Neck ultrasound segmentation method and application based on U-Net

Abstract

The present invention provides a U-Net-based neck ultrasound segmentation method and application, which relates to the field of medical auxiliary diagnosis technology. The method includes: constructing a U-Net segmentation model; the model includes a dual encoder structure, a jump connection structure and a decoder structure, wherein the front encoder and the decoder structure in the dual encoder structure correspond to the front fusion module and the back fusion module using the multi-layer fusion module, respectively, to promote the image features to be transmitted to different layers in the dual encoder structure and the decoder structure; the jump connection structure includes at least one GAG ​​module to ensure that the image features are transmitted from the corresponding layer of the back encoder in the dual encoder structure to the corresponding layer of the decoder structure; the model is used to segment and identify the neck ultrasound image, and the image features of the target tissue in the neck ultrasound image are extracted. The present invention can realize automatic segmentation and accurate identification of the target tissue in the neck ultrasound image, thereby providing medical auxiliary diagnosis when the doctor performs ultrasound guidance.

Description

Technical Field

[0001] The present invention relates to the field of medical auxiliary diagnosis, and in particular to a neck ultrasound segmentation method based on U-Net. Background Art

[0002] Ultrasound is economical, efficient, convenient, and real-time, making it a core tool for medical diagnosis and treatment. Ultrasound also plays a significant role in guiding radiofrequency ablation for papillary thyroid microcarcinoma, guiding thyroid nodule puncture, diagnosing neuromuscular atrophy, and guiding peripheral nerve nerve (PNB) anesthesia. Ultrasound offers the following advantages:

[0003] One of the advantages is that ultrasonic equipment is relatively economical and affordable. Compared with other high-end medical equipment (such as MRI, CT, etc.), its price is more moderate, which also makes ultrasonic technology easier to popularize and apply in primary medical institutions.

[0004] The second advantage is that ultrasound can propagate rapidly and vibrate at a high frequency, which enables ultrasound to quickly generate images and provide doctors with accurate diagnostic information or treatment guidance.

[0005] The third advantage is that ultrasound equipment is relatively small and easy to operate, allowing medical staff to quickly apply it on-site. Furthermore, ultrasound is relatively portable, making it easy to use in various medical scenarios.

[0006] The fourth advantage is that ultrasound can be transmitted through human tissue without any damage, without the need for invasive procedures such as cutting or puncturing. This means that patients will not feel pain or discomfort when undergoing ultrasound examinations or treatments, and there is a lower risk of complications.

[0007] In practice, despite the widespread clinical application of ultrasound guidance, efficient and accurate positioning has always been a challenge for anesthesiologists during intubation. During ultrasound guidance, inconsistent ultrasound tissue imaging can easily occur due to varying patient positions and the physician's varying probe tilt angles. For less experienced physicians, these inconsistent imaging can easily mislead them. For more experienced physicians, while they can identify the target location based on the anatomical features of the ultrasound image and the distribution of surrounding tissue, accurate identification cannot be guaranteed.

[0008] Therefore, how to automatically segment and accurately identify the target tissue in the acquired neck ultrasound image based on the use of ultrasound-guided technology and equipment to provide doctors with auxiliary diagnosis during ultrasound guidance is of great clinical significance.

[0009] To this end, the present invention provides a U-Net-based neck ultrasound segmentation method and application to solve the automatic segmentation and accurate identification of target tissue in neck ultrasound images, which is a technical problem that needs to be solved urgently. Summary of the Invention

[0010] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a U-Net-based neck ultrasound segmentation method and application. The present invention can automatically segment and accurately identify target tissues in neck ultrasound images, thereby providing medical auxiliary diagnosis when doctors perform ultrasound guidance.

[0011] In order to solve the existing technical problems, the present invention provides the following technical solutions:

[0012] A neck ultrasound segmentation method based on U-Net, comprising:

[0013] Constructing a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes several cervical ultrasound images processed by data cropping, data labeling, and data augmentation operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a skip connection structure, and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure respectively correspond to the front fusion module and the back fusion module using a multi-layer fusion module to promote the transmission of image features to different layers in the dual encoder structure and the decoder structure respectively; the skip connection structure includes at least one GAG ​​module to ensure that image features are transmitted from the corresponding layer of the back encoder in the dual encoder structure to the corresponding layer of the decoder structure;

[0014] The processed neck ultrasound image is used as sample data for training and testing the U-Net segmentation model, and the tested U-Net segmentation model is applied to the segmentation and recognition task of the neck ultrasound image;

[0015] Acquire neck ultrasound images using an ultrasound probe device;

[0016] The collected cervical ultrasound image is segmented and recognized using the aforementioned U-Net segmentation model to extract image features of the target tissue in the aforementioned cervical ultrasound image.

[0017] Furthermore, the U-Net model is trained using a preset proportion of sample data in the cervical ultrasound dataset;

[0018] After training, the remaining sample data in the aforementioned neck ultrasound dataset was tested, and the dice score was used as an evaluation indicator to measure the performance of the U-Net model in the ultrasound neck image segmentation task.

[0019] Furthermore, the dual encoder structure is divided into a front encoder and a back encoder, and the front encoder and the back encoder transmit image features through a jump connection; wherein,

[0020] The front encoder is capable of receiving a neck ultrasound image and sequentially executing a SUCK module and a SCAM attention mechanism module to capture image features corresponding to a target tissue; the target tissue comprises at least one of a sternocleidomastoid muscle, anterior scalene muscle, middle scalene muscle, brachial plexus, cervical nerve pathway, thyroid gland, and carotid artery;

[0021] The post-encoder receives the same neck ultrasound image as the front-encoder, and while processing the aforementioned neck ultrasound image, receives and fuses the image features transmitted by the front-encoder via the jump connection; wherein, the post-encoder consists of a three-layer architecture consisting of two 3x3 convolutions and a SCAM attention mechanism module; wherein the SCAM attention mechanism module is used to locate the image features of the aforementioned target tissue while suppressing other image features other than the image features of the aforementioned target tissue.

[0022] Furthermore, the SUCK module extracts image features of different scales through three different convolution branches; the features of small local areas are obtained by convolution with a small convolution kernel size of 3×3, and the features of large spatial areas are extracted by using 7x7 convolution and 13x13 convolution kernels.

