessing calculations.

[0040] S2. Based on a Gaussian filter, the original image is preprocessed according to a preset background grayscale threshold to obtain a processed image.

[0041] In a feasible implementation, since the grayscale distribution of the foreground and background in the near-infrared forearm vascular image is quite different, the threshold segmentation method can be used to remove the interference caused by the impure color of the background plate, thereby improving the adaptability and generalization ability of the algorithm.

[0042] Optionally, based on a Gaussian filter, according to a preset background grayscale threshold, the original image is preprocessed to obtain a processed image, including:

[0043] Based on a preset background grayscale threshold, background noise is removed from the original image to obtain a first image;

[0044] Input the first image into a Gaussian filter to remove noise, thereby obtaining a second image;

[0045] Contrast stretching is performed on the second image to obtain a processed image.

[0046] In a feasible implementation, since the grayscale intensity of the background part of the original image is significantly higher than that of the foreground part, the background noise can be removed by calculating the segmentation threshold. By adjusting the value of the threshold T, the degree of separation between the foreground and the background can be controlled. The background removal method based on the threshold is as follows (1):

[0047] (1)

[0048] Among them, O(x, y) represents the pixel value of the output image O at the coordinate (x, y); I(x, y) represents the pixel value of the input image I at the coordinate (x, y); T represents the threshold value, which is a user-defined value; for each pixel point (x, y) of the input image I, if its grayscale value is greater than the threshold T, the corresponding position in the output image O is set to 255, indicating the foreground; otherwise, it is set to 0, indicating the background. Otherwise represents the case where the pixel value of the input image I at the coordinate (x, y) is less than the threshold T.

[0049] In order to further suppress the original image noise and improve the image quality, the present invention adopts a Gaussian filter to remove noise. The Gaussian filter is a linear filter that can effectively suppress salt and pepper noise and Gaussian noise and smooth the image.

[0050] Use the Gaussian function as the filter kernel and perform a convolution operation on the first image. The Gaussian function is a bell-shaped curve with a peak, which represents the weight of the pixel from the center pixel. The farther the pixel is from the center pixel, the lower the weight. Therefore, when smoothing an image, the Gaussian filter emphasizes the center pixel more and blurs the image less than the mean filter. When the mean μ = 0, the two-dimensional form of the Gaussian function is as follows (2):

[0051] (2)

[0052] Among them, σ is the standard deviation of the normal distribution. The value of the standard deviation determines the decay speed of the function, which can control the width of the Gaussian function and affect the smoothness of the filter. A smaller standard deviation will produce a narrower curve and retain more details, while a larger standard deviation will produce a wider curve, resulting in a stronger smoothing effect.

[0053] The near-infrared forearm vascular image after background removal and Gaussian filtering can show the basic vascular contours, but the contrast between the vascular area and other parts of the arm is poor. The use of contrast stretching algorithm can initially improve this problem.

[0054] Contrast stretching can stretch the grayscale levels of dark and bright areas in an image, making the details more prominent and helping to visually identify objects or features in the image. Its principle is to remap the grayscale range of the image and stretch the grayscale values ​​in the original image to a wider range, thereby increasing the difference between different grayscale levels and making the image look clearer and more suitable for subsequent analysis or visualization. The stretching process is expressed as follows:

[0055] (3)

[0056] Among them, I out (x, y) is the pixel value of the output image; I in (x, y) is the pixel value of the input image; I min and I max are the minimum and maximum pixel values ​​of the input image; O min and O max are the desired minimum and maximum pixel values ​​of the output image. The pixel values ​​of the input image are mapped to a new range in the output image, which is represented by O min and O max Therefore, the darker parts in the original image will become brighter, and the brighter parts will become dimmer, expanding the contrast range between the blood vessels and other tissue parts.

[0057] S3. Based on the Reinforcement-CLAHE technology, a residual convolutional autoencoder is used to perform multi-stage image enhancement processing on the processed image to obtain an enhanced image.

[0058] In a feasible implementation, Figure 2 As shown in the near-infrared image enhancement stage, the present invention proposes a multi-stage image enhancement method to effectively enhance blood vessels and veins, so as to prepare for the model to automatically and accurately segment blood vessels. The basic idea is to input the preprocessed image into the Residual Convolutional Autoencoder (RCAE) network and enhance the above results using the Reinforcement-CLAHE enhancement method. The output results of the Residual Convolutional Autoencoder and Reinforcement-CLAHE are combined in proportion to obtain an enhanced image.

