Recognition method for recurrent laryngeal nerves and adjacent anatomical structures thereof in ultrasonic image

Abstract

The invention relates to the technical field of medical information processing, and particularly discloses a method for identifying a recurrent laryngeal nerve and an adjacent anatomical structure thereof in an ultrasonic image, and the method comprises the steps: obtaining a two-dimensional thyroid region ultrasonic image, and carrying out the fine marking of the recurrent laryngeal nerve, the true thyroid envelope, the false thyroid envelope and the connective tissue; performing uniform size adjustment, normalization and data enhancement on the marked image; performing semantic segmentation on the image to obtain a segmented image containing the recurrent laryngeal nerve and an adjacent anatomical structure region; extracting depth features of the local structure of the segmented image; extracting high-dimensional geometric structure features of the segmented image; fusing the extracted features to form a comprehensive feature vector; and inputting the fused comprehensive feature vector into a classification module to perform automatic classification and identification of the recurrent laryngeal nerve and the adjacent structure thereof. According to the invention, the recognition accuracy and stability in the ultrasonic image can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information processing, and more particularly, to a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image. Background Art

[0002] Accurate identification of the recurrent laryngeal nerve is of great significance for avoiding serious complications such as postoperative vocal cord paralysis in thyroid surgery. At present, the clinical methods for identifying the recurrent laryngeal nerve mainly include three categories: intraoperative naked eye identification, intraoperative electrophysiological monitoring, and ultrasound evaluation. Among them, intraoperative naked eye identification is limited by the experience of the surgeon, and intraoperative bleeding and surgical operation will affect the identification effect. The current measure to avoid recurrent laryngeal nerve injury during surgery is intraoperative electrophysiological monitoring. Whether the structure touched by the nerve monitoring probe can cause a change in the laryngeal point position is used to determine whether the structure is the recurrent laryngeal nerve. However, this method cannot help clinicians estimate the anatomical position of the recurrent laryngeal nerve, and its role in protecting the recurrent laryngeal nerve is limited. Therefore, from the perspective of clinical practice and demand, electrophysiological monitoring cannot replace anatomical evaluation.

[0003] Ultrasound imaging, as a non-invasive, safe, and real-time imaging method, is widely used in thyroid disease screening and preoperative evaluation. However, due to problems such as strong noise interference, blurred boundaries, and low tissue contrast in ultrasound images, the tiny structure of the recurrent laryngeal nerve is often difficult to identify in the image. At the same time, differences in imaging methods between different operators will cause inconsistent image quality, further increasing the difficulty of identifying the recurrent laryngeal nerve using traditional image processing methods. In addition, the spatial relationship between the recurrent laryngeal nerve and adjacent structures such as the trachea, thyroid gland, and blood vessels in ultrasound images is complex, which increases the challenges of structural segmentation and positioning. Therefore, how to accurately identify the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images is of great significance. Summary of the Invention

[0004] The present invention provides a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images, which solves the problem that the recurrent laryngeal nerve and its adjacent anatomical structures in existing ultrasound images have tiny and weak contrast, which easily leads to inaccurate identification. It can improve the accuracy and stability of identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images.

[0005] To achieve the above objectives, the present invention provides the following technical solutions:

[0006] A method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image, comprising:

[0007] Two-dimensional ultrasound images of the thyroid region were obtained, and experienced ultrasound physicians carefully marked the recurrent laryngeal nerve, true thyroid capsule, false thyroid capsule, and connective tissue.

[0008] Perform uniform resizing, normalization, and data augmentation on the annotated images;

[0009] The DeepLabV3+ network was used to perform semantic segmentation on the image to obtain a segmented image containing the recurrent laryngeal nerve and adjacent anatomical structures.

[0010] Using DenseNet to extract deep features of the local structure of the segmented image;

[0011] A Grassmann manifold method is introduced to extract high-dimensional geometric structural features of the segmented image;

[0012] The features extracted from DenseNet and Grassmann manifold are fused to form a comprehensive feature vector;

[0013] The fused comprehensive feature vector is input into the classification module for automatic classification and identification of the recurrent laryngeal nerve and its adjacent structures.

[0014] Preferably, the detailed marking of the recurrent laryngeal nerve, true thyroid capsule, false thyroid capsule and connective tissue by an experienced ultrasound physician includes:

[0015] The two layers of high echo lines on the dorsal side of the thyroid parenchyma from ventral to dorsal are marked as the true thyroid capsule and the false thyroid capsule respectively;

[0016] The oval, hyperechoic, cribriform structure on the dorsal side of the thyroid pseudocapsule is labeled as the recurrent laryngeal nerve;

[0017] The irregular, hyperechoic, non-cribriform structure on the dorsal side of the thyroid pseudocapsule is labeled as connective tissue.

[0018] Preferably, the uniform resizing, normalization and data enhancement of the labeled images include:

[0019] After reading, the original ultrasound images were uniformly resized to 224×224, and the pixel data was normalized. The images were then rotated by 20°, 50°, and 90° to simulate the angular changes of clinical ultrasound probes. At the same time, the geometric changes of the images were increased by translation and horizontal flipping, thereby expanding the training samples.

[0020] Preferably, the image semantic segmentation using the DeepLabV3+ network includes:

[0021] The DeepLabV3+ network structure is used to decode and segment the input image. The network structure includes: a backbone feature extraction network ResNet101, a void spatial pyramid pooling ASPP, a low-level feature guidance branch and a decoder.

