Ultrasonic-guided femoral nerve auxiliary puncture method

Abstract

The invention relates to the technical field of image processing, and provides an ultrasound-guided femoral nerve auxiliary puncture method. The method comprises the following steps: S1, constructing a femoral nerve ultrasound image data set; s2, carrying out image enhancement processing; s3, constructing a training set and a test set; s4, adaptive multi-scale feature extraction, attention guidance feature enhancement, high and low frequency fusion feature extraction and feature splicing are carried out; s5, dynamic guiding attention is introduced to carry out feature coding and decoding processing, and a pixel-level segmentation result is obtained; and S6, identifying a femoral nerve position based on a pixel-level segmentation result. According to the scheme of the invention, multi-mechanism fusion feature extraction improves the accuracy of femoral nerve recognition and reduces background interference; self-adaptive multi-scale feature extraction is carried out, and comprehensive perception of femoral nerves is enhanced; high and low frequency fusion feature extraction is combined, so that the stable recognition capability of femoral nerves is enhanced; a femoral nerve area is enhanced through an attention mechanism, and the recognition stability is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to an ultrasonic-guided femoral nerve assisted puncture method. Background Art

[0002] Femoral nerve block, as an important means in regional anesthesia, is widely used in scenarios such as hip replacement, knee surgery, and emergency trauma treatment. Since it can effectively relieve postoperative pain, reduce the use of postoperative analgesics, and promote early patient activity, it has become an important part of modern perioperative management. Traditional femoral nerve block techniques usually rely on anatomical landmark positioning. However, there are significant anatomical differences among individual patients, and the covering of adipose tissue and muscle structures increases the difficulty of femoral nerve positioning, which may lead to problems such as puncture failure, improper drug injection, or femoral nerve injury.

[0003] Although the ultrasonic-guided femoral nerve block technique has significantly improved the accuracy of femoral nerve block operation, ultrasound can clearly show the relative position of the femoral nerve and surrounding anatomical structures, helping doctors accurately place the puncture needle under visual guidance, thereby reducing the incidence of complications. However, the quality of ultrasound images is closely related to the doctor's operation experience. For experienced doctors, ultrasonic-guided femoral nerve block can be easily completed, but for beginners or doctors with limited operation experience, there are still certain technical barriers. Especially in patients with unclear or atypical anatomical structures, the risk of misoperation still exists. In addition, with the popularization of minimally invasive surgery and day surgery, patients' demands for rapid recovery and painless postoperative experience are continuously increasing. How to further improve the accuracy and safety of femoral nerve block, and help doctors complete the operation faster and more accurately, has become an urgent problem to be solved in current clinical practice. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an ultrasonic-guided femoral nerve assisted puncture method, aiming to solve the problems mentioned in the background art.

[0005] In the scenario of femoral nerve block, the application of artificial intelligence technology can not only assist doctors in quickly locating the femoral nerve, but also real-time monitor the relative positions of the needle tip, femoral nerve, and blood vessels during the puncture process, providing dynamic risk prompts, and significantly improving the safety of puncture. In addition, with the help of visualization capabilities, doctors can obtain intuitive guidance on femoral nerve path planning, further optimize the puncture path, and reduce the operation difficulty.

[0006] The present invention provides an ultrasonic-guided femoral nerve assisted puncture method, including the following steps:

[0007] Step S1: Collect multi-view ultrasonic images of the patient's inguinal region and construct a femoral nerve ultrasonic image dataset therefrom, where the femoral nerve ultrasonic image dataset includes femoral nerve ultrasonic images;

[0008] Step S2: The femoral nerve ultrasound image is input into the femoral nerve annotation module to annotate the potential area of the femoral nerve, and then input into the image enhancement module for image enhancement processing to obtain the enhanced femoral nerve ultrasound image;

[0009] Step S3: Construct a training set and a test set based on the enhanced femoral nerve ultrasound image for training and testing;

[0010] Step S4: The enhanced femoral nerve ultrasound image is input into the multi-mechanism fusion feature extraction module for adaptive multi-scale feature extraction, attention-guided feature enhancement, high-low frequency fusion feature extraction, and feature splicing processing to obtain the multi-mechanism fusion feature extraction feature map;

[0011] Step S5: The multi-mechanism fusion feature extraction feature map is input into the femoral nerve ultrasound image segmentation model based on the SegNet architecture, and dynamic guidance attention is introduced for feature encoding and decoding processing to obtain the pixel-level segmentation result;

[0012] Step S6: Deploy the femoral nerve annotation module and the trained and tested image enhancement module, multi-mechanism fusion feature extraction module, and femoral nerve ultrasound image segmentation model based on the SegNet architecture to the hospital server background. After the femoral nerve ultrasound image of the patient is input into the hospital server background and processed, the position of the femoral nerve is identified based on the pixel-level segmentation result.

[0013] Further, step S1 includes the following steps:

[0014] Step S101: Use an ultrasound system equipped with a probe. In the brightness mode, scan the probe horizontally above the inguinal fold of the patient, and find the typical position of the patient's femoral nerve by adjusting the position of the probe;

[0015] Step S102: The ultrasound system dynamically scans the typical position of the patient's femoral nerve to collect video data. The video data covers the entire femoral nerve area of the patient, and then several femoral nerve ultrasound images are extracted from the video data. Each femoral nerve ultrasound image includes the complete femoral nerve anatomical area, and finally several femoral nerve ultrasound images are stored as the femoral nerve ultrasound image dataset.

[0016] Further, step S2 includes the following steps:

[0017] Step S201: Use annotation software to open the femoral nerve ultrasound image dataset, manually identify the femoral nerve area in each femoral nerve ultrasound image, use the annotation tool to draw a closed curve along the boundary of the femoral nerve in the femoral nerve area, and perform preliminary edge detection through the Sobel operator during the process of drawing the closed curve to obtain the preliminarily annotated femoral nerve ultrasound image;

[0018] Step S202: The preliminarily annotated femoral nerve ultrasound image is input into the image enhancement module for histogram equalization, contrast-limited adaptive histogram equalization, Gaussian filtering, and Laplacian edge enhancement processing to obtain the enhanced femoral nerve ultrasound image.

[0019] Further, step S3 is specifically as follows:

[0020] The femoral nerve ultrasound image labels collected in step S1 are paired with the enhanced femoral nerve ultrasound image in step S2, and are divided into a training set and a test set in a ratio of 8:2 in a random grouping manner for the training and testing of the multi-mechanism fusion feature extraction module and the femoral nerve ultrasound image segmentation model based on the SegNet architecture; the femoral nerve ultrasound image collected in step S1 is paired with the enhanced femoral nerve ultrasound image in step S2, where the enhanced femoral nerve ultrasound image is used as a label, and is divided into a training set and a test set in a ratio of 8:2 in a random grouping manner for the training and testing of the image enhancement module.