[0059] Optionally, based on the Reinforcement-CLAHE technology, a residual convolutional autoencoder is used to perform multi-stage image enhancement processing according to the processed image to obtain an enhanced image, including:

[0060] According to the processed image, image enhancement processing is performed by a residual convolutional autoencoder to obtain a first enhanced image;

[0061] Based on the Reinforcement-CLAHE technology, the first enhanced image is subjected to image enhancement processing to obtain a second enhanced image;

[0062] The first enhanced image and the second enhanced image are weightedly fused to obtain an enhanced image.

[0063] In a feasible implementation, the Reinforcement-CLAHE of the present invention includes multi-scale limited contrast adaptive histogram equalization (CLAHE) and adaptive brightness contrast adjustment. By applying CLAHE on images of multiple scales, the large-scale structure and small-scale details of the image can be enhanced simultaneously, and excessive noise amplification can be suppressed. In this way, the model can extract vascular information to the maximum extent under the objective constraints of small sample data and realize pixel-level automatic segmentation. The method reasonably retains the contour features of vascular images at each stage, and experiments have proved that the effective combination ratio achieves the purpose of significantly enhancing the vascular area and suppressing noise interference. The calculation process of the multi-stage image enhancement method is shown as shown, and the calculation process of the linear weighting is shown in formula (4):

[0064] (4)

[0065] in, represents the target output image, I a represents the image enhanced only by the RCAE network, I b represents the image enhanced by RCAE and enhanced CLAHE, and η represents the weighted scaling coefficient.

[0066] Schematic diagram of the RCAE network structure. The RCAE network consists of three blocks, each of which includes a convolutional layer, an activation function layer, and a normalization layer. During the RCAE network training process, this paper considers the enhanced image to be a linear combination of the actual input image and the binary mask with vein annotations. If x represents the input forearm vascular image and h is the manually annotated vascular binary mask, the target image y is obtained as follows: y = αx +(1-α)h, where α is a fixed weight parameter.

[0067] Among them, x, h and y must have the same dimension. The generation process of the target image is shown as follows (5):

[0068] (5)

[0069] Among them, λ is a fixed weight parameter, I ori represents the input image, I mask Represents the binary mask image with vein annotations, O tar Represents the target image generated by the model. The three must have the same dimension.

[0070] S4. Construct the HAMU-Net blood vessel segmentation model based on the Hybrid Attention module, Triplet Attention module, Spatial attention module and U-Net network structure.

[0071] In one possible implementation, U-Net uses a unique U-shaped network structure, which is divided into two parts: downsampling (encoder) and upsampling (decoder) to effectively perform image segmentation. The task of downsampling is to gradually reduce the spatial resolution of the image while capturing the high-level features of the image, while upsampling expands the feature map to the resolution of the original input image through deconvolution operations. The encoder is the downsampling part of U-Net, which consists of a series of convolutional layers, pooling layers, and activation functions. These layers are used to extract the feature representation of the image. Each pooling layer reduces the spatial size of the feature map, but at the same time increases the abstraction level of the features. This allows the network to capture feature information at different scales, from lower-level textures to higher-level semantic information. The decoder part gradually increases the spatial resolution of the feature map through deconvolution operations. In addition, U-Net also uses a skip connection strategy to fuse the features in the encoder with the features in the decoder, which helps to retain the underlying feature information and improve the accuracy of the segmentation results.

[0072] Among them, the HAMU-Net blood vessel segmentation model includes the HAMU-Net backbone network, the skip connection part, the bottleneck layer and the output layer.

[0073] In a feasible implementation, Figure 3 As shown in the figure, the encoder part is responsible for extracting feature information from the input image. It contains a series of convolution and pooling layers, which gradually reduce the spatial size of the feature map and increase the number of channels to capture different features in the image, such as blood vessel edges, textures, and shapes. The output of the encoder is a set of high-dimensional feature representations that contain abstract feature information of the input image.

[0074] The decoder part restores the high-dimensional feature map extracted by the encoder to the same size as the input image. It contains a series of transposed convolutional layers that gradually increase the spatial size of the feature map and reduce the number of channels. The task of the decoder is to transform the high-dimensional feature map into pixel-level output for vascular image enhancement. Between the encoder and the decoder, a residual connection is introduced to generate a differential component that helps the network learn the difference between the input data and the target.

