An AI-based ultrasound-guided injection site segmentation method and system

Abstract

This invention relates to an artificial intelligence-based method and system for segmenting ultrasound-guided injection sites, belonging to the field of image processing technology. It includes the following steps: acquiring an image of the ultrasound-guided injection site and annotating the image with a segmentation mask; preprocessing the annotated image; constructing an ultrasound-guided injection site segmentation model, the model including a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit; inputting the preprocessed image into the ultrasound-guided injection site segmentation model for training; optimizing the model using a loss function during training to obtain a trained model; inputting the image of the ultrasound-guided injection site to be detected into the trained model to obtain the ultrasound-guided injection site segmentation result. This invention can enhance the robustness of the ultrasound-guided injection site segmentation model and improve the accuracy of ultrasound-guided injection site segmentation.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an artificial intelligence-based ultrasound-guided injection site segmentation method and system. Background Technology

[0002] Ultrasound-guided injection is a treatment method that involves injecting medication into the lesion site with the assistance of ultrasound anatomical images. It is primarily used for deep lesion injections and precise injections, targeting narrow areas such as deep muscles, joint cavities, nerves, tendons, or epidural spaces. For example, in treating cervical radiculopathy with nerve root injections, the neck structure is complex, surrounded by several important tissues such as the carotid artery, thyroid gland, and vagus nerve. Furthermore, the nerve root is located deep within the neck, making experience-based injections less accurate and requiring a long learning curve. Ultrasound imaging can display anatomical structures in real time, including bones, muscles, and blood vessels, allowing doctors to more accurately locate the nerve root to be injected. Compared to traditional blind injection techniques, this method significantly improves injection accuracy, effectively reduces the risk of complications, and improves patient treatment outcomes. In addition, in shoulder and hip joint injections, where the injection sites are deep and the spaces are narrow, ultrasound guidance is also necessary to enhance injection precision.

[0003] While ultrasound-guided injection technology can improve the localization accuracy of deep and narrow lesions, its effectiveness is highly dependent on the accurate segmentation of key anatomical structures in ultrasound images. Existing segmentation techniques still have the following problems: the inherent low resolution and speckle noise of ultrasound images cause traditional U-Net networks to lose small structures during downsampling; local focusing preserves the target but noise suppression is insufficient, making it difficult to meet the accuracy requirements of injection navigation; the high similarity between the target and background in ultrasound images leads to pixel-level consistency constraints that blur the boundaries, and standard perturbations cannot simulate real acoustic noise, resulting in missegmentation under strong noise. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an ultrasound-guided injection site segmentation method and system based on artificial intelligence.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] This invention provides an artificial intelligence-based ultrasound-guided injection site segmentation method, comprising the following steps:

[0007] S1. Acquire images of the ultrasound-guided injection site and mark segmentation masks in the images; preprocess the marked images;

[0008] S2. Construct an ultrasound-guided injection site segmentation model, which includes a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit; input the preprocessed image into the ultrasound-guided injection site segmentation model for training;

[0009] S3. Optimize the model using the loss function during training to obtain a well-trained model;

[0010] S4. The image of the ultrasound-guided injection site to be detected is input into the trained model to obtain the segmentation result of the ultrasound-guided injection site.

[0011] Furthermore, ultrasound images often suffer from segmentation difficulties due to low contrast, noise interference, or blurring of small targets (such as nerves), and traditional preprocessing methods are difficult to generalize to different injection sites. This invention uses dynamic morphological enhancement to simulate the morphological changes of real anatomical structures, contrast perturbation to enhance the visibility of small targets (such as tendon edges), and adds stripe noise to simulate common ultrasound artifacts, making the model more stable in noisy environments. Step S1 specifically includes:

[0012] The locations of the segmentation mask include deep muscles, joint cavities, nerves, and tendons;

[0013] The preprocessing of the labeled images specifically involves:

[0014] Random structural elements are generated based on the morphology of the injection site and subjected to dilation or erosion operations to obtain a dynamically morphologically enhanced image. Nonlinear grayscale transformation is applied to perturb the contrast, thereby enhancing the contrast of small targets in the dynamically morphologically enhanced image to obtain a contrast-enhanced image. Random directional stripe noise is added to the contrast-enhanced image to obtain a preprocessed image.