[0021] Further, in step S4, the adaptive multi-scale feature extraction and attention-guided feature enhancement processing are specifically as follows:

[0022] Step S401: Perform adaptive multi-scale feature extraction processing: The enhanced femoral nerve ultrasound image is sequentially input into convolutional layers, batch normalization layers, and sampling layers of different sizes for processing; among them, the enhanced femoral nerve ultrasound image is input into convolutional layers of different sizes for multi-scale feature extraction processing; the extracted multi-scale features are input into the batch normalization layer for feature standardization processing; the standardized features are input into the sampling layer for max pooling processing to retain significant features and obtain the femoral nerve ultrasound image with significant features retained.

[0023] Step S402: Perform attention-guided feature enhancement processing: The femoral nerve ultrasound image with significant features retained is sequentially input into a channel attention layer, a spatial attention layer, and a splicing layer for processing; among them, the femoral nerve ultrasound image with significant features retained is input into the channel attention layer, and useful features are enhanced through channel attention, and then input into the spatial attention layer, and the saliency of the femoral nerve region is improved through spatial attention, and finally input into the splicing layer to fuse the channel attention mechanism and the spatial attention mechanism to ensure the importance of features in different dimensions and obtain the femoral nerve ultrasound image with attention-guided feature enhancement.

[0024] Further, in step S4, high-frequency and low-frequency fusion feature extraction and feature splicing processing are performed to obtain a multi-mechanism fusion feature extraction feature map, specifically as follows:

[0025] Step S403: Perform high-low frequency fusion feature extraction processing: The femoral nerve ultrasound images with attention-guided feature enhancement are successively subjected to high-frequency information extraction, low-frequency information extraction, and feature fusion processing to obtain femoral nerve ultrasound images that fuse high-frequency and low-frequency information;

[0026] Step S404: Perform feature splicing processing: The femoral nerve ultrasound images that fuse high-frequency and low-frequency information are input into the feature splicing layer for feature splicing processing, and finally a feature map of multi-mechanism fusion feature extraction is obtained.

[0027] Furthermore, in step S5, the femoral nerve ultrasound image segmentation model based on the SegNet architecture includes: an encoder, a dynamic feature-guided decoder, and a Softmax activation function layer. Step S5 is specifically as follows:

[0028] Step S501: The encoder includes convolutional layers, batch normalization layers, and max pooling layers of the same scale. The feature map of multi-mechanism fusion feature extraction is input into the convolutional layers of the same scale of the encoder to extract local femoral nerve features, and then input into the batch normalization layer to normalize the feature distribution of each channel. Finally, it is input into the max pooling layer to reduce the resolution through max pooling;

[0029] Step S502: The dynamic feature-guided decoder includes a dynamic guidance attention layer, an anti-pooling layer, a transposed convolutional layer, and a feature adjustment layer. After being processed by the encoder, the feature map of multi-mechanism fusion feature extraction is input into the dynamic guidance attention layer of the dynamic feature-guided decoder to selectively restore features by calculating the dynamic correlation between the encoded features and the decoded features, and then input into the anti-pooling layer to calculate the upsampling result by combining the dynamic guidance attention to ensure that information is not lost. Then it is input into the transposed convolutional layer to further improve the resolution of the decoded features and restore the detailed information at the same time. Finally, it is input into the feature adjustment layer to further optimize the decoded features, ensure that the boundary of the femoral nerve is clear, and reduce the influence of background noise at the same time;

[0030] Step S503: After being processed by the dynamic feature-guided decoder, the feature map of multi-mechanism fusion feature extraction is input into the Softmax activation function layer for calculation to generate a pixel-level segmentation result.

[0031] Furthermore, step S6 is specifically as follows:

[0032] Step S601: Use the training set and test set constructed in step S3 to train and test the image enhancement module, the multi-mechanism fusion feature extraction module, and the femoral nerve ultrasound image segmentation model based on the SegNet architecture;

[0033] Step S602: Deploy the femoral nerve annotation module, the image enhancement module, the multi-mechanism fusion feature extraction module, and the femoral nerve ultrasound image segmentation model based on the SegNet architecture after training and testing to the hospital server background;

[0034] Step S603: Input the femoral nerve ultrasound image of the patient into the hospital server background for processing. Finally, in the femoral nerve ultrasound image segmentation model based on the SegNet architecture, classify and identify each pixel in the femoral nerve ultrasound image to determine the position of the femoral nerve in the femoral nerve ultrasound image, providing a basis for the doctor's femoral nerve puncture decision.

[0035] The present invention has the following technical effects:

[0036] (1) Through a series of image processing techniques, the present invention optimizes the quality of ultrasound images, improves the visualization effect of the femoral nerve, reduces background noise, and improves the recognition accuracy of subsequent AI models. By using histogram equalization and contrast-limited adaptive histogram equalization, the local and global contrast of ultrasound images is enhanced, making the low-gray femoral nerve region more prominent and improving the visibility of the femoral nerve. Contrast-limited adaptive histogram equalization can enhance details in local areas while avoiding excessive enhancement that leads to an increase in background noise, enabling the femoral nerve to be clearly recognized in a complex background. The artifacts and noise in ultrasound images are reduced, improving the image quality. Gaussian filtering is used to smooth the femoral nerve ultrasound image, removing high-frequency noise while retaining the main structural features of the femoral nerve tissue, improving the image quality. Combining Laplacian edge enhancement makes the boundary of the femoral nerve sharper, improving the edge resolution ability of ultrasound images and making the femoral nerve region more distinguishable before segmentation.

[0037] (2) By adopting multi-mechanism fusion feature extraction, key features of ultrasound images are extracted in the deep learning model to ensure the accurate recognition of the femoral nerve, improve the accuracy of femoral nerve recognition, and reduce background interference. Adaptive multi-scale feature extraction is used, with convolutional kernels of different sizes to enhance the overall perception of the femoral nerve and ensure the complete presentation of the femoral nerve region. Combining high-frequency and low-frequency fusion feature extraction, while retaining the overall shape (low-frequency information) and boundary details (high-frequency information) of the femoral nerve, enhances the stable recognition ability of the femoral nerve. The femoral nerve region is strengthened through the attention mechanism, improving the recognition stability.