[0075] The size of the convolution kernel is an important factor in determining the receptive field. Since the size and position of the effective area in the vascular image and the width of the vascular contour may vary randomly within a given range, the size of the convolution kernel cannot be estimated intuitively. The convolution kernel of the encoder part gradually decreases from 9×9 to 3×3, while the size of the convolution kernel of the decoder part gradually increases.

[0076] A fully convolutional network (FCN) is added to the decoder near the output layer as an auxiliary segmentation head to increase the network's perception and improve segmentation performance. Figure 4 As shown in the figure, the detailed composition of the hybrid attention module, triplet attention module, spatial attention module, Deeplabv3Head and FCN Head in the model.

[0077] The HAMU-Net backbone network includes an encoder and a decoder; the encoder includes 9 3×3 convolutional layers and 4 2×2 downsampling layers; the decoder includes 9 3×3 convolutional layers, 4 2×2 upsampling layers and 1 Spatialattention module;

[0078] The jump connection part includes three first jump connections and a second jump connection; the first jump connection introduces the Hybrid Attention module;

[0079] The bottleneck layer is a Triplet Attention module; the Triplet Attention module includes the first branch, the second branch, and the third branch;

[0080] The output layer includes the main segmentation head and the auxiliary segmentation head; the main segmentation head includes 1 3×3 convolutional layer, 1 Deeplabv3 Head and 1 1×1 convolutional layer; the auxiliary segmentation head includes 1 3×3 convolutional layer and 1 FCN Head.

[0081] In a feasible implementation, the encoder side includes 9 3×3 convolutional layers and 4 2×2 downsampling layers; the decoder side includes 9 3×3 convolutional layers, 4 2×2 upsampling layers and 1 Spatial attention module; there are 4 jump connections, of which the first three introduce the Hybrid Attention module; the bottleneck layer is a Triplet Attention module; the output layer includes 2 segmentation head branches, the main segmentation head branch includes 1 3×3 convolutional layer, 1 Deeplabv3 Head and 1 1×1 convolutional layer, and the auxiliary segmentation head includes 1 3×3 convolutional layer and 1 FCN Head.

[0082] The Triplet Attention module is introduced in the part deep in the encoder and connected to the bottleneck layer to help the model capture global contextual information from the three dimensions (height, width, and channel) of the feature map. This is very useful for complex scenes in the blood vessel segmentation task, because blood vessels usually extend throughout the image and global information is needed to understand their structure.

[0083] Traditional Convolutional Block Attention Module (CBAM) and Squeeze-and-Excitation Networks (SENet) both require a certain number of learnable parameters to establish inter-channel dependencies. The Triplet Attention used in the present invention adopts a completely new approach, using an attention mechanism with almost no parameters to model channel attention and spatial attention.

[0084] The architecture of the Triplet Attention module consists of three parallel branches. The first branch is similar to the traditional CBAM and is used to establish channel attention. However, unlike CBAM, Triplet Attention does not require a large number of parameters to implement this branch. The second and third branches are responsible for capturing the cross-dimensional interactions between the channel dimension C and the spatial dimension H or W. These branches are implemented in a way that has almost no parameters, making the model more lightweight. The Triplet Attention module simply averages the outputs of the three branches to obtain comprehensive attention information in the channel and spatial dimensions. This approach not only improves the computational efficiency of the model, but also makes it more suitable for resource-constrained environments while maintaining good performance.

[0085] Using the Spatial Attention module in the shallow output layer of the decoder allows the model to focus on key spatial locations in the feature map, especially the segmentation of blood vessel edges and small branches. The core idea of ​​Spatial Attention is to generate a weight map by analyzing the information on the spatial dimensions (i.e., height and width) of the input feature map to enhance the model's attention to important spatial locations.

[0086] Assume that the input feature map has a dimension of (C × H × W), that is, a height of H, a width of W, and a number of channels of C. At this stage, the model has not yet selected the importance of a specific spatial position. Spatial Attention aggregates the spatial information of the feature map through a pooling operation and extracts globally representative information.

[0087] Perform global average pooling on the feature map in the channel dimension and average the channel values ​​at each position. The average feature response at each position can be obtained to form an average feature map in space. Perform maximum pooling on the feature map in the channel dimension to extract the maximum activation value in the channel at each position, which can capture the strongest response at each spatial position and form the maximum feature map in space.