[0015] Furthermore, the injection site features vary in scale (e.g., large areas of the joint cavity, small neural structures), and traditional U-Net networks ignore multi-scale information, leading to incomplete segmentation. The low resolution of ultrasound images easily causes feature loss, affecting the initial segmentation quality. This invention ensures that small targets (e.g., nerves) and large areas (e.g., muscles) are covered by embedding an adaptive multi-scale convolution module, improving the completeness of segmentation. It optimizes feature fusion through adaptive weight calculation, reducing redundant calculations and making the training process more efficient. The coarse segmentation module in step S2 includes an encoder and a decoder. The encoder uses a ResNet-34 network as the backbone network, including convolutional layers, a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer. The decoder includes a first upsampling layer, a second upsampling layer, a third upsampling layer, and a fourth upsampling layer. An adaptive multi-scale convolution module is embedded after upsampling at each layer.

[0016] The adaptive multi-scale convolution module includes a multi-scale convolution unit, an adaptive weight calculation unit, and a first feature fusion unit; in the skip connection, the encoded features output by the fourth downsampling layer... The first convolutional feature is obtained after multi-scale convolutional units. Second convolution features The first and second convolutional features are processed by an adaptive weight calculation unit to obtain adaptive weight coefficients. First convolutional feature Second convolution feature and adaptive weight coefficients The fused features are obtained after processing by the first feature fusion unit. The formula is expressed as follows:

[0017] ,

[0018] ,

[0019] ,

[0020] ,

[0021] in, Indicates the kernel size as Convolution operations; Indicates the kernel size as Convolution operations; This indicates a global average pooling operation; Represents the ReLU activation function; Indicates the kernel size as Convolution operations; The Sigmoid activation function is used. The fused feature is concatenated with the decoded feature output from the first upsampling layer to obtain the concatenated feature. This concatenated feature serves as the input to the second upsampling layer. Similarly, the encoded feature output from the third downsampling layer is processed by an adaptive multi-scale convolution module to obtain the second fused feature. This second fused feature is concatenated with the decoded feature output from the second upsampling layer to obtain the second concatenated feature, which serves as the input to the third upsampling layer. This process continues until the decoder operation is complete, resulting in a coarse segmentation mask. ;

[0022] Furthermore, ultrasound image boundaries are often discontinuous (e.g., tendons and surrounding tissues), and coarse segmentation results may have jagged edges; this invention can capture subtle edge changes through multi-scale edge-aware weighted fusion; the edge refinement unit in step S2 includes a Gaussian smoothing layer and a multi-scale edge-aware module:

[0023] The coarse segmentation mask The input is fed into the edge refinement unit and processed by a Gaussian smoothing layer to obtain a smooth coarse segmentation mask, as shown in the following formula:

[0024] ,

[0025] in, Represents the Gaussian kernel; This represents the convolution operation; The smooth coarse segmentation mask is input into the multi-scale edge sensing module, where it undergoes gradient calculation using Sobel operators at three different scales, resulting in three gradient maps. These three gradient maps are then weighted and fused to generate a comprehensive gradient map. The formula is as follows:

[0026] ,

[0027] in, , , These represent the first multi-scale weighting coefficient, the second multi-scale weighting coefficient, and the third multi-scale weighting coefficient, respectively. These weighting coefficients are adaptively adjusted through wavelet energy. , , These represent the first gradient plot, the second gradient plot, and the third gradient plot, respectively. This represents the integrated gradient plot.

[0028] Furthermore, the comprehensive gradient map and the coarse segmentation mask described in step S2 are fused together by the second feature fusion unit to obtain the fine segmentation mask; the formula is expressed as follows:

[0029] ,

[0030] in, This indicates a finely segmented mask.

[0031] Furthermore, the three Sobel operators with different scales respectively adopt a scale size of , , The Sobel core.

[0032] Furthermore, in the process of generating the comprehensive gradient map, the gradient map is decomposed into different frequency sub-bands using Haar wavelet transform.