[0038] (3) Adopt a femoral nerve ultrasound image segmentation model based on the SegNet architecture to finally achieve automatic pixel-level segmentation of the femoral nerve, ensure clear boundaries of the femoral nerve, and improve the accuracy of femoral nerve block. Improve the segmentation accuracy of the femoral nerve and ensure the integrity of the femoral nerve boundary. Perform pixel-level segmentation using the SegNet architecture, introduce dynamic features to guide the decoder in the decoding stage, optimize the feature recovery process, and make the boundary of the femoral nerve clearer. Through dynamic guided attention, ensure the optimal matching of encoded features and decoded features, improve the segmentation accuracy, and reduce the problem of blurred boundaries. Brief Description of the Drawings

[0039] By referring to the following drawings, the exemplary embodiments of the present invention can be more fully understood:

[0040] Figure 1 It is a processing flow chart of an ultrasound-guided femoral nerve assisted puncture method provided by an embodiment of the present invention.

[0041] Figure 2 It is a processing flow chart of an image enhancement module provided by an embodiment of the present invention.

[0042] Figure 3 It is a processing flow chart of a multi-mechanism fusion feature extraction module provided by an embodiment of the present invention.

[0043] Figure 4 It is a processing flow chart of a femoral nerve ultrasound image segmentation model based on the SegNet architecture provided by an embodiment of the present invention. Detailed Embodiments

[0044] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs; the terms used herein are only for the purpose of describing specific embodiments and do not limit the present invention.

[0046] An embodiment of the present invention provides an ultrasound-guided femoral nerve assisted puncture method, including the following steps:

[0047] Step S1: Collect multi-view ultrasound images of the patient's inguinal region and construct a femoral nerve ultrasound image dataset therefrom, where the femoral nerve ultrasound image dataset includes femoral nerve ultrasound images;

[0048] Step S2: The femoral nerve ultrasound image is input into the femoral nerve annotation module to annotate the potential area of the femoral nerve, and then input into the image enhancement module for image enhancement processing to obtain the enhanced femoral nerve ultrasound image, which can improve the segmentation accuracy of the femoral nerve in different environments.

[0049] Step S3: Based on the enhanced femoral nerve ultrasound image, a training set and a test set are constructed for training and testing.

[0050] Step S4: The enhanced femoral nerve ultrasound image is input into the multi-mechanism fusion feature extraction module for adaptive multi-scale feature extraction, attention-guided feature enhancement, high-low frequency fusion feature extraction, and feature splicing processing to obtain the multi-mechanism fusion feature extraction feature map.

[0051] Among them, the recognition ability of the femoral nerve boundary and overall morphology is improved through adaptive multi-scale perception. The attention mechanism is combined to guide feature enhancement, reduce background interference, and strengthen the target femoral nerve area. High-low frequency fusion feature extraction is performed to improve the stability of the ultrasound image and enhance the robustness of the segmentation model.

[0052] Step S5: The multi-mechanism fusion feature extraction feature map is input into the femoral nerve ultrasound image segmentation model based on the SegNet architecture, and dynamic guidance attention is introduced for feature encoding and decoding processing to obtain the pixel-level segmentation result.

[0053] Among them, the optimal mapping between the encoded feature and the decoded feature is calculated through dynamic guidance attention to ensure clear boundaries of the femoral nerve and reduce artifact interference at the same time.

[0054] Step S6: The femoral nerve annotation module, the trained and tested image enhancement module, the multi-mechanism fusion feature extraction module, and the femoral nerve ultrasound image segmentation model based on the SegNet architecture are deployed on the hospital server background. After the femoral nerve ultrasound image of the patient is input into the hospital server background and processed, the position of the femoral nerve is recognized based on the pixel-level segmentation result.

[0055] In some embodiments, Step S1 includes the following steps:

[0056] Step S101: Using an ultrasound system equipped with a probe, in the brightness mode, the probe is scanned horizontally above the patient's inguinal fold, and the typical position of the patient's femoral nerve is found by adjusting the probe position.

[0057] Among them, the Sonosite SII ultrasound system is used, equipped with an HFL38xi high-frequency linear probe with a frequency of 13 - 6 MHz and a frame rate of 30 frames per second. The B mode (Brighteness Mode) is set to display high-contrast images of anatomical structures; the probe scanning angle is maintained within the range of 15 - 30 degrees to ensure the best incidence of the ultrasonic beam on the femoral nerve area, enhance the reflection signal, and at the same time apply appropriate probe pressure to ensure good acoustic coupling and avoid image artifacts;

[0058] Step S102: The ultrasound system dynamically scans the typical position of the femoral nerve of the patient to collect video data. The video data covers the entire femoral nerve area of the patient, and then several femoral nerve ultrasound images are extracted from the video data. Each femoral nerve ultrasound image includes the complete femoral nerve anatomical area. Finally, several femoral nerve ultrasound images are stored as a femoral nerve ultrasound image dataset;

[0059] Among them, video data is collected for 10 seconds to ensure coverage of the entire femoral nerve area. A computer script is used to extract 30 frames of femoral nerve ultrasound images per second from the video. Each frame of femoral nerve ultrasound image contains the complete femoral nerve anatomical area, and several femoral nerve ultrasound images are stored as a femoral nerve ultrasound image dataset P original ={I1, I2, …, I n}, where: P original is the femoral nerve ultrasound image dataset, and I1, I2, …, I n are the first femoral nerve ultrasound image, the second femoral nerve ultrasound image, and the nth femoral nerve ultrasound image.

[0060] In some embodiments, step S2 includes the following steps:

[0061] Step S201: Use annotation software to open the femoral nerve ultrasound image dataset, manually identify the femoral nerve area in each femoral nerve ultrasound image, use an annotation tool to draw a closed curve along the femoral nerve boundary in the femoral nerve area, and perform preliminary edge detection through the Sobel operator during the process of drawing the closed curve to obtain a preliminarily annotated femoral nerve ultrasound image;

[0062] Among them, the preliminary edge detection by the Sobel operator includes the following operations:

[0063] The horizontal direction gradient formula is expressed as:

[0064]

[0065] The vertical direction gradient formula is expressed as:

[0066]

[0067] The gradient magnitude formula is expressed as:

[0068]

[0069] Wherein: is a convolution operation; G is a measure of the edge strength at a pixel; I is a femoral nerve ultrasound image; is a symbol for taking the derivative of the femoral nerve ultrasound image; is the symbol for taking the derivative in the horizontal direction; is the symbol for taking the derivative in the vertical direction; G x is the gradient of the femoral nerve ultrasound image in the horizontal direction; G y is the gradient of the femoral nerve ultrasound image in the vertical direction;

[0070] Then, Canny edge detection is adopted to further perform closing processing on the boundary, and the formula is expressed as:

[0071] R = {(x, y)|G(x, y)>T low , G(x, y)<T high};