[0088] After average pooling and maximum pooling, the two feature maps represent the global information of each spatial position. The results of maximum pooling and average pooling are added element by element to obtain a fused spatial feature map, which retains the average response value of each position and contains the maximum response value of each position.

[0089] The spatial feature map is further processed by convolution operation to generate weight values ​​for each spatial position. The convolution kernel size includes 7×7 and 5×7, and multiple receptive fields capture spatial context information. The value of each position in the feature map after convolution represents the importance of the position. After batch normalization and Sigmoid activation function, the spatial weight map is obtained. Finally, the generated spatial weight map is multiplied element by element with the original input feature map, so that the model weights important spatial positions and weakens unimportant positions. The calculation process is as follows (6) and (7):

[0090] (6)

[0091] (7)

[0092] Among them, X represents the input feature map, A represents the attention weight, Y represents the output feature map, σ is the Sigmoid activation function, ﹒ represents element multiplication, Conv is the convolution operation, and Pool is the pooling operation.

[0093] The Hybrid Attention module is used in the jump connection part. Triplet Attention and SpatialAttention run in series. The model can capture global multi-dimensional information and refine local spatial features, so that these features are strengthened both globally and locally, allowing the decoder to better utilize these detailed information. In the HybridAttention module, the attention weight and feature map calculation process are similar to the two types of attention introduced earlier, and will not be repeated here. The Hybrid Attention module attention weight and feature map calculation process are as follows (8) and (9):

[0094] (8)

[0095] (9)

[0096] Among them, M is the input feature map, N is the feature map after Triplet Attention, and P is the output feature map after SpatialAttention.

[0097] like Figure 5 As shown in the figure, the output layer contains two branches: Deeplabv3 Head and FCN Head. As the main segmentation head of the model, Deeplabv3 Head receives features from the backbone network and is responsible for generating the final segmentation results. One of the core principles of the Deeplabv3 network is the use of dilated convolution, also known as dilated convolution.

[0098] The traditional convolution operation is to perform convolution operations on the input based on a fixed convolution kernel size, while the dilated convolution introduces a dilation rate parameter, allowing the convolution kernel to operate on the input with different jump steps. This enables the network to capture information in the image with a multi-scale receptive field and better handle vascular structures of different sizes. The Atrous Spatial Pyramid Pooling (ASPP) module is a key component in the Deeplabv3 network. It includes multiple parallel dilated convolution branches, each with a different dilation rate. These branches extract features from the input image under different receptive fields and merge the results together. This enables the network to simultaneously capture the contextual information of vascular images at different scales, increase the network's perception, and improve the model's resolution of the boundaries and details of vascular contours.

[0099] This part includes multiple branches, of which three branches are modules with the same structure (named as module 3 in the figure), mainly composed of convolution layer (the convolution kernel can be set to different sizes according to the expansion rate, 3×3, 5×5, 7×7 in the present invention), batch normalization layer and linear rectification function layer. Another branch is composed of adaptive two-dimensional average pooling layer, module 1 and linear rectification function layer. The output of the above branches is passed through the splicing layer, module 1, random inactivation layer, module 3 and convolution layer 1×1 to obtain the final image segmentation result.

[0100] The FCN Head auxiliary head is another part of the decoder. In this model, it is used to perform tasks in parallel with the main segmentation head, but does not output results. It contains multiple 1×1 convolutional layers to adjust the number of channels to the required number of categories, and is connected to a batch normalization layer, a linear rectifier activation function, and a random dropout layer (one 1×1 convolutional layer, one normalization layer, and one linear rectifier function layer form a module, named Module 1 in the figure). During training, the difference between the output of the auxiliary head and the true label is used to calculate the loss and combined with the loss of the main segmentation head. This multi-segmentation head training strategy helps the network better learn the semantic information of the image. At the same time, it provides an auxiliary supervision signal that helps to propagate gradients faster and alleviate the gradient vanishing problem. In addition, it can make full use of the high-resolution features of the intermediate layers to improve the performance of the model, especially in terms of blood vessel edges and details.

[0101] S5. Input the enhanced image into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image.

[0102] In a feasible implementation, the present invention addresses the problems of low contrast between blood vessel lines and other tissues and unclear blood vessel contours in near-infrared blood vessel images. By effectively combining the advantages of traditional image enhancement methods and deep learning image enhancement methods, the processing effect is more suitable for the learning of vascular regional features by neural networks, and the vascular structure is significantly enhanced. The proposed HAMU-Net blood vessel segmentation model can efficiently fuse global and local information and reasonably utilize the correlation between pixels in the image. It solves the problem that the network is unable to achieve accurate segmentation of blood vessel images due to insufficient multi-scale feature learning under small sample conditions.