[0033] Furthermore, the loss function in step S3 includes coarse segmentation loss. Fine-grained loss and edge consistency loss Specifically, it includes:

[0034] The coarse segmentation loss The Dice-cross-entropy mixed loss is expressed by the following formula:

[0035] ,

[0036] in, Represents the Dice loss function; Indicates the first balance coefficient; This represents the actual annotation mask used in the coarse segmentation stage; Represents the cross-entropy loss function;

[0037] The fine-segmentation loss The formula is expressed as follows:

[0038] ,

[0039] in, This represents the Focal loss function; This represents the actual annotation mask used in the fine segmentation stage; Indicates the edge gradient alignment strength; Represents the L1 norm; Represents the gradient operator;

[0040] The edge consistency loss The formula is expressed as follows:

[0041] ,

[0042] in, This represents the total number of pixels in the gradient map; This represents the vector value of the predicted gradient map at the i-th pixel. This represents the vector value of the true gradient map at the i-th pixel. Indicates the angle between the gradient directions; Represents the cosine function; This represents the L2 norm.

[0043] Furthermore, the formula for the total loss function of the model in step S3 is as follows:

[0044] ,

[0045] in, Represents the total loss function; , , These represent the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. Represents the predicted gradient map; This represents the true gradient map.

[0046] This invention also provides an artificial intelligence-based ultrasound-guided injection site segmentation system, comprising:

[0047] Data acquisition and preprocessing module: used to acquire images of the ultrasound-guided injection site and mark segmentation masks in the images; preprocess the marked images;

[0048] Model building module: used to build an ultrasound-guided injection site segmentation model, which includes a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit; the preprocessed image is input into the ultrasound-guided injection site segmentation model for training;

[0049] Model training module: During the training process, the model is optimized using a loss function to obtain a trained model;

[0050] Ultrasound-guided injection site segmentation module: This module is used to input the image of the ultrasound-guided injection site to be detected into the trained model to obtain the segmentation result of the ultrasound-guided injection site.

[0051] The advantages of this invention are:

[0052] This invention significantly improves the model's contrast perception and anti-interference ability for small targets such as deep muscles, joint cavities, nerves, and tendons in ultrasound images through a preprocessing strategy of dynamic morphological enhancement and random stripe noise injection. It employs a two-stage segmentation architecture: in the coarse segmentation stage, an adaptive multi-scale convolution module is introduced to accurately capture cross-scale anatomical structures by dynamically weighting and fusing multi-scale features; in the fine segmentation stage, wavelet energy-weighted multi-scale gradient fusion technology is combined to achieve sub-pixel-level edge refinement. Simultaneously, a triple loss supervision mechanism is designed to collaboratively optimize the overall segmentation accuracy and anatomically reasonable boundary continuity. This invention enhances the robustness of ultrasound-guided injection site segmentation models and improves the accuracy of ultrasound-guided injection site segmentation. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0054] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0055] Figure 2 This is a comparison image of the original image and the preprocessed image obtained by the method of this invention.

[0056] Figure 3 The training loss curves are for different combinations of loss functions in the method of this invention.

[0057] Figure 4 This is a comparison of the segmentation performance of the method of the present invention with that of existing methods. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example 1

[0060] In this embodiment, as Figure 1 As shown, this invention provides an ultrasound-guided injection site segmentation method based on artificial intelligence, the specific steps of which include:

[0061] S1. Obtain an image of the ultrasound-guided injection site and mark the segmentation mask in the image; preprocess the marked image.

[0062] Specifically, the locations of the labeled segmentation mask include deep muscles, joint cavities, nerves, and tendons;

[0063] The preprocessing of the labeled images specifically involves:

[0064] Dynamic morphological enhancement simulates the morphological changes of real anatomical structures. Random structural elements are generated based on the morphology of the injection site and subjected to dilation or erosion operations to obtain the dynamically enhanced image. The formula is as follows:

[0065] ,

[0066] in, This represents an image after dynamic morphological enhancement. Represents the original image. Structural elements representing morphological operations. This indicates an expansion operation. Indicates an etching operation;

[0067] Ultrasound images have a low signal-to-noise ratio, and small targets such as nerves and tendons have little gray-level difference from the background. Existing technologies mostly use linear gray-level transformation to enhance contrast, which cannot specifically improve the gray-level difference of small targets, resulting in small targets (such as the cross-section of a nerve bundle) being blurry and indistinguishable in low signal-to-noise ratio environments. This invention uses nonlinear gray-level transformation to perturb the contrast, enhancing the contrast of small targets in the dynamically morphologically enhanced image, thus obtaining a contrast-enhanced image.