[0023] Furthermore, the SCAM attention mechanism module can sequentially execute:

[0024] Given an input X∈R B×C×H×W Use global average pooling to split X∈R by rows and columns B×C×H×1 and X∈R B×C×1×W Then, the aforementioned X∈R B×C×H×1 and X∈R B×C×1×W Split into 8 parts according to the number of channels, and obtain sub-features X i ∈R B ×C / 8×H×1 and sub-feature X j ∈R B×C / 8×1×W ;

[0025] The aforementioned sub-feature X i ∈R B×C / 8×H×1 and sub-feature X j ∈R B×C / 8×1×W After passing the SUCK module, the aforementioned X i ∈RB×C / 8×H×1 With X j ∈R B×C / 8×1×W Aggregate by rows and columns separately and normalize them using group normalization with K groups to obtain X∈R B×C×H×1 and X∈R B×C×1×W , where K is a positive integer;

[0026] After normalization, X∈R B×C×H×1 and X∈R B×C×1×W After passing through the sigmoid function respectively, combined with the aforementioned given input X∈R B×C×H×W Generate spatial attention X S ;

[0027] In the aforementioned spatial attention X S After using average pooling and group normalization, the query Q, key K, and value V based on self-attention are divided into three branches. After performing a depth-wise separable convolution operation on each branch, the query attention weight, key attention weight, and value attention weight are obtained by calculating the query Q, key K, and value V respectively.

[0028] Combined with the aforementioned query attention weight, key attention weight and value attention weight, after the context-aware self-attention mechanism, after average pooling and Sigmoid function, combined with the aforementioned spatial attention X S Perform multiplication operation to obtain the output feature map.

[0029] Furthermore, the multi-layer fusion module includes a front fusion module corresponding to each layer of the front encoder of the dual encoder structure and a back fusion module corresponding to each layer of the decoder structure; wherein,

[0030] The front fusion module is capable of receiving the image features input by each layer in the front encoder, and after gradually changing the number of channels and image size through multiple convolutions, performing a residual connection with the output layer of the front encoder;

[0031] The post-fusion module can gradually adjust the number of channels and image size of the image features output by each layer in the decoder structure through multiple deconvolution operations, and then perform residual connection with the image features output by the bottom layer.

[0032] Further, the GAG ​​module includes query embedding, key embedding and value embedding;

[0033] When executing the GAG ​​module, it specifically includes:

[0034] Use the aforementioned key embedding and query embedding to operate on the input image features to obtain the key attention weight and query attention weight;

[0035] The key attention weight and query attention weight are connected with residual connection, and after Gelu function, 1×1 convolution and sigmoid function, they are multiplied with the value embedding;

[0036] Among them, when there are multiple GAG ​​modules, the image features output by the upper-layer GAG module are input into the lower-layer GAG module after adjusting the image size through deconvolution operation. After adjusting the number of channels through 1×1 convolution, the residual connection is performed with the result obtained by the aforementioned multiplication operation to highlight the global image features of the cervical ultrasound image.

[0037] A neck ultrasound segmentation device based on U-Net, comprising:

[0038] A data construction unit is used to construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes a plurality of cervical ultrasound images processed by data cropping, data labeling, and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a skip connection structure, and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure respectively correspond to the front fusion module and the back fusion module using the multi-layer fusion module to promote the image features to be transmitted to different layers in the dual encoder structure and the decoder structure respectively; the skip connection structure includes at least one GAG ​​module to ensure that the image features are transmitted from the corresponding layer of the back encoder in the dual encoder structure to the corresponding layer of the decoder structure;

[0039] A model building unit, configured to use the processed neck ultrasound image as sample data for training and testing the U-Net segmentation model, and apply the tested U-Net segmentation model to the segmentation and recognition task of the neck ultrasound image;

[0040] An image acquisition unit, used for acquiring an ultrasonic image of the neck through an ultrasonic probe device;

[0041] The image segmentation and recognition unit is used to segment and recognize the collected cervical ultrasound image using the aforementioned U-Net segmentation model to extract image features of the target tissue in the aforementioned cervical ultrasound image.

[0042] A U-Net-based neck ultrasound segmentation system, comprising:

[0043] Network node, used to send and receive neck ultrasound images;

[0044] An image segmentation module, used to identify target tissue in the aforementioned neck ultrasound image and perform segmentation processing;

[0045] A system server, the system server connecting the network nodes and the image segmentation module;

[0046] The system server is configured to: construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes a number of cervical ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a skip connection structure and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure correspond to the front fusion module and the back fusion module using the multi-layer fusion module, respectively, to promote Image features are transferred to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that image features are transferred from the corresponding layer of the post-encoder in the dual encoder structure to the corresponding layer of the decoder structure; the processed neck ultrasound image is used as sample data in the training and testing of the U-Net segmentation model, and the U-Net segmentation model that has completed the test is applied to the segmentation and recognition task of the neck ultrasound image; the neck ultrasound image is collected by an ultrasound probe device; the collected neck ultrasound image is segmented and recognized using the U-Net segmentation model to extract the image features of the target tissue in the neck ultrasound image.

[0047] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the implementation steps of any of the above methods.

[0048] Based on the above advantages and positive effects, the advantages of the present invention are: an improved U-net segmentation model, which adopts a dual encoder structure with a front encoder and a post encoder to improve the ability to extract image features, wherein the front encoder sequentially executes a spatial understanding convolution kernel module and a spatial channel attention mechanism module to capture the image features corresponding to the target tissue to improve the recognition accuracy of small tissues, and a spatial channel attention mechanism module is introduced in the post encoder to locate the image features of the target tissue; and the front encoder in the dual encoder structure and the aforementioned decoder structure correspond to the front fusion module and the post fusion module of the multi-layer fusion module to promote the image features to be transferred to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that the image features are transferred from the corresponding layer of the post encoder in the dual encoder structure to the corresponding layer of the decoder structure.

[0049] Furthermore, the improved U-net segmentation model can automatically segment and accurately identify target tissues in neck ultrasound images, especially the sternocleidomastoid muscle, anterior scalene muscle, middle scalene muscle, brachial plexus, cervical nerve pathway, thyroid gland and carotid artery, thereby providing medical auxiliary diagnosis when doctors perform ultrasound guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a process flow provided by an embodiment of the present invention.