[0103] Optionally, the enhanced image is input into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image, including:

[0104] Input the enhanced image into the encoder for feature extraction to obtain an input image feature map;

[0105] According to the input image feature map, multi-dimensional features are captured through the bottleneck layer to obtain the output feature map;

[0106] Based on the jump connection part, the output feature map is input into the decoder for image generation to obtain the forearm blood vessel image;

[0107] According to the forearm vascular image, image segmentation is performed through the output layer to obtain a vascular segmentation image.

[0108] In a feasible implementation, the present invention introduces multiple attention modules as the feature extraction network of the blood vessel segmentation model based on the original U-Net, so that the model can better focus on the overall distribution state of the blood vessels and the distribution information of the blood vessel edges. The Deeplabv3 architecture containing the ASPP module is selected as the main segmentation head of the blood vessel segmentation model, so that the network can better learn context information, effectively capture the spatial relationship and semantic relationship between features, and improve the feature representation ability and information transmission effect.

[0109] Optionally, according to the input image feature map, multi-dimensional feature capture is performed through a bottleneck layer to obtain an output feature map, including:

[0110] The input image feature map is input into the first branch for convolution operation to obtain the first attention feature map; according to the first attention feature map, the Sigmoid function is used for data processing to obtain the first channel attention weight;

[0111] The input image feature map is input into the second branch for convolution operation to obtain the second attention feature map; according to the second attention feature map, the Sigmoid function is used to process the data to obtain the second channel attention weight;

[0112] The input image feature map is input into the third branch for convolution operation to obtain the third attention feature map; according to the third attention feature map, the Sigmoid function is used for data processing to obtain the third channel attention weight;

[0113] According to the first channel attention weight, the second channel attention weight and the third channel attention weight, the first attention feature map, the second attention feature map and the third attention feature map are weighted fused to obtain the output feature map.

[0114] In a feasible implementation, assume that the size of the input feature map is (C × H × W), that is, the height is H, the width is W, and the number of channels is C. The model will process the input feature map in parallel with three branches. The first branch pools the feature map in the channel dimension, retains the spatial information of height × width, generates a channel-compressed feature map, and uses a 7×7 convolution to process the pooled feature map to obtain the first attention feature map. After the Sigmoid function, the first channel attention weight is generated.

[0115] The second branch pools the feature map in the width dimension, retains the height × channel information, generates a width-compressed feature map, and uses a 7 × 7 convolution to process the pooled feature map to obtain the second attention feature map. After passing the Sigmoid function, the second channel attention weight is generated.

[0116] The third branch pools the feature map in the height dimension, retains the width × channel information, generates a highly compressed feature map, uses a 7 × 7 convolution to process the pooled feature map, obtains the third attention feature map, and generates the third channel attention weight after the Sigmoid function.

[0117] The attention weights of the second and third branches are dimensionally permuted, and the attention weights obtained by the three branches are element-wise multiplied with the input feature map to obtain the overall feature map. The feature maps of the three branches are fused and averaged to obtain the final output feature map. The output feature map calculation process is as follows (10), (11), (12), (13):

[0118] (10)

[0119] (11)

[0120] (12)

[0121] (13)

[0122] Among them, T a is the input feature map, σ is the Sigmoid function, p represents the dimension permutation, A channel is the attention weight of the first branch, A width is the attention weight of the second branch, A height is the attention weight of the third branch, Avg means averaging, and ﹒ means element-wise multiplication.

[0123] The present invention proposes a forearm vascular image segmentation method based on the HAMU-Net deep learning network. Through multiple effective segmentation network architectures and modules, including encoder-decoder structure, U-Net backbone network, HybridAttention, Triplet Attention and Spatial Attention modules, as well as Deeplabv3 segmentation head and FCN auxiliary segmentation head, the HAMU-Net vascular segmentation network model has outstanding advantages such as multi-level feature extraction, multi-scale receptive field, and accurate segmentation, which improves the robustness and generalization ability of the model in the forearm vascular image segmentation task. The present invention is an efficient and accurate near-infrared forearm vascular image segmentation method based on a hybrid attention module.