[0068] Probe movement can produce artifacts. To improve the robustness of the model, random directional stripe noise is added to the contrast-enhanced image to simulate probe movement artifacts, resulting in a preprocessed image.

[0069] In one embodiment, such as Figure 2 As shown in the figure, the effect of the preprocessing method of the present invention is verified. It can be seen from the figure that the present invention can display the continuity of deep muscle boundaries by performing dynamic morphological enhancement through the expansion / erosion operation of random structural elements, and enhance the contrast of small targets by performing nonlinear grayscale transformation, so that the cross-section of nerve bundles can be clearly distinguished in a low signal-to-noise ratio environment, and can solve the problem of blurred edges in ultrasound images.

[0070] S2. Construct an ultrasound-guided injection site segmentation model, which includes a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit; the preprocessed image is input into the ultrasound-guided injection site segmentation model for training.

[0071] Specifically,

[0072] The coarse segmentation module includes an encoder and a decoder; the encoder uses a ResNet-34 network as the backbone network and includes convolutional layers, a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer; the decoder includes a first upsampling layer, a second upsampling layer, a third upsampling layer, and a fourth upsampling layer; an adaptive multi-scale convolutional module is embedded after upsampling at each layer;

[0073] Traditional U-Net networks use fixed-scale convolutional kernels, which cannot adapt to the mixed-scale structure of "large-scale deep muscle + small-scale nerve bundles" in ultrasound images. This results in incomplete extraction of large-scale structural features and loss of small-scale structural features, leading to low coarse segmentation accuracy. This invention embeds an adaptive multi-scale convolutional module after upsampling at each layer of the decoder. This adaptive multi-scale convolutional module includes a multi-scale convolutional unit, an adaptive weight calculation unit, and a first feature fusion unit. The multi-scale convolutional unit captures structural features at different scales, the adaptive weight calculation unit assigns high weights to important features, and the first feature fusion unit optimizes the feature fusion effect, enabling the coarse segmentation mask to accurately distinguish between deep muscle (large-scale) and nerve bundles (small-scale).

[0074] In the skip connection, the encoded features output from the fourth downsampling layer are processed by a multi-scale convolutional unit to obtain first and second convolutional features. These first and second convolutional features are then processed by an adaptive weight calculation unit to obtain adaptive weight coefficients. The first convolutional features, second convolutional features, and adaptive weight coefficients are then processed by a first feature fusion unit to obtain fused features. The formula is as follows:

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] in, Indicates the first convolutional feature; Indicates the kernel size as Convolution operations; This represents the encoded features of the output of the fourth downsampling layer; Indicates the second convolutional feature; Indicates the kernel size as Convolution operations; Indicates the adaptive weighting coefficient; This indicates a global average pooling operation; Represents the ReLU activation function; Indicates the kernel size as Convolution operations; This represents the Sigmoid activation function; The first upsampling layer outputs a fused feature; this fused feature is concatenated with the decoded feature output by the first upsampling layer to obtain a concatenated feature; this concatenated feature serves as the input to the second upsampling layer. Similarly, the encoded feature output by the third downsampling layer is processed by an adaptive multi-scale convolution module to obtain a second fused feature, which is then concatenated with the decoded feature output by the second upsampling layer to obtain a second concatenated feature, which serves as the input to the third upsampling layer; and so on. After the decoder operation is completed, a coarse segmentation mask is obtained. .

[0080] Existing segmentation models mostly focus on "region segmentation" and lack dedicated edge refinement units, resulting in blurred edges in the segmentation results (such as joint cavity boundaries and nerve edges). Since the edge accuracy at the injection site directly affects injection safety, blurred edges increase the risk of injection deviation. To achieve edge refinement, this invention designs an edge refinement unit, including a Gaussian smoothing layer and a multi-scale edge sensing module. The specific operation is as follows:

[0081] The coarse segmentation mask is input into the edge refinement unit and processed by a Gaussian smoothing layer to reduce isolated noise points in the coarse segmentation, resulting in a smooth coarse segmentation mask, as expressed in the following formula:

[0082] ,

[0083] in, Represents the Gaussian kernel; This represents the convolution operation; The smooth coarse segmentation mask is represented by three gradient maps calculated using Sobel operators at three different scales. These three gradient maps are then weighted and fused to generate a combined gradient map. The formula is as follows:

[0084] ,

[0085] in, , , These represent the first multi-scale weighting coefficient, the second multi-scale weighting coefficient, and the third multi-scale weighting coefficient, respectively. These weighting coefficients are adaptively adjusted through wavelet energy. , , These represent the scale size used. The first gradient map generated by the Sobel kernel, using a scale of The second gradient map generated by the Sobel kernel, using a scale of The third gradient map generated by the Sobel kernel; This invention uses a three-scale Sobel operator and wavelet energy adaptive weights to generate a comprehensive gradient map, which can capture edges of different thicknesses.