[0051] Figure 2 A schematic diagram of the structure of the U-Net segmentation model provided in an embodiment of the present invention.

[0052] Figure 3 A schematic diagram of the structure of the SUCK module in the U-Net segmentation model provided in an embodiment of the present invention.

[0053] Figure 4 Schematic diagram of the structure of the SCAM attention mechanism module in the U-Net segmentation model provided in an embodiment of the present invention.

[0054] Figure 5 A schematic diagram of the structure of the GAG ​​module in the U-Net segmentation model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following is a further detailed description of a U-Net-based neck ultrasound segmentation method and application disclosed in the present invention in conjunction with the accompanying drawings and specific embodiments. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated, and they can be combined with each other to achieve better technical effects. In the drawings of the following embodiments, the same reference numerals appearing in each drawing represent the same features or components, which can be applied to different embodiments. Therefore, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0056] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not intended to limit the conditions under which the invention can be implemented. Any structural modification, change in proportional relationship, or adjustment of size should fall within the scope of the technical content disclosed in the invention without affecting the efficacy and purpose of the invention. The scope of the preferred embodiments of the present invention includes alternative implementations, in which the functions can be performed in a non-described or discussed order, including performing the functions in a substantially simultaneous manner or in a reverse order according to the functions involved, which should be understood by those skilled in the art of the art to which the embodiments of the present invention belong.

[0057] Technologies, methods, and apparatus known to persons of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0058] Example

[0059] See also Figure 1 FIG. 1 is a flow chart of the present invention. The implementation step S100 of the method is as follows:

[0060] S101, build a neck ultrasound dataset and U-Net segmentation model.

[0061] The neck ultrasound dataset includes a plurality of neck ultrasound images that have been processed by data cropping, data labeling, and data enhancement operations.

[0062] The neck ultrasound image dataset in this embodiment is obtained by those skilled in the art. Specifically, neck ultrasound images of 500 patients are collected and organized to construct the neck ultrasound dataset, which includes approximately 3,000 images taken by neck ultrasound examinations.

[0063] During the construction of the aforementioned cervical ultrasound dataset, the acquired cervical ultrasound images were preferably divided into multiple categories, and the target tissue and surrounding tissue in each cervical ultrasound image were labeled using the Labelme software tool. Finally, data augmentation was used to complete the construction of the cervical ultrasound dataset. Since data augmentation is a well-known technical method in this field, it will not be further elaborated here.

[0064] In actual operation, the original neck ultrasound images were renumbered to protect the patient's privacy information, and neck ultrasound images with target structures were selected from them, totaling 1,500.

[0065] The target tissue includes at least one of the sternocleidomastoid muscle, the anterior scalene muscle, the middle scalene muscle, the brachial plexus, the cervical nerve pathway, the thyroid gland, and the carotid artery.

[0066] The surrounding tissues are other tissues that can be collected by neck ultrasound except for the aforementioned target tissues.

[0067] Specifically, the surrounding tissues include jugular vein, lymph nodes, esophagus, trachea, vocal cords, laryngeal structures, cervical vertebrae, parathyroid glands, adipose tissue, parathyroid glands, skin and subcutaneous tissue.

[0068] In a neck ultrasound image, when the sternocleidomastoid muscle, anterior scalene muscle, middle scalene muscle, brachial plexus, cervical nerve pathway, thyroid gland, and carotid artery are considered target tissues, the jugular vein, lymph nodes, esophagus, trachea, vocal cords, laryngeal structures, cervical vertebrae, parathyroid gland, adipose tissue, parathyroid gland, skin, and subcutaneous tissue can be considered as surrounding tissues to provide background information and anatomical reference, thereby facilitating the location of the target tissue and providing a more comprehensive anatomical structure relationship. Since the above-mentioned classification of target tissues and surrounding tissues is a conventional technical means in the field, it will not be further elaborated here.

[0069] It is also worth noting that, in the neck ultrasound image, the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure.

[0070] The small target tissue is preferably a target tissue with a small volume or a thin structure that requires high-resolution ultrasound to be clearly displayed, such as the brachial plexus, cervical nerve pathway, and thyroid gland.

[0071] The brachial plexus is the terminal branch of the brachial plexus, responsible for movement and sensation in the upper limbs. Because the brachial plexus is small and easily obscured by other tissues during ultrasound, high resolution is required for clear visualization. The cervical nerve pathways are also relatively small, with thin nerve tissue, requiring careful adjustment of the ultrasound probe for accurate positioning. While the thyroid gland is smaller than nerves and muscles, it is a very important small organ within the entire neck, and its specific nodules or lesions often require detailed ultrasound evaluation.

[0072] The large target tissues occupy a large space in the neck area and are relatively easy to identify by ultrasound, such as the sternocleidomastoid muscle, the anterior and middle scalene muscles, and the carotid artery.

[0073] Among them, the sternocleidomastoid muscle is one of the largest muscles in the neck, running along the side of the neck. It has a clear position, a large size, and is easy to display on ultrasound; the anterior scalene muscle and the middle scalene muscle occupy a larger area during cross-sectional scanning and are easy to distinguish from other structures; the carotid artery is one of the largest blood vessels in the neck, with a deep anatomical position and a large diameter, making it easy to detect and evaluate by ultrasound.

[0074] Dividing the aforementioned target tissues according to the size and range of the anatomical structure can help to more specifically adjust the probe resolution and scanning mode during ultrasound evaluation to ensure that target tissues of different sizes can be clearly identified.

[0075] In order to ensure the invariance of the characteristics of the tissue region and reduce the amount of calculation, this embodiment performs cropping and occlusion processing on the original neck ultrasound image.

[0076] By way of example and not limitation, Figure 1 For example, for each neck ultrasound image containing the target structure, several physicians with many years of clinical intubation experience in this field segmented the anatomical structures of different target tissues and labeled them using the LabelMe tool.

[0077] As one of the preferred implementations of this embodiment, step S102 is executed, and the processed neck ultrasound image is used as sample data for training and testing of the U-Net segmentation model, and the U-Net segmentation model that has completed the test is applied to the segmentation and recognition task of the neck ultrasound image.