[0124] Figure 6 1 is a block diagram of a forearm vascular image segmentation device based on a HAMU-Net deep learning network according to an exemplary embodiment. The device is used in a forearm vascular image segmentation method based on a HAMU-Net deep learning network. Figure 6 The device includes an image acquisition module 610, an image preprocessing module 620, an image enhancement module 630, a model building module 640 and an image segmentation module 650. Among them:

[0125] An image acquisition module 610 is used to acquire a forearm blood vessel image through a near-infrared acquisition device to obtain an original image;

[0126] An image preprocessing module 620 is used to preprocess the original image based on a Gaussian filter according to a preset background grayscale threshold to obtain a processed image;

[0127] An image enhancement module 630 is used to perform multi-stage image enhancement processing using a residual convolutional autoencoder based on the processed image to obtain an enhanced image based on the Reinforcement-CLAHE technology;

[0128] A model building module 640 is used to build a HAMU-Net blood vessel segmentation model based on a Hybrid Attention module, a Triplet Attention module, a Spatial attention module and a U-Net network structure;

[0129] The image segmentation module 650 is used to input the enhanced image into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image.

[0130] Optionally, the image preprocessing module 620 is further used to:

[0131] Based on a preset background grayscale threshold, background noise is removed from the original image to obtain a first image;

[0132] Input the first image into a Gaussian filter to remove noise, thereby obtaining a second image;

[0133] Contrast stretching is performed on the second image to obtain a processed image.

[0134] Optionally, the image enhancement module 630 is further configured to:

[0135] According to the processed image, image enhancement processing is performed by a residual convolutional autoencoder to obtain a first enhanced image;

[0136] Based on the Reinforcement-CLAHE technology, the first enhanced image is subjected to image enhancement processing to obtain a second enhanced image;

[0137] The first enhanced image and the second enhanced image are weightedly fused to obtain an enhanced image.

[0138] Among them, the HAMU-Net blood vessel segmentation model includes the HAMU-Net backbone network, the skip connection part, the bottleneck layer and the output layer.

[0139] The HAMU-Net backbone network includes an encoder and a decoder; the encoder includes 9 3×3 convolutional layers and 4 2×2 downsampling layers; the decoder includes 9 3×3 convolutional layers, 4 2×2 upsampling layers and 1 Spatialattention module;

[0140] The jump connection part includes three first jump connections and a second jump connection; the first jump connection introduces the Hybrid Attention module;

[0141] The bottleneck layer is a Triplet Attention module; the Triplet Attention module includes the first branch, the second branch, and the third branch;

[0142] The output layer includes the main segmentation head and the auxiliary segmentation head; the main segmentation head includes 1 3×3 convolutional layer, 1 Deeplabv3 Head and 1 1×1 convolutional layer; the auxiliary segmentation head includes 1 3×3 convolutional layer and 1 FCN Head.

[0143] Optionally, the image segmentation module 650 is further configured to:

[0144] Input the enhanced image into the encoder for feature extraction to obtain an input image feature map;

[0145] According to the input image feature map, multi-dimensional features are captured through the bottleneck layer to obtain the output feature map;

[0146] Based on the jump connection part, the output feature map is input into the decoder for image generation to obtain the forearm blood vessel image;

[0147] According to the forearm vascular image, image segmentation is performed through the output layer to obtain a vascular segmentation image.

[0148] Optionally, the image segmentation module 650 is further configured to:

[0149] The input image feature map is input into the first branch for convolution operation to obtain the first attention feature map; according to the first attention feature map, the Sigmoid function is used for data processing to obtain the first channel attention weight;

[0150] The input image feature map is input into the second branch for convolution operation to obtain the second attention feature map; according to the second attention feature map, the Sigmoid function is used to process the data to obtain the second channel attention weight;

[0151] The input image feature map is input into the third branch for convolution operation to obtain the third attention feature map; according to the third attention feature map, the Sigmoid function is used for data processing to obtain the third channel attention weight;

[0152] According to the first channel attention weight, the second channel attention weight and the third channel attention weight, the first attention feature map, the second attention feature map and the third attention feature map are weighted fused to obtain the output feature map.

[0153] The present invention proposes a forearm vascular image segmentation method based on the HAMU-Net deep learning network. Through multiple effective segmentation network architectures and modules, including encoder-decoder structure, U-Net backbone network, HybridAttention, Triplet Attention and Spatial Attention modules, as well as Deeplabv3 segmentation head and FCN auxiliary segmentation head, the HAMU-Net vascular segmentation network model has outstanding advantages such as multi-level feature extraction, multi-scale receptive field, and accurate segmentation, which improves the robustness and generalization ability of the model in the forearm vascular image segmentation task. The present invention is an efficient and accurate near-infrared forearm vascular image segmentation method based on a hybrid attention module.