[0086] In the process of generating the comprehensive gradient map, the gradient map is decomposed into different frequency sub-bands by combining Haar wavelet transform, which enhances texture details and suppresses irrelevant background, making the edge information of the comprehensive gradient map more accurate.

[0087] The gradient map and the coarse segmentation mask are fused together by the second feature fusion unit to obtain the fine segmentation mask; the formula is as follows:

[0088] ,

[0089] in, This indicates a finely segmented mask.

[0090] S3. During the training process, the loss function is used to optimize the model to obtain a well-trained model.

[0091] Specifically, the loss function includes coarse segmentation loss. Fine-grained loss and edge consistency loss ;

[0092] coarse segmentation loss The Dice-cross-entropy mixed loss is expressed by the following formula:

[0093] ,

[0094] in, This represents the Dice loss function, used to address the problem of uneven injection site proportions. This represents the first balance coefficient. ; This represents the actual annotation mask used in the coarse segmentation stage; This represents the cross-entropy loss function, used to enhance small targets such as deep muscles and nerves;

[0095] Fine-segmentation loss The formula is expressed as follows:

[0096] ,

[0097] in, This represents the Focal loss function; This represents the actual annotation mask used in the fine segmentation stage; Indicates the edge gradient alignment strength. ; Represents the L1 norm; Represents the gradient operator;

[0098] Edge consistency loss The formula is expressed as follows:

[0099] ,

[0100] in, This represents the total number of pixels in the gradient map. This represents the vector value of the predicted gradient map at the i-th pixel. This represents the vector value of the true gradient map at the i-th pixel. Indicates the angle between the gradient directions; Represents the cosine function; This represents the L2 norm.

[0101] The formula for the total loss function is as follows:

[0102] ,

[0103] in, Represents the total loss function; This represents the first weighting coefficient. ; This represents the second weighting coefficient. ; This represents the third weighting coefficient. ; Represents the predicted gradient map; This represents the true gradient map.

[0104] In one embodiment, such as Figure 3 As shown, the training loss curves for different combinations of loss functions are presented. Ablation experiments are used to verify the rationality of the loss function design and reveal the synergistic effect of the triple loss supervision mechanism on model optimization. Curve 1: Blue line represents coarse segmentation loss; Curve 2: Green line represents a combination of coarse and fine segmentation loss; Curve 3: Red line represents a combination of coarse, fine, and edge consistency loss. Curve 3 has the fastest convergence speed, with the loss value decreasing to 0.07 after 50 iterations and stabilizing after 100 iterations with small fluctuations. Curve 1 has the slowest convergence speed, with the loss value reaching 0.12 after 120 iterations and large fluctuations after stabilization. Curve 2 only reaches a loss value of 0.07 after 80 iterations, and its convergence speed is also slower than Curve 3. This demonstrates that the edge consistency loss of this invention constrains the gradient direction deviation of boundary pixels by calculating the angle between the predicted gradient map and the true gradient map, avoiding the model from ignoring boundary continuity due to excessive focus on the target region, thereby accelerating the convergence speed and improving training stability.

[0105] S4. The image of the ultrasound-guided injection site to be detected is input into the trained model to obtain the segmentation result of the ultrasound-guided injection site.

[0106] Example 2

[0107] This invention uses a constructed dataset of ultrasound-guided injection site images as experimental data to compare the method of this invention with existing methods. Existing methods include a baseline model, U-Net++ network, At-U-Net network, and Trans-U-Net network. The baseline model is a standard U-Net network; the U-Net++ network nests dense connections and deep supervision on top of the U-Net network; the At-U-Net network introduces an attention mechanism into the skip connections of the U-Net network; and the Trans-U-Net network integrates the Transformer network and U-Net network architectures. During training, the image resolution was uniformly set to 224×224, the initial learning rate was set to 0.01, the training iterations were 100 epochs, and the batch size was 8. Cosine annealing was used for learning rate decay during the experiment. Dice coefficient, recall, and accuracy were used as evaluation metrics during the experiment. The experimental comparison results are shown in Table 1. Figure 4 .