[0078] Specifically, the U-Net model is trained using a preset proportion (for example, 80%) of sample data in the neck ultrasound dataset; after training, the remaining proportion (for example, 20%) of sample data in the aforementioned neck ultrasound dataset is tested, and the dice score is used as an evaluation indicator to measure the performance of the U-Net model in the ultrasound neck image segmentation task.

[0079] The U-Net segmentation model can retain the image features of the target tissue when extracting image features. In actual operation, the above image features are specifically determined in combination with the doctor's analysis requirements for the neck ultrasound image.

[0080] See Figure 2 As shown, the U-Net segmentation model preferably includes a dual encoder structure, a skip connection structure, and a decoder structure. The pre-encoder in the dual encoder structure and the decoder structure use a pre-fusion module and a post-fusion module of a multi-layer fusion module to promote the transfer of image features to different layers in the dual encoder and decoder structures, respectively. The skip connection structure includes at least one GAG ​​module to ensure that image features are transferred from the corresponding layer of the post-encoder in the dual encoder structure to the corresponding layer of the decoder structure.

[0081] The U-Net segmentation model in this embodiment preferably adopts a dual encoder structure to improve the ability to extract image features.

[0082] The dual encoder structure is divided into a front encoder and a back encoder, and the front encoder and the back encoder transmit image features through a jump connection.

[0083] Among them, the front encoder is able to receive a neck ultrasound image and sequentially execute a spatial understanding convolution kernel module and a spatial channel attention mechanism module to capture image features corresponding to the target tissue.

[0084] The post-encoder receives the same neck ultrasound image as the front-encoder, and while processing the neck ultrasound image, receives and fuses features transmitted by the front-encoder via the skip connection.

[0085] The post-encoder consists of a three-layer architecture consisting of two 3x3 convolutions and a spatial channel attention mechanism module; the spatial channel attention mechanism module is used to locate the image features of the aforementioned target tissue while suppressing other image features other than the image features of the aforementioned target tissue.

[0086] It is worth noting that the pre-encoder is able to capture the image features of both small and large target tissues through the synergistic effect of the spatial understanding convolution kernel module and the spatial channel attention mechanism module. Furthermore, the pre-encoder is responsible for initially extracting features from the neck ultrasound image and inputting them into the post-encoder for more refined processing, thereby improving the ability to extract image features from neck ultrasound images.

[0087] This embodiment takes into account that large target tissues (such as veins or arteries) are more sensitive to semantic information, while small target tissues (such as nerves) are more sensitive to texture information. Therefore, a spatial understanding convolution kernel module is proposed to extract features at different scales.

[0088] The spatial understanding convolution kernel module is called Spatial Understanding Convolution Kernel in English, abbreviated as SUCK module.

[0089] The front encoder in the dual encoder structure adopts the SUCK module to improve the segmentation accuracy of small target tissues.

[0090] Specifically, the spatial understanding convolution kernel module, namely the SUCK module, extracts image features of different scales through three different convolution branches; among them, the features of small local areas are obtained by convolution with a small convolution kernel size of 3×3, and the features of large spatial areas are extracted by using 7x7 convolution and 13x13 convolution kernels.

[0091] As an example and not a limitation, the structural diagram of the SUCK module is as follows: Figure 3 As shown in the figure, for a given input feature map, the output result of the SUCK module is composed of the results of three convolution branches, specifically through three different convolution branches to extract feature information of different scales.

[0092] For the first branch, the input feature x passes through a 3x3 convolution with a dilation rate of 1 and a batch normalization (BN) layer before being directly output. This means that features for small local regions are obtained through convolution with a small kernel size (3×3). For the second branch, the input x passes through dilation rates of 1 and 2, respectively, and is then residually connected to the first branch before being output. For the third branch, the input feature x passes through 3x3 convolutions with dilation rates of 1, 2, and 3, and is then concatenated with the residuals of the first and second branches. This means that features for large spatial regions are extracted using simulated 7x7 and 13x13 convolution kernels, respectively.

[0093] Finally, the different features extracted by the three branches are connected with the different anatomical structures and then input into the SCAM module of the next layer.

[0094] For the SUCK module, since the Gelu activation function is used after each convolution, the expression of the SUCK module is as follows:

[0095] x 0 =Concat(W R1 x i +W R2 (W R1 x i ),W R3 (W R2 (W R1 z i ))+b1)

[0096] Among them, Concat means concatenating the matrices on the left and right sides according to the channel dimension, x 0 , x i Represent the output features and input features respectively, W R1 , W R2 , W R3 They represent 3x3 convolutions with expansion rates of 1, 2, and 3 respectively, and b1 is the bias.

[0097] The expression for the dilated convolution with input x, convolution kernel k, and dilation rate r is:

[0098] (x*k r )(p)=∑ s+rt＝p x(s)k(t)

[0099] Among them, x represents the feature map, k r is the convolution kernel with an expansion rate of r, p is the value at p after the feature map and the convolution kernel are operated, t represents the current time point, and is the output (x*k r )(p) is a time variable.

[0100] The spatial channel attention mechanism module is called Spatial-Channel Attention Module in English, abbreviated as SCAM, that is, SCAM attention mechanism module.

[0101] As another preferred embodiment of this embodiment, Figure 4 As shown, the SCAM attention mechanism module can execute step S110 in sequence:

[0102] S111, the given input X∈R B×C×H×W Use global average pooling to split X∈R by rows and columns B ×C×H×1 and X∈R B×C×1×W Then, the aforementioned X∈R B×C×H×1 and X∈R B×C×1×W Split into 8 parts according to the number of channels, and obtain sub-features X i ∈R B×C / 8×H×1 and sub-feature X j ∈R B×C / 8×1×W .

[0103] It is worth noting here that for the aforementioned X∈R B×C×H×W , this embodiment preferably uses global average pooling for each dimension, thereby creating two unidirectional one-dimensional sequence structures, namely X∈R B×C×H×1 and X∈R B×C×1×W Then, the aforementioned X∈R B ×C×H×1 and X∈R B×C×1×W Split into 8 parts according to the number of channels, and obtain sub-features X i ∈R B×C / 8×H×1 and sub-feature X j ∈R B ×C / 8×1×W This operation enables the model to learn more spatial distribution and contextual connections. In addition, since the SUCK module can obtain different receptive fields, S112 is executed.