[0072] Wherein: R is the preliminarily labeled femoral nerve region, that is, the set of detected edge pixel points; G(x, y) is the representation of the edge strength G at the pixel point at the pixel; T low is the set low threshold; T high is the set high threshold; (x, y) is the coordinate of the femoral nerve ultrasound image, where x is the abscissa of the femoral nerve ultrasound image and y is the ordinate of the femoral nerve ultrasound image;

[0073] Finally, preliminary femoral nerve region annotation is performed, and the formula is expressed as:

[0074] R = {(x, y)|I(x, y)∈[I min , I max , Area(R)≥A min};

[0075] Wherein: [I min , I max is the gray scale range of the femoral nerve region; Area(R) is the area of the femoral nerve region; A min is the set minimum femoral nerve region area threshold; I(x, y) is the pixel point located at (x, y);

[0076] After the femoral nerve ultrasound image I is processed by the femoral nerve annotation module, the preliminarily labeled femoral nerve region R is obtained, and its form is a binary mask, where the value range 1 represents the femoral nerve region and 0 represents the non-femoral nerve region, and the femoral nerve ultrasound image dataset P original ={I1, I2, …, I n}The preliminary labeled femoral nerve ultrasound image dataset obtained through preliminary edge detection Wherein: and are the first preliminary labeled femoral nerve ultrasound image, the second preliminary labeled femoral nerve ultrasound image, and the nth preliminary labeled femoral nerve ultrasound image, and R1, R2, and R n are the first preliminary labeled femoral nerve region, the second preliminary labeled femoral nerve region, and the nth preliminary labeled femoral nerve region;

[0077] Step S202: The preliminary labeled femoral nerve ultrasound image is input into the image enhancement module for histogram equalization, contrast-limited adaptive histogram equalization (CLAHE), Gaussian filtering, and Laplacian edge enhancement processing to obtain the enhanced femoral nerve ultrasound image; it can improve the visualization effect of the femoral nerve region, reduce background interference, enable the segmentation model to more accurately identify the position of the femoral nerve, thereby improving the success rate of ultrasound-guided femoral nerve puncture, reducing the operation difficulty, reducing the risk of accidental puncture during the operation, and enhancing the safety and effectiveness of femoral nerve block;

[0078] Among them, histogram equalization is to enhance the local details of the ultrasound image with low contrast by adjusting the distribution of pixel gray values, so as to improve the visibility of the femoral nerve region. Specifically:

[0079] Calculate the histogram within the preliminary labeled femoral nerve region R and equalize this region. The formula is expressed as:

[0080]

[0081] Where: T(k) is the transformed pixel gray value, L is the number of gray levels, and p R (q) is the normalized histogram probability density of the gray level q within the kth preliminary labeled femoral nerve region R k inside, k is the pixel value index after equalization transformation of the preliminary labeled femoral nerve region R; q is the pixel value index within the kth preliminary labeled femoral nerve region R k inside;

[0082] The formula for calculating the normalized histogram probability density is expressed as:

[0083]

[0084] Where: |R k | is the total number of pixels in the preliminary labeled femoral nerve region; δ(I * (x, y) = q) is the gray-scale transformation operation of the qth pixel point. When the gray value at the pixel (x, y) is equal to q, the value is 1, otherwise it is 0,

[0085] Performing histogram equalization within the femoral nerve region can enhance the brightness contrast between the femoral nerve region and the surrounding tissues, avoid the influence of background noise, and make the boundary of the femoral nerve more obvious;

[0086] Apply the histogram equalization transformation to the preliminarily annotated femoral nerve ultrasound image I * :

[0087]

[0088] where: T(I * (x,y)) is the pixel value of each pixel in the preliminarily annotated femoral nerve ultrasound image I * ; I HE (x,y) is the result of the histogram equalization transformation of the pixel at the pixel point located at (x,y);

[0089] That is, map the pixel value I i of each pixel within the i-th preliminarily annotated femoral nerve region R * (x,y) to the equalized gray value T(I * (x,y)), thereby enhancing the contrast of the femoral nerve;

[0090] Secondly, contrast-limited adaptive histogram equalization (CLAHE) is to perform CLAHE within the annotated femoral nerve region R to ensure the enhancement of details within the femoral nerve region, prevent the increase of background noise caused by global enhancement, thereby accurately improving the visibility of the femoral nerve, while keeping the background tissue stable and avoiding artifact interference. The formula is expressed as:

[0091]

[0092] where: I CLAHE represents the pixel value after CLAHE processing; C max is the contrast gain limit threshold; if(x,y)∈R i means if the pixel at the coordinate (x,y) belongs to R i ; otherwise means otherwise;

[0093] Thirdly, Gaussian filtering is to perform smoothing processing on the preliminarily annotated femoral nerve region R using a two-dimensional Gaussian kernel to reduce the high-frequency noise of the ultrasound image. The formula is expressed as:

[0094]

[0095] where: I smooth is the femoral nerve ultrasound image after Gaussian filtering; σ is the standard deviation, used to balance noise suppression and edge preservation; u is the abscissa offset of the filtering window; v is the ordinate offset of the filtering window;

[0096] By using Gaussian filtering, ultrasonic artifacts and noises contained in the preliminarily labeled femoral nerve region can be removed, while keeping the boundary of the femoral nerve clear and avoiding affecting the texture features of the background region;

[0097] Finally, Laplacian edge enhancement is adopted. That is, the Laplacian operator is used to enhance the edge in the preliminarily labeled femoral nerve region, and the formula is expressed as:

[0098]

[0099] Among them, the Laplace transform is calculated as follows:

[0100]

[0101] In the formula: I sharp is the femoral nerve ultrasound image after Laplacian edge enhancement; is the Laplace transform;

[0102] After Laplacian edge enhancement, the boundary of the femoral nerve can be clearly visible, while not affecting the structure of the background tissue and avoiding unnecessary misidentification;

[0103] After processing based on histogram equalization, CLAHE, Gaussian filtering and Laplacian edge enhancement, the enhanced femoral nerve ultrasound image dataset P = {(I′1, R1), (I′2, R2), …, (I′ n , R n )} is obtained. I′1, I′ i and I′ n are the first enhanced femoral nerve ultrasound image, the second enhanced femoral nerve ultrasound image and the nth enhanced femoral nerve ultrasound image. Compared with the preliminarily labeled femoral nerve ultrasound image, the enhanced femoral nerve ultrasound image has higher contrast and clearer femoral nerve boundary.

[0104] In some embodiments, step S3 is specifically as follows:

[0105] Pair the femoral nerve ultrasound image labels collected in step S1 with the enhanced femoral nerve ultrasound images in step S2, and divide them into a training set and a test set in a random grouping manner according to a ratio of 8:2 for the training and testing of the multi-mechanism fusion feature extraction module and the femoral nerve ultrasound image segmentation model based on the SegNet architecture; pair the femoral nerve ultrasound images collected in step S1 with the enhanced femoral nerve ultrasound images in step S2, where the enhanced femoral nerve ultrasound image is used as the label, and divide them into a training set and a test set in a random grouping manner according to a ratio of 8:2 for the training and testing of the image enhancement module.