[0154] Figure 7 FIG. 1 is a schematic diagram of a forearm blood vessel image segmentation device provided by an embodiment of the present invention. Figure 7 As shown, the forearm blood vessel image segmentation device may include the above Figure 6 The forearm blood vessel image segmentation device based on the HAMU-Net deep learning network is shown. Optionally, the forearm blood vessel image segmentation device 710 may include a first processor 2001.

[0155] Optionally, the forearm blood vessel image segmentation device 710 may further include a memory 2002 and a transceiver 2003 .

[0156] The first processor 2001, the memory 2002 and the transceiver 2003 may be connected via a communication bus.

[0157] Combine the following Figure 7 The components of the forearm blood vessel image segmentation device 710 are specifically introduced as follows:

[0158] The first processor 2001 is the control center of the forearm blood vessel image segmentation device 710, and may be a processor or a general term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, such as one or more microprocessors (digital signal processors, DSPs), or one or more field programmable gate arrays (field programmable gate arrays, FPGAs).

[0159] Optionally, the first processor 2001 may perform various functions of the forearm blood vessel image segmentation device 710 by running or executing a software program stored in the memory 2002 and calling data stored in the memory 2002 .

[0160] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 7 CPU0 and CPU1 are shown in FIG.

[0161] In a specific implementation, as an embodiment, the forearm blood vessel image segmentation device 710 may also include multiple processors, such as Figure 7 The first processor 2001 and the second processor 2004 are shown in FIG. Each of these processors may be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). The processor here may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0162] The memory 2002 is used to store the software program for executing the solution of the present invention, and is controlled to be executed by the first processor 2001. The specific implementation method can refer to the above method embodiment, which will not be repeated here.

[0163] Optionally, the memory 2002 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 may be integrated with the first processor 2001, or may exist independently and access the first processor 2001 through the interface circuit ( Figure 7 (not shown) is coupled to the first processor 2001, which is not specifically limited in this embodiment of the present invention.

[0164] The transceiver 2003 is used to communicate with a network device or a terminal device.

[0165] Optionally, the transceiver 2003 may include a receiver and a transmitter ( Figure 7 The receiver is used to implement a receiving function, and the transmitter is used to implement a sending function.

[0166] Optionally, the transceiver 2003 may be integrated with the first processor 2001, or may exist independently and be transmitted through the interface circuit ( Figure 7 (not shown) is coupled to the first processor 2001, which is not specifically limited in this embodiment of the present invention.

[0167] It should be noted that Figure 7 The structure of the forearm blood vessel image segmentation device 710 shown in the figure does not constitute a limitation on the router, and the actual knowledge structure recognition device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0168] In addition, the technical effects of the forearm vascular image segmentation device 710 can refer to the technical effects of the forearm vascular image segmentation method based on the HAMU-Net deep learning network described in the above method embodiment, and will not be repeated here.

[0169] It should be understood that the first processor 2001 in the embodiment of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0170] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0171] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0172] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0173] In the present invention, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0174] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0175] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0176] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0177] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0178] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0179] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0180] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0181] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A forearm vascular image segmentation method based on HAMU-Net deep learning network, characterized in that: The method comprises: The forearm blood vessel image is collected by a near-infrared acquisition device to obtain an original image; Based on a Gaussian filter, the original image is preprocessed according to a preset background grayscale threshold to obtain a processed image; Based on the Reinforcement-CLAHE technology, according to the processed image, a residual convolutional autoencoder is used to perform multi-stage image enhancement processing to obtain an enhanced image; Construct the HAMU-Net blood vessel segmentation model based on the Hybrid Attention module, Triplet Attention module, Spatial attention module and U-Net network structure; The enhanced image is input into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image.

2. The forearm vascular image segmentation method based on the HAMU-Net deep learning network according to claim 1, characterized in that: The method of preprocessing the original image based on the Gaussian filter according to a preset background grayscale threshold to obtain a processed image includes: Based on a preset background grayscale threshold, background noise is removed from the original image to obtain a first image; Input the first image into a Gaussian filter to remove noise, thereby obtaining a second image; Contrast stretching is performed on the second image to obtain a processed image.