[0108] Table 1. Experimental comparison between the method of the present invention and existing methods.

[0109]

[0110] As can be seen, in the ultrasound-guided injection site segmentation task, the proposed solution outperforms existing methods in all indicators. The proposed solution significantly improves the model's contrast perception and anti-interference ability for small targets such as deep muscles, joint cavities, nerves and tendons in ultrasound images through image preprocessing. The two-stage segmentation architecture enables accurate segmentation. Ultimately, the proposed solution improves the overall performance of the model in the ultrasound-guided injection site segmentation task, with all indicators outperforming existing methods.

[0111] Example 3

[0112] This embodiment provides an artificial intelligence-based ultrasound-guided injection site segmentation system, including:

[0113] Data acquisition and preprocessing module: used to acquire images of the ultrasound-guided injection site and mark segmentation masks in the images; preprocess the marked images;

[0114] Model building module: used to build an ultrasound-guided injection site segmentation model, which includes a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit; the preprocessed image is input into the ultrasound-guided injection site segmentation model for training;

[0115] Model training module: During the training process, the model is optimized using a loss function to obtain a trained model;

[0116] Ultrasound-guided injection site segmentation module: This module is used to input the image of the ultrasound-guided injection site to be detected into the trained model to obtain the segmentation result of the ultrasound-guided injection site.

[0117] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based ultrasound-guided injection site segmentation method, characterized in that, Includes the following steps: S1. Acquire images of the ultrasound-guided injection site and mark segmentation masks in the images; preprocess the marked images; S2. Construct an ultrasound-guided injection site segmentation model, which includes a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit. The preprocessed image is input into the ultrasound-guided injection site segmentation model for training; the specific process is as follows: The coarse segmentation module includes an encoder and a decoder; the encoder uses a ResNet-34 network as the backbone network and includes convolutional layers, a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer; the decoder includes a first upsampling layer, a second upsampling layer, a third upsampling layer, and a fourth upsampling layer; an adaptive multi-scale convolutional module is embedded after upsampling at each layer; The adaptive multi-scale convolution module includes a multi-scale convolution unit, an adaptive weight calculation unit, and a first feature fusion unit. In skip connections, the encoded features output by the fourth downsampling layer The first convolutional feature is obtained after multi-scale convolutional units. Second convolution features The first and second convolutional features are processed by an adaptive weight calculation unit to obtain adaptive weight coefficients. First convolutional feature Second convolution feature and adaptive weight coefficients The fused features are obtained after processing by the first feature fusion unit. The fused feature is concatenated with the decoded feature output from the first upsampling layer to obtain a concatenated feature; this concatenated feature serves as the input to the second upsampling layer. Similarly, the encoded feature output from the third downsampling layer is processed by an adaptive multi-scale convolution module to obtain a second fused feature, which is then concatenated with the decoded feature output from the second upsampling layer to obtain a second concatenated feature, which serves as the input to the third upsampling layer. This process continues until the decoder operation completes, resulting in a coarse segmentation mask. ; The edge refinement unit includes a Gaussian smoothing layer and a multi-scale edge sensing module: The coarse segmentation mask The input is fed into the edge refinement unit and processed by the Gaussian smoothing layer to obtain a smooth coarse segmentation mask. The smooth coarse segmentation mask is then input into the multi-scale edge perception module, where the gradients are calculated by the Sobel operator at three different scales to obtain three gradient maps. The three gradient maps are then weighted and fused to generate a comprehensive gradient map. The integrated gradient map and the coarse segmentation mask are fused together by the second feature fusion unit to obtain the fine segmentation mask; S3. During training, the model is optimized using a loss function to obtain a trained model; the loss function includes coarse segmentation loss. Fine-grained loss and edge consistency loss ; S4. The image of the ultrasound-guided injection site to be detected is input into the trained model to obtain the segmentation result of the ultrasound-guided injection site.