[0106] In some embodiments, in step S4, adaptive multi-scale feature extraction and attention-guided feature enhancement processing are performed. Among them, for adaptive multi-scale feature extraction, convolutional layers of different sizes are used to simultaneously capture local details and overall contours to ensure the complete presentation of the femoral nerve region. For attention-guided feature enhancement, channel attention and spatial attention are combined to enhance the saliency of the femoral nerve region and suppress background interference.

[0107] Specifically:

[0108] Step S401: Perform adaptive multi-scale feature extraction processing: The enhanced femoral nerve ultrasound images are sequentially input into convolutional layers, batch normalization layers, and sampling layers of different sizes for processing. Among them, the enhanced femoral nerve ultrasound images are input into convolutional layers of different sizes for multi-scale feature extraction processing. The extracted multi-scale features are input into the batch normalization layer for feature standardization processing. The standardized features are input into the sampling layer for max pooling processing to retain significant features and obtain femoral nerve ultrasound images with significant features retained.

[0109] Among them, in the ultrasound image, the morphology, size, and boundary information of the femoral nerve vary due to individual differences among patients. To capture this information at different levels, the enhanced femoral nerve ultrasound images are sequentially input into convolutional layers of different sizes for multi-scale feature extraction processing. The convolutional layers of different sizes include 3×3 convolutional layer, 5×5 convolutional layer, and 7×7 convolutional layer, which is expressed by the formula:

[0110] F MSC = ReLU(W3 * I′ + W5 * I′ + W7 * I′ + b);

[0111] In the formula: I′ is the enhanced femoral nerve ultrasound image; W3, W5, and W7 respectively represent the convolutional kernel parameter weights of 3×3, 5×5, and 7×7, b is the bias term, ReLU is the ReLU non-linear activation function, and F MSC is the output feature map of the convolutional layer of different sizes;

[0112] Since the femoral nerve ultrasound images of different patients have different brightness distributions, direct input may lead to unstable training. Then batch normalization is used to perform feature standardization on the output features of the convolutional layer, which is expressed by the formula:

[0113]

[0114] In the formula: μ is the feature mean; σ is the feature standard deviation; γ and β are both trainable parameters; is the output feature of the batch normalization layer for adaptive multi-scale feature extraction;

[0115] To reduce the computational complexity while retaining the most significant features, the standardized features are input into the sampling layer, and dimensionality reduction is performed through max pooling, which is expressed by the formula:

[0116]

[0117] In the formula: k represents the pooling window size; s represents the stride, and F pool is the output feature map of the sampling layer, and MaxPool is the max pooling operation.

[0118] Step S402: Perform attention-guided feature enhancement processing: The femoral nerve ultrasound images retaining significant features are sequentially input into the channel attention layer, the spatial attention layer, and the splicing layer for processing; among them, the femoral nerve ultrasound images retaining significant features are input into the channel attention layer, and through channel attention, useful features are enhanced, and then input into the spatial attention layer. Through spatial attention, the saliency of the femoral nerve region is improved. Finally, it is input into the splicing layer to fuse the channel attention mechanism and the spatial attention mechanism to ensure the importance of features in different dimensions, and obtain the femoral nerve ultrasound image with attention-guided feature enhancement; the attention mechanism is used to enhance the femoral nerve region information while reducing background interference;

[0119] Specifically:

[0120] In the channel attention layer, the channel attention mechanism is used to help automatically learn which feature channels are important for femoral nerve recognition, so as to enhance useful features and suppress irrelevant information. The formula is expressed as:

[0121] F CA = sigmoid(W c *F pool );

[0122] In the formula: W c is the channel attention weight, sigmoid is the sigmoid activation function used to normalize the attention weight, and F CA is the output feature map of the channel attention layer;

[0123] In the spatial attention layer, the spatial attention mechanism is used to improve the saliency of the femoral nerve region in the femoral nerve ultrasound image, which is conducive to focusing on the target region. The formula is expressed as:

[0124] F SA = sigmoid(W s *F pool );

[0125] In the formula: W s is the spatial attention weight, and F SA is the output feature map of the spatial attention layer.

[0126] In the splicing layer, the channel attention mechanism and the spatial attention mechanism are fused to ensure the importance of the femoral nerve features in different dimensions, which is expressed by the formula:

[0127] F AGFE =F CA *F pool +F SA *F pool ;

[0128] In the formula: F AGFE is the femoral nerve ultrasound image with attention-guided feature enhancement;

[0129] In some embodiments, in step S4, high-low frequency fusion feature extraction and feature splicing processing are performed. Among them, high-low frequency fusion feature extraction is to separate high-frequency edge information and low-frequency morphological information to improve the stability and robustness of the femoral nerve features; it can optimize the input features of the enhanced femoral nerve ultrasound image and provide a more accurate and robust femoral nerve segmentation basis for the femoral nerve ultrasound image segmentation model, thereby improving puncture safety and operation success rate;

[0130] Specifically:

[0131] Step S403: Perform high-low frequency fusion feature extraction processing: The femoral nerve ultrasound image with attention-guided feature enhancement is sequentially subjected to high-frequency information extraction, low-frequency information extraction, and feature fusion processing to obtain a femoral nerve ultrasound image that fuses high-frequency information and low-frequency information; after attention-guided feature enhancement, the high-frequency information (femoral nerve edge details) and low-frequency information (overall femoral nerve morphology) of the fused femoral nerve ultrasound image are fused;

[0132] Specifically:

[0133] The femoral nerve edge information in the femoral nerve ultrasound image belongs to the high-frequency component. The high-frequency information is extracted by Gaussian filtering, and the formula is expressed as:

[0134] I h =F AGFE -G σ *F AGFE ;

[0135] In the formula: G σ is the Gaussian smoothing filter for separating high-frequency information; I h is the high-frequency information;

[0136] The overall structural information of the femoral nerve in the femoral nerve ultrasound image belongs to the low-frequency component. The low-frequency information is extracted by Gaussian filtering, and the formula is expressed as:

[0137] I l =G σ *F AGFE ;

[0138] Where: I l is the low-frequency information;

[0139] By weighted fusion, the high-frequency information and the low-frequency information are fused, and the formula is expressed as:

[0140] F HLFF =W h I h +W l I l ;

[0141] Where: F HLFF is the feature map that fuses the high-frequency information and the low-frequency information; W h is the weight of the high-frequency information; W l is the weight of the low-frequency information;