[0104] S112, the aforementioned sub-feature X i ∈R B×C / 8×H×1 and sub-feature X j ∈R B×C / 8×1×W After passing the SUCK module, the aforementioned X i ∈R B×C / 8×H×1 With X j ∈R B×C / 8×1×W Aggregate by rows and columns respectively, and use group normalization (GN) with K groups to normalize them, and get X∈R B×C×H×1 and X∈R B×C×1×W .

[0105] Wherein, K is a positive integer. In this embodiment, it is preferred to set the value of K to 8, that is, for the aforementioned X i With X j Aggregate by rows and columns separately and use group normalization with 8 groups to normalize the above X i With X j Normalize them separately.

[0106] In addition, it is worth emphasizing that the aforementioned X i With X j After passing through the SUCK module, the aggregation operation is performed according to rows and columns, that is, under the concat connection operation and GN-8 group normalization, the 8 sub-features X are combined i and 8 molecular features X j Reaggregate into a complete feature X∈R B×C×H×1 And 1 complete feature X∈R B×C×1×W .

[0107] S113, after normalization, X∈R B×C×H×1 and X∈R B×C×1×W After passing through the sigmoid function respectively, combined with the aforementioned given input X∈R B×C×H×W Generate spatial attention X S .

[0108] The described spatial attention extracts multi-semantic spatial information from each feature, thus providing spatial priors for channel attention computation.

[0109] The spatial attention X S Expressed as:

[0110]

[0111] Among them, σ is the Sigmoid function, and They represent the group normalization of the 8 groups of sub-features separated along the height and width directions respectively, and Concat means splicing the left and right matrices according to the channel dimension.

[0112] S114, in the aforementioned spatial attention X S After using average pooling and group normalization, the query Q, key K and value V based on self-attention are divided into three branches. After performing a depth-wise separable convolution operation on each branch, the query attention weight, key attention weight and value attention weight are obtained respectively by calculating the query Q, key K and value V.

[0113] Among them, group normalization (GN) can divide a feature into several groups according to different channels, and then normalize the channels within each group. This can reduce the computational complexity of the model while maintaining the normalization effect.

[0114] The depthwise separable convolution (DWConv) is a convolution operation that can reduce the computational complexity and the number of parameters of the model.

[0115] Specifically, the depthwise separable convolution operation includes a depthwise convolution step and a pointwise convolution step.

[0116] In the depthwise convolution step, convolution is performed on each channel of the input separately. That is, for each input channel, a separate convolution kernel is used to operate, rather than convolving all input channels together. For example, if the input feature map has N channels (N is a positive integer), depthwise convolution will use a convolution kernel for each channel. Assuming that the size of each convolution kernel is k×k (k is a positive integer), there are N k×k convolution kernels in the depthwise convolution operation.

[0117] In the pointwise convolution (also known as 1x1 convolution) step, the channel information after the depthwise convolution is fused. Specifically, the output channels at each position are linearly combined using a 1x1 convolution kernel to adjust the number of channels. Then, the results of the depthwise convolution are combined to form the final output channel number.

[0118] S115, combining the query attention weight, key attention weight and value attention weight, after the context-aware self-attention mechanism, after average pooling and Sigmoid function, combined with the spatial attention X S Perform multiplication operation to obtain the output feature map.

[0119] Considering that channel attention focuses on the feature correlation between channels and ignores the information of spatial dimension, this embodiment preferably adopts the self-attention Q,K,V∈R calculated along the channel dimension. B×C×N , where N = WH, the segmentation performance of the model is enhanced by combining channel attention and spatial attention.

[0120] Specifically, Q, K, and V represent query, key, and value, respectively, which are the three core elements in the self-attention mechanism and are used to calculate attention weights and weighted output information.

[0121] B represents the batch size, that is, the number of samples in the batch; C represents the number of feature channels, which refers to the depth or dimension of the input features; N is the number of spatial locations of each feature map, which is equal to W×H, that is, the product of the width and height of the feature map.

[0122] The above structure applies the context-aware self-attention mechanism in the channel dimension. By jointly calculating the attention weights using Q, K, and V, it can capture the information relationship between different channels in the global context, thereby enhancing the segmentation performance of the model. In particular, in pixel-level tasks (such as image segmentation), it improves the ability to capture detailed information in feature maps.

[0123] Among them, the context-aware multi-head self-attention mechanism (CA-MHSA), that is, the multi-head self-attention mechanism based on channel attention, is used to enhance the context understanding ability of the self-attention model. It calculates attention on multiple heads and combines contextual information to further improve the model's representation ability, especially when processing inputs with complex dependencies.

[0124] Based on this, it can be determined that the output feature map Y of the SCAM attention mechanism module is expressed as:

[0125] Among them, σ represents the Sigmoid function, Pool preferably represents 7X7 average pooling, are mapping functions that generate queries, keys, and values, respectively, and the CA-MHSA attention mechanism is represented by Att, so that

[0126]

[0127] It is worth noting that 7x7 average pooling is preferred in this embodiment. It downsamples the image by taking the average of each 7x7 region in the input feature map, reducing the resolution from (H, W) to (H', W'). This operation helps reduce the spatial dimension of the feature map while preserving important contextual information in the image.

[0128] As another preferred implementation of this embodiment, in order to enable the information extracted by each layer of the encoder to flow better, the U-Net segmentation model in this embodiment preferably uses a multi-layer fusion module to enable the upper and lower layers to preserve more image features, and uses the GAG ​​attention mechanism to improve the model segmentation accuracy.

[0129] Preferably, the multi-layer fusion module includes a front fusion module provided for each layer of the front encoder corresponding to the dual encoder structure and a back fusion module provided for each layer of the aforementioned decoder structure.

[0130] The front fusion module is capable of receiving the image features inputted by each layer in the front encoder, and after gradually changing the number of channels and image size through multiple convolutions, performing a residual connection with the output layer of the front encoder.

[0131] Specific combination Figure 2 As shown in FIG, the front fusion module can receive the image features of each layer input, and after gradually changing the number of channels and image size through four convolutions, perform a residual connection with the output layer of the front encoder.

[0132] In this embodiment, it is preferred that in the 5-layer structure of the front encoder, the SUCK module, the SCAM attention mechanism module and the activation function Gelu are used sequentially for each layer structure, so that the features obtained by the residual connection of all layers include the image features of the first four layers, and are residually connected with the last layer.