2. The method for segmenting injection sites based on artificial intelligence according to claim 1, characterized in that, Step S1 specifically includes: The locations of the segmentation mask include deep muscles, joint cavities, nerves, and tendons; The preprocessing of the labeled images specifically involves: Random structural elements are generated based on the morphology of the injection site and subjected to dilation or erosion operations to obtain a dynamically morphologically enhanced image. Nonlinear grayscale transformation is applied to perturb the contrast, thereby enhancing the contrast of small targets in the dynamically morphologically enhanced image to obtain a contrast-enhanced image. Random directional stripe noise is added to the contrast-enhanced image to obtain a preprocessed image.

3. The method for segmenting injection sites based on artificial intelligence according to claim 2, characterized in that, The fusion features are obtained in step S2. The formula is expressed as follows: ， ， ， ， in, Indicates the kernel size as Convolution operations; Indicates the kernel size as Convolution operations; This indicates a global average pooling operation; Represents the ReLU activation function; Indicates the kernel size as Convolution operations; This represents the Sigmoid activation function.

4. The method for segmenting injection sites based on artificial intelligence according to claim 3, characterized in that, The formula for obtaining the smooth coarse segmentation mask in step S2 is expressed as follows: , in, Represents the Gaussian kernel; This represents the convolution operation; The formula for generating a combined gradient map by weighted fusion of the three gradient maps is as follows: ， in, , , These represent the first multi-scale weighting coefficient, the second multi-scale weighting coefficient, and the third multi-scale weighting coefficient, respectively. These weighting coefficients are adaptively adjusted through wavelet energy. , , These represent the first gradient plot, the second gradient plot, and the third gradient plot, respectively. This represents the integrated gradient plot.

5. The method for segmenting an ultrasound-guided injection site based on artificial intelligence according to claim 4, characterized in that, The formula for obtaining the fine-segmented mask in step S2 is expressed as follows: ， in, This indicates a finely segmented mask.

6. The method for segmenting an ultrasound-guided injection site based on artificial intelligence according to claim 5, characterized in that, The three Sobel operators with different scales respectively adopt a scale of... , , The Sobel core.

7. The method for segmenting an ultrasound-guided injection site based on artificial intelligence according to claim 6, characterized in that, In the process of generating the comprehensive gradient map, the gradient map is decomposed into different frequency sub-bands using Haar wavelet transform.

8. The method for segmenting an ultrasound-guided injection site based on artificial intelligence according to claim 7, characterized in that, The loss function in step S3 specifically includes: The coarse segmentation loss The Dice-cross-entropy mixed loss is expressed by the following formula: ， in, Represents the Dice loss function; Indicates the first balance coefficient; This represents the actual annotation mask used in the coarse segmentation stage; Represents the cross-entropy loss function; The fine-segmentation loss The formula is expressed as follows: ， in, This represents the Focal loss function; This represents the actual annotation mask used in the fine segmentation stage; Indicates the edge gradient alignment strength; Represents the L1 norm; Represents the gradient operator; The edge consistency loss The formula is expressed as follows: ， in, This represents the total number of pixels in the gradient map; This represents the vector value of the predicted gradient map at the i-th pixel. This represents the vector value of the true gradient map at the i-th pixel. Indicates the angle between the gradient directions; Represents the cosine function; This represents the L2 norm.

9. The method for segmenting an ultrasound-guided injection site based on artificial intelligence according to claim 8, characterized in that, The formula for the total loss function of the model in step S3 is as follows: ， in, Represents the total loss function; , , These represent the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. Represents the predicted gradient map; This represents the true gradient map.

10. An AI-based ultrasound-guided injection site segmentation system, executing the AI-based ultrasound-guided injection site segmentation method as described in claim 1, characterized in that... include: Data acquisition and preprocessing module: used to acquire images of the ultrasound-guided injection site and mark segmentation masks in the images; Preprocess the labeled images; Model building module: used to build an ultrasound-guided injection site segmentation model, which includes a coarse segmentation module and a fine segmentation module; the coarse segmentation module adopts a U-Net network architecture, and the fine segmentation module includes an edge refinement unit and a second feature fusion unit; The preprocessed image is input into the ultrasound-guided injection site segmentation model for training. Model training module: Used to optimize the model using a loss function during training to obtain a trained model; Ultrasound-guided injection site segmentation module: This module is used to input the image of the ultrasound-guided injection site to be detected into the trained model to obtain the segmentation result of the ultrasound-guided injection site.

Images