[0142] The femoral nerve ultrasound image that fuses the high-frequency and low-frequency information is obtained, and the formula is expressed as:

[0143] P HLFF ={(F HLFF , R i )|(F AGFE , R i ) ∈ P AGFE};

[0144] Where: P HLFF is the femoral nerve ultrasound image dataset that fuses the high-frequency information and the low-frequency information; P AGFE is the femoral nerve ultrasound image dataset with attention-guided feature enhancement;

[0145] Step S404: Perform feature splicing processing: The femoral nerve ultrasound image that fuses the high-frequency and low-frequency information is input into the feature splicing layer for feature splicing processing, and finally a multi-mechanism fusion feature extraction feature map is obtained; specifically:

[0146] In the feature splicing layer, the features are spliced, and finally a multi-mechanism fusion feature extraction feature map after feature extraction is obtained. The formula is expressed as:

[0147] F final =F AGFE +F HLFF ;

[0148] Where: F final is the multi-mechanism fusion feature extraction feature map;

[0149] The final multi-mechanism fusion feature extraction feature map dataset P feature , the formula is expressed as:

[0150] P feature ={(F final , R1), (Ffinal , R2), …, (F final , R n )};

[0151] In some embodiments, to improve the accuracy and robustness of ultrasound-guided femoral nerve segmentation, based on the SegNet architecture, a Dynamic Feature Guided Decoder (DFG-Decoder) is set up to optimize the feature recovery ability during the decoding process. The SegNet structure is used to segment the feature map of multi-mechanism fusion features. At the same time, a dynamic feature guided decoder is introduced in the decoding stage. The optimal mapping between the encoded features and the decoded features is calculated through Dynamic Guided Attention (DGA), and combined with the Feature Adjustment Layer (FRL) to ensure clear femoral nerve boundaries while reducing artifact interference;

[0152] In step S5, the femoral nerve ultrasound image segmentation model based on the SegNet architecture includes: an encoder, a dynamic feature guided decoder, and a Softmax activation function layer. Step S5 is specifically as follows:

[0153] Step S501: The encoder includes convolutional layers, batch normalization layers, and max pooling layers of the same scale. The feature map of multi-mechanism fusion features is input into the convolutional layers of the same scale of the encoder to extract local femoral nerve features, and then input into the batch normalization layer to normalize the feature distribution of each channel. Finally, it is input into the max pooling layer to reduce the resolution through max pooling;

[0154] Among them, the convolutional layers of the same scale include three 3×3 convolutional layers, which can extract local femoral nerve features and improve the boundary resolution ability. By stacking convolutional layers, the network depth is increased to learn more complex features. The formula is expressed as:

[0155]

[0156] In the formula: W l is the convolutional kernel weight of the l-th layer, learning the edge and morphological features of the femoral nerve region; b l represents the bias term, improving the flexibility of the model; is the output feature map of the convolutional layer of the same scale; is the feature map of multi-mechanism fusion features output by the (l - 1)-th layer in the stacked convolutional layers.

[0157] In the batch normalization layer, the feature distribution of each channel is normalized to accelerate model convergence and improve training stability. The formula is expressed as:

[0158]

[0159] In the formula: and θ are the mean and standard deviation of the current batch input, and τ and ∈ represent trainable scaling and translation parameters; The output feature map of the batch normalization layer of the encoder;

[0160] Finally, in the max pooling layer, the resolution of the output feature map of the batch normalization layer of the encoder is reduced through max pooling, reducing the computational complexity while retaining the most significant feature information. The formula is expressed as:

[0161]

[0162] In the formula: k is the size of the pooling window, s is the stride. In the encoder stage, maxpool is the max pooling operation, is the output feature map of the max pooling layer;

[0163] The encoder output feature map F encoded , and the formula is expressed as:

[0164]

[0165] Step S502: The dynamic feature-guided decoder includes a dynamic guidance attention layer, an unpooling layer, a transposed convolution layer, and a feature adjustment layer. After the multi-mechanism fusion feature extraction feature map is processed by the encoder, it is input into the dynamic guidance attention layer of the dynamic feature-guided decoder. The dynamic correlation between the encoded features and the decoded features is calculated to selectively restore the features, and then input into the unpooling layer. The upsampling result is calculated by combining the dynamic guidance attention to ensure that the information is not lost. Then it is input into the transposed convolution layer to further increase the resolution of the decoded features and restore the detailed information. Finally, it is input into the feature adjustment layer to further optimize the decoded features, ensure that the femoral nerve boundary is clear, and reduce the influence of background noise at the same time. The dynamic feature-guided decoder can dynamically adjust the restored features during the decoding process, improve the accuracy of femoral nerve segmentation, calculate the optimal feature mapping through dynamic guidance attention, and perform boundary optimization in combination with the feature adjustment layer;

[0166] Specifically:

[0167] In the dynamic guidance attention layer, calculate the dynamic correlation between the encoded features and the decoded features, select the optimal restored features, and adjust the restored information in real time during the decoding process to avoid feature loss caused by fixed pooling indices;

[0168] According to the encoder output feature map F encoded , calculate the dynamic guidance attention matrix A DGA , and the formula is expressed as:

[0169] A DGA = sigmoid(W d *F encoded +W u *F decoded );

[0170] Among them, W d and W u are trainable weight matrices for calculating feature importance; F decoded is the feature recovered during the decoding process;

[0171] Then, in the unpooling layer, combined with the dynamic guidance attention matrix A DGA , calculate the optimal upsampling result to ensure that information is not lost during the decoding process. The formula is expressed as:

[0172]

[0173] In the formula: represents the decoded feature of the decoded feature output by the (l + 1)-th deconvolution layer in the decoder, is the output feature map of the unpooling layer; P l is the max pooling index; Unpool is the reference of the unpooling layer;

[0174] In the deconvolution layer, further improve the resolution of the decoded feature and at the same time restore the detailed information. The formula is expressed as:

[0175]

[0176] In the formula: is the deconvolution kernel weight; is the feature map output by the l-th deconvolution layer in the decoder; b l is the bias of the l-th layer;

[0177] Finally, in the feature adjustment layer, further optimize the decoded feature to ensure that the femoral nerve boundary is clear and at the same time reduce the influence of background noise. The formula is expressed as:

[0178]

[0179] In the formula: W f is the trainable convolution kernel of the feature adjustment layer; b f is the bias term; is the output feature map of the feature adjustment layer;

[0180] Step S503: After the feature map of multi-mechanism fusion feature extraction is processed by the dynamic feature guidance decoder, it is input into the Softmax activation function layer for calculation to generate a pixel-level segmentation result;

[0181] Specifically:

[0182]