[0133] The post-fusion module can gradually adjust the number of channels and image size of the image features output by each layer in the decoder structure through multiple deconvolution operations, and then perform residual connection with the image features output by the bottom layer.

[0134] Specific, combined Figure 2 As shown, the post-fusion module can receive the image features output by each layer. In this embodiment, four deconvolution operations are preferably used to gradually change the number of channels, and finally a residual connection is performed with the bottom output of the decoder structure.

[0135] As another preferred implementation of this embodiment, the post-fusion module can gradually adjust the number of channels and image size of the image features output by each layer in the decoder structure through multiple deconvolution operations and convolution operations, and then perform residual connection with the image features output by the bottom layer.

[0136] It is worth noting that the convolution operation and deconvolution operation in the post-fusion module preferably have different configuration requirements according to different layers, wherein the deconvolution operation can adjust the image size. For example, the deconvolution operation is used in each layer of the post-fusion module to expand the height and width of the image size by twice the original. For example, the height and width of the image size are adjusted from 32×32 to 512×512 through 4 deconvolution operations, and the convolution operation can adjust the number of channels and refine the image features. In the post-fusion module, it is preferred to set a convolution operation for each deconvolution operation to gradually change the number of channels. When performing a residual connection with the bottom output of the decoder structure, ensure that the image size and the number of channels match the image size and the number of channels output in the jump connection structure, thereby completing the residual connection.

[0137] Among them, the convolution operation can be one convolution or multiple convolutions. The convolution operation in the post-fusion module preferably uses multiple convolutions. This is because the gradual transformation of channels by multiple convolutions will be smoother, which avoids the problems of gradient disappearance, gradient explosion or information loss in the U-Net segmentation model designed in this embodiment during image processing, and at the same time enhances the stability of the model.

[0138] Preferably, the GAG ​​(Global Attention Gate) module is a mechanism for enhancing the model's ability to perceive global information during feature learning. The GAG ​​module adjusts and optimizes the selection of local image features by capturing global image features, thereby enabling the model to more effectively focus on globally relevant information.

[0139] In this embodiment, the GAG ​​module includes query embedding, key embedding and value embedding. Figure 5 As shown, step S120 is included:

[0140] S121, use the aforementioned key embedding and query embedding to operate on the input image features to obtain the key attention weight and query attention weight.

[0141] S122: A residual connection is adopted for the aforementioned key attention weight and query attention weight, and after passing through the Gelu function, 1×1 convolution and sigmoid function, multiplication operation is performed with the value embedding.

[0142] Among them, when there are multiple GAG ​​modules, the image features output by the upper-layer GAG module are input into the lower-layer GAG module after adjusting the image size through deconvolution operation, and after adjusting the number of channels through 1×1 convolution, a residual connection is performed with the result obtained by the aforementioned multiplication operation to highlight the global image features of the cervical ultrasound image.

[0143] In addition, combined Figure 2 As shown, two GAG modules are added to the skip connection structure. The GAG ​​modules can receive the image features output by the post-encoder in the dual-encoder structure and the image features output by the decoder structure, and perform the aforementioned step S120, thereby adjusting the attention weights calculated between the local image features and the global image features to determine the importance of the local image features in the final decision. Furthermore, the GAG ​​modules can suppress irrelevant features while highlighting relevant features. In addition, the latter GAG module can simultaneously receive the image features output by the previous GAG module and further highlight the global image features of the cervical ultrasound image.

[0144] In actual operation, steps S103 and S104 are specifically performed.

[0145] S103, collecting a neck ultrasound image using an ultrasound probe device.

[0146] S104: Use the aforementioned U-Net segmentation model to segment and identify the collected cervical ultrasound image to extract image features of the target tissue in the aforementioned cervical ultrasound image.

[0147] In addition, to ensure that the U-Net segmentation model is suitable for recognizing cervical ultrasound images in actual operations, the performance of the improved U-Net segmentation model in this embodiment on the cervical ultrasound image dataset was verified and the resulting test results were analyzed. Specifically, the following operations were performed:

[0148] First, the improved U-Net segmentation model was trained on the personalized cervical ultrasound image dataset used in this example, and its performance was evaluated. Second, the experimental results were compared and analyzed with the pre-improved U-Net segmentation algorithm. Finally, the improved model was compared and evaluated with mainstream segmentation algorithms.

[0149] Taking the improved U-Net segmentation model in this embodiment, the classic Unet model, and advanced models (such as MACUnet, MEGANet, ACCUnet, MultiResUnet, and Unet++) as examples, 100 rounds of iterative training were performed on the personalized cervical ultrasound image dataset of this embodiment. Subsequently, the DSC evaluation index of the target tissue segmented and identified was compared, and finally the Mean DSC (MDSC) and Mean IOU (MIOU) results were given, as shown in Table 1.

[0150] The target tissues for segmentation and identification include the thyroid gland represented by T, the carotid artery represented by CA, the cervical nerve pathway represented by CNP, the brachial plexus represented by BP, the middle scalene muscle represented by MSM, the anterior scalene muscle represented by ASM, and the sternocleidomastoid muscle represented by SCM.

[0151] Table 1 Experimental comparison results of multiple groups of models

[0152]

[0153] The experimental results show that the improved U-Net segmentation model in this embodiment can accurately segment the target tissue, with a Mean DSC of 0.933 and a Mean IOU of 0.874, which have high confidence levels. Moreover, for rotated neck ultrasound images, segmentation and recognition operations can be performed more accurately for small target tissues (such as cervical nerve pathways, brachial plexus nerves, etc.).

[0154] Other technical features are described in the previous embodiments and will not be repeated here.

[0155] In addition, the present invention also provides an embodiment, which provides a neck ultrasound segmentation device based on U-Net, including:

[0156] A data construction unit is used to construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes several cervical ultrasound images that have been processed by data cropping, data labeling and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a jump connection structure and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure respectively correspond to the front fusion module and the back fusion module using a multi-layer fusion module to promote the image features to be transmitted to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that the image features are transmitted from the corresponding layer of the back encoder in the dual encoder structure to the corresponding layer of the decoder structure.