3. The forearm vascular image segmentation method based on the HAMU-Net deep learning network according to claim 1, characterized in that: The Reinforcement-CLAHE technology is based on the processed image and a residual convolutional autoencoder is used to perform multi-stage image enhancement processing to obtain an enhanced image, including: According to the processed image, performing image enhancement processing by a residual convolutional autoencoder to obtain a first enhanced image; Based on the Reinforcement-CLAHE technology, performing image enhancement processing on the first enhanced image to obtain a second enhanced image; The first enhanced image and the second enhanced image are weightedly fused to obtain an enhanced image.

4. The forearm vascular image segmentation method based on the HAMU-Net deep learning network according to claim 1, characterized in that: The HAMU-Net blood vessel segmentation model includes a HAMU-Net backbone network, a skip connection part, a bottleneck layer and an output layer.

5. The forearm vascular image segmentation method based on the HAMU-Net deep learning network according to claim 4, characterized in that: The HAMU-Net backbone network includes an encoder and a decoder; the encoder includes 9 3×3 convolutional layers and 4 2×2 downsampling layers; the decoder includes 9 3×3 convolutional layers, 4 2×2 upsampling layers and 1 Spatialattention module; The jump connection part includes three first jump connections and a second jump connection; the first jump connection introduces a Hybrid Attention module; The bottleneck layer is a Triplet Attention module; the Triplet Attention module includes a first branch, a second branch and a third branch; The output layer includes a main segmentation head and an auxiliary segmentation head; the main segmentation head includes a 3×3 convolutional layer, a Deeplabv3 Head and a 1×1 convolutional layer; the auxiliary segmentation head includes a 3×3 convolutional layer and a FCN Head.

6. The forearm vascular image segmentation method based on the HAMU-Net deep learning network according to claim 5, characterized in that: The step of inputting the enhanced image into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image comprises: Inputting the enhanced image into the encoder for feature extraction to obtain an input image feature map; According to the input image feature map, multi-dimensional feature capture is performed through the bottleneck layer to obtain an output feature map; Based on the jump connection part, the output feature map is input into the decoder to generate an image, so as to obtain a forearm blood vessel image; According to the forearm blood vessel image, image segmentation is performed through the output layer to obtain a blood vessel segmentation image.

7. The forearm vascular image segmentation method based on the HAMU-Net deep learning network according to claim 6, characterized in that: The step of capturing multi-dimensional features through the bottleneck layer according to the input image feature map to obtain an output feature map includes: Input the input image feature map into the first branch for convolution operation to obtain a first attention feature map; according to the first attention feature map, use the Sigmoid function to perform data processing to obtain a first channel attention weight; Input the input image feature map into the second branch for convolution operation to obtain a second attention feature map; according to the second attention feature map, use the Sigmoid function to perform data processing to obtain a second channel attention weight; Input the input image feature map into the third branch for convolution operation to obtain a third attention feature map; according to the third attention feature map, use the Sigmoid function to perform data processing to obtain a third channel attention weight; According to the first channel attention weight, the second channel attention weight and the third channel attention weight, the first attention feature map, the second attention feature map and the third attention feature map are weighted fused to obtain an output feature map.

8. A forearm vascular image segmentation device based on a HAMU-Net deep learning network, the forearm vascular image segmentation device based on a HAMU-Net deep learning network is used to implement the forearm vascular image segmentation method based on a HAMU-Net deep learning network as claimed in any one of claims 1 to 7, characterized in that: The device comprises: An image acquisition module is used to acquire forearm blood vessel images through a near-infrared acquisition device to obtain an original image; An image preprocessing module, used to preprocess the original image based on a Gaussian filter according to a preset background grayscale threshold to obtain a processed image; An image enhancement module, for performing multi-stage image enhancement processing using a residual convolutional autoencoder based on the processed image to obtain an enhanced image based on the Reinforcement-CLAHE technology; Model building module, used to build HAMU-Net blood vessel segmentation model based on Hybrid Attention module, Triplet Attention module, Spatial Attention module and U-Net network structure; The image segmentation module is used to input the enhanced image into the HAMU-Net blood vessel segmentation model to obtain a blood vessel segmentation image.

9. A forearm blood vessel image segmentation device, characterized in that: The forearm blood vessel image segmentation device comprises: processor; A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, which can be called by a processor to execute the method according to any one of claims 1 to 7.

Images