[0183] In the formula: P(Y (x,y)= c|X) represents the probability that the pixel (x, y) belongs to the category c (femoral nerve / background); F decoded (x, y, c) In the decoder feature adjustment layer, it is the feature value corresponding to the category c for the pixel (x, y); c is the category index, 0 represents the background, and 1 represents the femoral nerve; is the final segmentation result; exp is the exponential function with the natural constant e as the base; is the index position operator of the maximum value in the return sequence;

[0184] Since the femoral nerve area accounts for a relatively small proportion in the ultrasound image, the Dice Loss is used to optimize the model, and the formula is expressed as:

[0185]

[0186] In the formula: is the predicted femoral nerve area mask prediction result; is the smoothing factor; to prevent division by zero error; L Dice is the Dice Loss value;

[0187] Calculate the probability P corresponding to the femoral nerve for each pixel point in the feature map of the multi-mechanism fusion feature extraction seg , and the formula is expressed as:

[0188]

[0189] In the formula: P seg is the probability corresponding to the femoral nerve for each pixel point in the feature map of the multi-mechanism fusion feature extraction. Classify the points with a probability greater than 0.5 in P seg as the pixel points that make up the femoral nerve, and obtain the pixel-level segmentation result; is the femoral nerve area mask prediction result corresponding to the first femoral nerve ultrasound image, is the femoral nerve area mask prediction result corresponding to the second femoral nerve ultrasound image, is the femoral nerve area mask prediction result corresponding to the nth femoral nerve ultrasound image; F refined is the feature map of the multi-mechanism fusion feature extraction.

[0190] In some embodiments, step S6 is specifically:

[0191] Step S601: Use the training set and test set constructed in step S3 to train and test the image enhancement module, the multi-mechanism fusion feature extraction module, and the femoral nerve ultrasound image segmentation model based on the SegNet architecture;

[0192] Step S602: Deploy the femoral nerve annotation module, the image enhancement module, the multi-mechanism fusion feature extraction module, and the femoral nerve ultrasound image segmentation model based on the SegNet architecture after training and testing to the hospital server background;

[0193] Step S603: Input the femoral nerve ultrasound image of the patient into the hospital server background for processing. Finally, in the femoral nerve ultrasound image segmentation model based on the SegNet architecture, classify and identify each pixel in the femoral nerve ultrasound image to determine the position of the femoral nerve in the femoral nerve ultrasound image, providing a basis for the doctor's femoral nerve puncture decision.

[0194] In summary, through a series of image processing techniques, the present invention optimizes the quality of ultrasound images, improves the visualization effect of the femoral nerve, reduces background noise, and improves the recognition accuracy of subsequent AI models. By using histogram equalization and contrast-limited adaptive histogram equalization, the local and global contrast of ultrasound images is enhanced, making the low-gray femoral nerve area more prominent and improving the visibility of the femoral nerve. Contrast-limited adaptive histogram equalization can enhance details in local areas while avoiding increased background noise caused by over-enhancement, enabling the femoral nerve to be clearly recognized in complex backgrounds. The artifacts and noise in ultrasound images are reduced, improving the image quality. Gaussian filtering is used to smooth the femoral nerve ultrasound image, removing high-frequency noise while retaining the main structural features of the femoral nerve tissue, improving the image quality. Combined with Laplacian edge enhancement, the boundary of the femoral nerve becomes sharper, improving the edge resolution ability of ultrasound images and making the femoral nerve area more distinguishable before segmentation.

[0195] Multi-mechanism fusion feature extraction is adopted to extract key features of ultrasound images in the deep learning model to ensure the accurate recognition of the femoral nerve, improve the accuracy of femoral nerve recognition, and reduce background interference. Adaptive multi-scale feature extraction is used, with convolution kernels of different sizes to enhance the comprehensive perception of the femoral nerve and ensure the complete presentation of the femoral nerve area. Combining high-frequency and low-frequency fusion feature extraction, while retaining the overall shape (low-frequency information) and boundary details (high-frequency information) of the femoral nerve, enhances the stable recognition ability of the femoral nerve. The femoral nerve area is strengthened through the attention mechanism to improve the recognition stability.

[0196] Adopt the femoral nerve ultrasound image segmentation model based on the SegNet architecture to finally achieve pixel-level automatic segmentation of the femoral nerve, ensure the clear boundary of the femoral nerve, and improve the accuracy of femoral nerve block. Improve the femoral nerve segmentation accuracy and ensure the integrity of the femoral nerve boundary. Pixel-level segmentation is performed using the SegNet architecture, and a dynamic feature-guided decoder is introduced in the decoding stage to optimize the feature recovery process, making the femoral nerve boundary clearer. Through dynamic guided attention, the optimal matching of encoded features and decoded features is ensured, improving the segmentation accuracy and reducing the problem of fuzzy boundaries.

[0197] The present invention has been described in detail with reference to the embodiments accompanied by drawings. Those of ordinary skill in the art can make various variations to the present invention based on the above description. Therefore, certain details in the embodiments should not constitute a limitation to the present invention, and the present invention will be protected by the scope defined by the appended claims.

Claims

1. An ultrasound-guided femoral nerve assisted puncture method, characterized in that: The following steps are involved: Step S1: acquiring multi-view ultrasound images of the patient's groin area and constructing a femoral nerve ultrasound image dataset based on the images, wherein the femoral nerve ultrasound image dataset includes femoral nerve ultrasound images; Step S2: the femoral nerve ultrasound image is input into the femoral nerve annotation module to annotate the potential area of ​​the femoral nerve, and then input into the image enhancement module for image enhancement processing to obtain an enhanced femoral nerve ultrasound image; Step S3: constructing a training set and a test set based on the enhanced femoral nerve ultrasound images for training and testing; Step S4: the enhanced femoral nerve ultrasound image is input into a multi-mechanism fusion feature extraction module to perform adaptive multi-scale feature extraction, attention-guided feature enhancement, high- and low-frequency fusion feature extraction, and feature splicing processing to obtain a multi-mechanism fusion feature extraction feature map; Step S5: The multi-mechanism fusion feature extraction feature map is input into the femoral nerve ultrasound image segmentation model based on the SegNet architecture, and dynamic guided attention is introduced to perform feature encoding and decoding processing to obtain pixel-level segmentation results; Step S6: The femoral nerve labeling module and the trained and tested image enhancement module, the multi-mechanism fusion feature extraction module and the femoral nerve ultrasound image segmentation model based on the SegNet architecture are deployed on the hospital server backend. After the patient's femoral nerve ultrasound image is input into the hospital server backend and processed, the femoral nerve position is identified based on the pixel-level segmentation results.