[0157] The model building unit is used to input the processed neck ultrasound image as sample data into the training and testing of the U-Net segmentation model, and apply the tested U-Net segmentation model to the segmentation and recognition task of the neck ultrasound image.

[0158] The image acquisition unit is used to acquire an ultrasonic image of the neck through an ultrasonic probe device.

[0159] The image segmentation and recognition unit is used to segment and recognize the collected cervical ultrasound image using the aforementioned U-Net segmentation model to extract image features of the target tissue in the aforementioned cervical ultrasound image.

[0160] Other technical features are described in the previous embodiments and will not be repeated here.

[0161] In addition, the present invention also provides an embodiment, which provides a neck ultrasound segmentation system based on U-Net, including:

[0162] Network node, used to send and receive neck ultrasound images.

[0163] The image segmentation module is used to identify the target tissue in the aforementioned cervical ultrasound image and perform segmentation processing.

[0164] A system server is connected to the network node and the image segmentation module.

[0165] The system server is configured to: construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes a number of cervical ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a skip connection structure and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure correspond to the front fusion module and the back fusion module using the multi-layer fusion module, respectively, to promote Image features are transferred to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that image features are transferred from the corresponding layer of the post-encoder in the dual encoder structure to the corresponding layer of the decoder structure; the processed neck ultrasound image is used as sample data in the training and testing of the U-Net segmentation model, and the U-Net segmentation model that has completed the test is applied to the segmentation and recognition task of the neck ultrasound image; the neck ultrasound image is collected by an ultrasound probe device; the collected neck ultrasound image is segmented and recognized using the U-Net segmentation model to extract the image features of the target tissue in the neck ultrasound image.

[0166] For other technical features, please refer to the previous embodiments and will not be repeated here.

[0167] In addition, an embodiment of the present invention also provides a computer-readable storage medium on which a program is stored for use in the aforementioned U-Net-based neck ultrasound segmentation system. When the program is executed by a processor, it can implement the steps of any of the above-mentioned U-Net-based neck ultrasound segmentation methods.

[0168] For other technical features, please refer to the previous embodiments and will not be repeated here.

[0169] In the above description, the components may be selectively and operatively combined in any number within the scope of the intended protection of the present disclosure. In addition, terms such as "include," "encompass," and "have" should be interpreted as inclusive or open-ended rather than exclusive or closed by default, unless expressly defined to the contrary. All technical, technological, or other terms have the meanings understood by those skilled in the art, unless they are defined to the contrary. Common terms found in dictionaries should not be interpreted in an overly idealized or unrealistic manner in the context of the relevant technical documentation, unless expressly defined to that extent by the present disclosure.

[0170] Although example aspects of the present disclosure have been described for illustrative purposes, those skilled in the art will appreciate that the foregoing description is merely a description of preferred embodiments of the present invention and does not limit the scope of the present invention in any way. The scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order in which they appear or are discussed. Any changes or modifications made by those skilled in the art based on the foregoing disclosure are intended to fall within the scope of the claims.

Claims

1. A neck ultrasound segmentation method based on U-Net, characterized in that: include: Construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes several cervical ultrasound images that have been processed by data cropping, data labeling and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a jump connection structure and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure respectively correspond to the front fusion module and the back fusion module using the multi-layer fusion module to promote the image features to be transmitted to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that the image features are transmitted from the back encoder in the dual encoder structure The corresponding layer is passed to the corresponding layer of the decoder structure; the GAG ​​module is a global attention gating module; the GAG ​​module includes query embedding, key embedding and value embedding; when executing the GAG ​​module, it specifically includes: using the aforementioned key embedding and query embedding to operate on the input image features to obtain key attention weights and query attention weights; the aforementioned key attention weights and query attention weights are subjected to residual connection, and after passing through the Gelu function, 1×1 convolution and sigmoid function, they are multiplied with the value embedding; wherein, when there are multiple GAG ​​modules, the image features output by the upper-layer GAG module are adjusted to the image size through the deconvolution operation and then input into the lower-layer GAG module, and after adjusting the number of channels through 1×1 convolution, the residual connection is performed with the result obtained by the aforementioned multiplication operation to highlight the global image features of the cervical ultrasound image; The processed neck ultrasound image is used as sample data for training and testing the U-Net segmentation model, and the tested U-Net segmentation model is applied to the segmentation and recognition task of the neck ultrasound image; Acquire neck ultrasound images using an ultrasound probe device; The collected cervical ultrasound image is segmented and recognized using the aforementioned U-Net segmentation model to extract image features of the target tissue in the aforementioned cervical ultrasound image.

2. The method according to claim 1, characterized in that The U-Net model is trained using a preset proportion of sample data in the cervical ultrasound dataset; After completing the training, the remaining sample data in the aforementioned neck ultrasound dataset was tested, and dicescore was used as an evaluation indicator to measure the performance of the U-Net model in the ultrasound neck image segmentation task.

3. The method according to claim 1, characterized in that The dual encoder structure is divided into a front encoder and a back encoder, and the front encoder and the back encoder transmit image features through a jump connection; wherein, The front encoder is capable of receiving a neck ultrasound image and sequentially executing a SUCK module and a SCAM attention mechanism module to capture image features corresponding to a target tissue; the SUCK module is a spatial understanding convolution kernel module; the SCAM attention mechanism module is a spatial channel attention mechanism module; the target tissue comprises at least one of a sternocleidomastoid muscle, anterior scalene muscle, middle scalene muscle, brachial plexus, cervical nerve pathway, thyroid gland, and carotid artery; The post-encoder receives the same neck ultrasound image as the front-encoder, and while processing the aforementioned neck ultrasound image, receives and fuses the image features transmitted by the front-encoder via the jump connection; wherein, the post-encoder consists of a three-layer architecture consisting of two 3x3 convolutions and a SCAM attention mechanism module; wherein the SCAM attention mechanism module is used to locate the image features of the aforementioned target tissue while suppressing other image features other than the image features of the aforementioned target tissue.

4. The method according to claim 3, characterized in that The SUCK module extracts image features of different scales through three different convolution branches. The features of small local areas are obtained by convolution with a small convolution kernel size of 3×3, and the features of large spatial areas are extracted by using 7x7 convolution and 13x13 convolution kernels.