2. The ultrasound-guided femoral nerve assisted puncture method according to claim 1, characterized in that: Step S1 includes the following steps: Step S101: using an ultrasound system equipped with a probe, in brightness mode, the probe is placed above the patient's inguinal fold for transverse scanning, and the typical position of the patient's femoral nerve is found by adjusting the position of the probe; Step S102: The ultrasound system dynamically scans the typical position of the patient's femoral nerve and collects video data, where the video data covers the entire femoral nerve area of ​​the patient. Several femoral nerve ultrasound images are then extracted from the video data, where each femoral nerve ultrasound image includes a complete femoral nerve anatomical area. Finally, the several femoral nerve ultrasound images are stored as a femoral nerve ultrasound image data set.

3. The ultrasound-guided femoral nerve assisted puncture method according to claim 2, characterized in that: Step S2 includes the following steps: Step S201: using annotation software to open a femoral nerve ultrasound image dataset, manually identifying a femoral nerve region in each femoral nerve ultrasound image, using an annotation tool to draw a closed curve along a femoral nerve boundary in the femoral nerve region, and performing preliminary edge detection using a Sobel operator during the process of drawing the closed curve, to obtain a preliminary annotated femoral nerve ultrasound image; Step S202: The initially annotated femoral nerve ultrasound image is input into an image enhancement module, and histogram equalization, contrast-limited adaptive histogram equalization, Gaussian filtering and Laplace edge enhancement are performed to obtain an enhanced femoral nerve ultrasound image.

4. The ultrasound-guided femoral nerve assisted puncture method according to claim 3, characterized in that: Step S3 is specifically as follows: The femoral nerve ultrasound image label collected in step S1 is paired with the femoral nerve ultrasound image enhanced in step S2, and divided into training set and test set in a ratio of 8:2 by random grouping, which is used for training and testing of the multi-mechanism fusion feature extraction module and the femoral nerve ultrasound image segmentation model based on the SegNet architecture; the femoral nerve ultrasound image collected in step S1 is paired with the femoral nerve ultrasound image enhanced in step S2, wherein the enhanced femoral nerve ultrasound image is used as a label, and divided into training set and test set in a ratio of 8:2 by random grouping, which is used for training and testing of the image enhancement module.

5. The ultrasound-guided femoral nerve assisted puncture method according to claim 4, characterized in that: In step S4, the adaptive multi-scale feature extraction and attention-guided feature enhancement processing are specifically as follows: Step S401: performing adaptive multi-scale feature extraction processing: the enhanced femoral nerve ultrasound image is sequentially input into convolution layers, batch normalization layers and sampling layers of different sizes for processing; wherein the enhanced femoral nerve ultrasound image is input into convolution layers of different sizes for multi-scale feature extraction processing; the extracted multi-scale features are input into the batch normalization layer for feature standardization processing; the standardized features are input into the sampling layer for maximum pooling processing, significant features are retained, and a femoral nerve ultrasound image retaining significant features is obtained; Step S402: Performing attention-guiding feature enhancement processing: The femoral nerve ultrasound image with significant features retained is sequentially input into the channel attention layer, the spatial attention layer and the splicing layer for processing; The femoral nerve ultrasound image with retained significant features is input into the channel attention layer, and the useful features are enhanced through channel attention. Then it is input into the spatial attention layer, and the significance of the femoral nerve area is improved through spatial attention. Finally, it is input into the splicing layer, and the channel attention mechanism and the spatial attention mechanism are integrated to ensure the importance of features in different dimensions, and obtain the femoral nerve ultrasound image with attention-guided feature enhancement.

6. The ultrasound-guided femoral nerve assisted puncture method according to claim 4, characterized in that: In step S4, high- and low-frequency fusion feature extraction and feature splicing processing are performed to obtain a multi-mechanism fusion feature extraction feature map, specifically: Step S403: performing high- and low-frequency fusion feature extraction processing: the femoral nerve ultrasound image enhanced with attention guidance features is sequentially subjected to high-frequency information extraction, low-frequency information extraction and feature fusion processing to obtain a femoral nerve ultrasound image that fuses high-frequency and low-frequency information; Step S404: perform feature stitching processing: input the femoral nerve ultrasound image that integrates high-frequency and low-frequency information into the feature stitching layer, perform feature stitching processing, and finally obtain a multi-mechanism fusion feature extraction feature map.

7. The ultrasound-guided femoral nerve assisted puncture method according to claim 5, characterized in that: In step S5, the femoral nerve ultrasound image segmentation model based on the SegNet architecture includes: an encoder, a dynamic feature guided decoder and a Softmax activation function layer. Step S5 is specifically as follows: Step S501: the encoder includes a convolution layer, a batch normalization layer and a maximum pooling layer of the same scale. The multi-mechanism fusion feature extraction feature map is input into the convolution layer of the same scale of the encoder to extract the local femoral nerve features, and then input into the batch normalization layer to normalize the feature distribution of each channel, and finally input into the maximum pooling layer to reduce the resolution through maximum pooling; Step S502: The dynamic feature guided decoder includes a dynamic guided attention layer, an anti-pooling layer, a deconvolution layer and a feature adjustment layer. After the feature map of the multi-mechanism fusion feature extraction is processed by the encoder, it is input into the dynamic guided attention layer of the dynamic feature guided decoder. The dynamic correlation between the encoded feature and the decoded feature is calculated to selectively restore the feature, and then it is input into the anti-pooling layer. By combining the dynamic guided attention to calculate the upsampling result, it is ensured that the information is not lost, and then it is input into the deconvolution layer to further improve the resolution of the decoded feature and restore the detail information at the same time. Finally, it is input into the feature adjustment layer to further optimize the decoded feature, ensure the clear boundary of the femoral nerve, and reduce the influence of background noise; Step S503: After the multi-mechanism fusion feature extraction feature map is processed by the dynamic feature guided decoder, it is input into the Softmax activation function layer for calculation to generate a pixel-level segmentation result.

8. The ultrasound-guided femoral nerve assisted puncture method according to claim 6, characterized in that: Step S6 is specifically as follows: Step S601: using the training set and test set constructed in step S3, the image enhancement module, the multi-mechanism fusion feature extraction module and the femoral nerve ultrasound image segmentation model based on the SegNet architecture are trained and tested; Step S602: deploying the femoral nerve annotation module, the image enhancement module after training and testing, the multi-mechanism fusion feature extraction module, and the femoral nerve ultrasound image segmentation model based on the SegNet architecture on the hospital server backend; Step S603: The patient's femoral nerve ultrasound image is input into the hospital server backend for processing. Finally, in the femoral nerve ultrasound image segmentation model based on the SegNet architecture, each pixel in the femoral nerve ultrasound image is classified and identified to determine the position of the femoral nerve in the femoral nerve ultrasound image, thereby providing a basis for the doctor's femoral nerve puncture decision.

Images