5. The method according to claim 3, characterized in that The SCAM attention mechanism module can execute sequentially: Given an input X∈R B×C×H×W Use global average pooling to split X∈R by rows and columns B×C×H×1 and X∈R B ×C×1×W Then, the aforementioned X∈R B×C×H×1 and X∈R B×C×1×W Split into 8 parts according to the number of channels, and obtain sub-features X i ∈R B ×C / 8×H×1 and sub-feature X j ∈R B×C / 8×1×W ; The aforementioned sub-feature X i ∈R B×C / 8×H×1 and sub-feature X j ∈R B×C / 8×1×W After passing the SUCK module, the aforementioned X i ∈R B ×C / 8×H×1 With X j ∈R B×C / 8×1×W Aggregate by rows and columns separately and normalize them using group normalization with K groups to obtain X∈R B×C×H×1 and X∈R B×C×1×W , where K is a positive integer; After normalization, X∈R B×C×H×1 and X∈R B×C×1×W After passing through the sigmoid function respectively, combined with the aforementioned given input X∈R B×C×H×W Generate spatial attention X S ; In the aforementioned spatial attention X S After using average pooling and group normalization, the query Q, key K, and value V based on self-attention are divided into three branches. After performing a depth-wise separable convolution operation on each branch, the query attention weight, key attention weight, and value attention weight are obtained by calculating the query Q, key K, and value V respectively. Combined with the aforementioned query attention weight, key attention weight and value attention weight, after the context-aware self-attention mechanism, after average pooling and Sigmoid function, combined with the aforementioned spatial attention X S Perform multiplication operation to obtain the output feature map.

6. The method according to claim 1, characterized in that The multi-layer fusion module includes a front fusion module corresponding to each layer of the front encoder of the dual encoder structure and a back fusion module corresponding to each layer of the decoder structure; wherein, The front fusion module is capable of receiving the image features input by each layer in the front encoder, and after gradually changing the number of channels and image size through multiple convolutions, performing a residual connection with the output layer of the front encoder; The post-fusion module can gradually adjust the number of channels and image size of the image features output by each layer in the decoder structure through multiple deconvolution operations, and then perform residual connection with the image features output by the bottom layer.

7. A U-Net-based neck ultrasound segmentation device according to any one of claims 1 to 6, characterized in that include: A data construction unit is used to construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes several cervical ultrasound images that have been processed by data cropping, data labeling and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a jump connection structure and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure respectively correspond to the front fusion module and the back fusion module using the multi-layer fusion module to promote the image features to be transmitted to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that the image features are extracted from the dual encoder structure. The corresponding layer of the post-encoder is passed to the corresponding layer of the decoder structure; the GAG ​​module is a global attention gating module; the GAG ​​module includes query embedding, key embedding and value embedding; when executing the GAG ​​module, it specifically includes: using the aforementioned key embedding and query embedding to operate on the input image features to obtain key attention weights and query attention weights; the aforementioned key attention weights and query attention weights are subjected to residual connection, and after passing through the Gelu function, 1×1 convolution and sigmoid function, they are multiplied with the value embedding; wherein, when there are multiple GAG ​​modules, the image features output by the upper-layer GAG module are adjusted to the image size through the deconvolution operation and then input into the lower-layer GAG module, and after adjusting the number of channels through 1×1 convolution, the residual connection is performed with the result obtained by the aforementioned multiplication operation to highlight the global image features of the cervical ultrasound image; A model building unit, configured to use the processed neck ultrasound image as sample data for training and testing the U-Net segmentation model, and apply the tested U-Net segmentation model to the segmentation and recognition task of the neck ultrasound image; An image acquisition unit, used for acquiring an ultrasonic image of the neck through an ultrasonic probe device; The image segmentation and recognition unit is used to segment and recognize the collected cervical ultrasound image using the aforementioned U-Net segmentation model to extract image features of the target tissue in the aforementioned cervical ultrasound image.

8. A U-Net-based neck ultrasound segmentation system according to any one of claims 1 to 6, characterized in that include: Network node, used to send and receive neck ultrasound images; An image segmentation module, used to identify target tissue in the aforementioned neck ultrasound image and perform segmentation processing; A system server, the system server connecting the network nodes and the image segmentation module; The system server is configured to: construct a cervical ultrasound dataset and a U-Net segmentation model; the cervical ultrasound dataset includes several cervical ultrasound images that have been processed by data cropping, data labeling and data enhancement operations; the U-Net segmentation model can retain the image features of the target tissue when extracting image features; the target tissue is divided into small target tissue and large target tissue according to the size and range of the anatomical structure; the U-Net segmentation model includes a dual encoder structure, a jump connection structure and a decoder structure, wherein the front encoder in the dual encoder structure and the aforementioned decoder structure respectively correspond to the front fusion module and the back fusion module using a multi-layer fusion module to promote the image features to be transmitted to different layers in the dual encoder structure and the decoder structure respectively; the jump connection structure includes at least one GAG ​​module to ensure that the image features are transmitted from the corresponding layer of the post encoder in the dual encoder structure to the corresponding layer of the decoder structure; the GAG ​​module is a global attention gating module; the GAG ​​module includes query embedding, key embedding and value embedding; when executing the GAG ​​module When , it specifically includes: using the aforementioned key embedding and query embedding to operate on the input image features to obtain key attention weights and query attention weights; taking residual connection for the aforementioned key attention weights and query attention weights, and after passing through Gelu function, 1×1 convolution and sigmoid function, multiplying them with value embedding; wherein, when there are multiple GAG ​​modules, the image features output by the upper-layer GAG module are input into the lower-layer GAG module after adjusting the image size through deconvolution operation, and after adjusting the number of channels through 1×1 convolution, residual connection is performed with the result obtained by the aforementioned multiplication operation to highlight the global image features of the neck ultrasound image; the processed neck ultrasound image is used as sample data into the training and testing of the U-Net segmentation model, and the U-Net segmentation model that has completed the test is applied to the segmentation and recognition task of the neck ultrasound image; the neck ultrasound image is collected by an ultrasound probe device; the collected neck ultrasound image is segmented and recognized using the aforementioned U-Net segmentation model to extract the image features of the target tissue in the aforementioned neck ultrasound image.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps of any one of claims 1 to 6 are implemented.

Images