[0022] Preferably, the DeepLabV3+ network structure is used to decode and segment the input image, including:

[0023] First, the input image is passed through the backbone network for feature extraction. The backbone network uses the pre-trained ResNet101. The input image first undergoes 7×7 convolution, batch normalization, ReLU activation function and maximum pooling layer to complete preliminary feature extraction and downsampling. Then it passes through four residual blocks in sequence to obtain feature information from shallow to deep layers.

[0024] Then, the high-level features from the layer4 residual block are fed into the hole-space pyramid pooling (ASPP) to continue multi-scale feature extraction. The multi-scale features output by ASPP are then fused through 1×1 convolution to obtain a representation with rich contextual information.

[0025] A low-level feature guidance branch is introduced to pass the low-level features extracted from the layer 1 residual block of the backbone network through the convolution layer, batch normalization and ReLU in sequence to form a concise and efficient edge feature representation;

[0026] The multi-scale features output by ASPP are upsampled to keep their spatial size consistent with the low-level features, and the two are concatenated in the channel dimension to form a fused feature map.

[0027] Preferably, the method of decoding and segmenting the input image using the DeepLabV3+ network structure further includes:

[0028] In the decoder, the fused feature map is sequentially passed through two convolutional layers, batch normalization, ReLU, and a third convolutional layer to further extract the nonlinear relationship between the fused features and improve the boundary recognition capability.

[0029] Finally, the decoded output is upsampled back to the original image size to obtain a semantic segmentation result of the same size as the input image, enabling accurate identification and labeling of the recurrent laryngeal nerve and its adjacent anatomical structures in the image.

[0030] Preferably, the extracting the deep features of the local structure of the segmented image using DenseNet includes:

[0031] The initial convolutional block of the DenseNet network is used as the initial feature extraction layer of the network to receive the input image. A 5×5 convolution kernel is used to expand the number of channels to 64, and batch normalization, ReLU, and Dropout are combined to extract edge and texture features.

[0032] Next, the number of output channels is controlled by the densely connected blocks of the DenseNet network;

[0033] Finally, the output convolution block of the DenseNet network compresses and maps the high-dimensional feature map output by the densely connected block.

[0034] Preferably, the method of introducing the Grassmann manifold method to extract high-dimensional geometric structural features of the image includes:

[0035] By capturing the inherent low-dimensional structural features of the input image data and mapping it to the Grassmann manifold space for analysis, the input data X is first subjected to singular value decomposition: , get the direction of its principal component; by retaining the previous Main direction , construct the input sample in The projection of the dimensional subspace to Each column vector of , construct the inner product matrix , and finally flatten the upper triangular part of the inner product matrix into a vector as the final Grassmann feature.

[0036] Preferably, the fusing of features extracted from DenseNet and Grassmann manifolds comprises:

[0037] First, the two features are flattened and converted into one-dimensional vectors. Next, the two features are directly concatenated to form a higher-dimensional joint feature vector to preserve the diversity of the original information, or the two features are aligned and fused to enhance the correlation and complementarity between different features.

[0038] Preferably, the step of inputting the fused comprehensive feature vector into a classification module for automatic classification and identification of the recurrent laryngeal nerve and its adjacent structures includes:

[0039] The fused comprehensive features are input into the fully connected layer, and the multi-category probabilities are output through the softmax function, and the category with the highest probability score is taken as the final classification category.

[0040] This paper provides a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images. This method leverages the image segmentation and feature extraction capabilities of DeepLabV3+ and DenseNet, along with the modeling advantages of Grassmann manifolds for image geometry, to achieve more accurate positioning and classification in ultrasound images with complex backgrounds. This method addresses the problem of inaccurate identification of the recurrent laryngeal nerve and its adjacent anatomical structures, often caused by subtle and weak contrast in existing ultrasound images. It improves the accuracy and stability of ultrasound image recognition of the recurrent laryngeal nerve and its adjacent anatomical structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments.

[0042] Figure 1 The present invention provides a flow chart of a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image.

[0043] Figure 2 1 is a schematic structural diagram of a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image provided by an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the process of labeling ultrasound images by an ultrasound physician according to an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of the image processing process provided by an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the structure of the DeepLabV3+ image segmentation module provided by the present invention;

[0047] Figure 6 This is a schematic diagram of the structure of the DenseNet feature extraction module provided by the present invention;

[0048] Figure 7 A schematic diagram of the structure of the Grassmann manifold feature extraction module provided by the present invention;

[0049] Figure 8 A schematic diagram of the feature fusion process provided by the present invention; DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings and implementation methods.

[0051] In view of the problem that the recurrent laryngeal nerve and adjacent anatomical structures in current ultrasound images have tiny and weak contrast, which easily leads to inaccurate identification, the present invention provides a method for identifying the recurrent laryngeal nerve and adjacent anatomical structures in ultrasound images, which solves the problem that the recurrent laryngeal nerve and adjacent anatomical structures in current ultrasound images have tiny and weak contrast, which easily leads to inaccurate identification, and can improve the accuracy and stability of identifying the recurrent laryngeal nerve and adjacent anatomical structures in ultrasound images.

[0052] like Figure 1 As shown, a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image includes:

[0053] S1: Obtain a two-dimensional ultrasound image of the thyroid region, and have an experienced ultrasound physician carefully mark the recurrent laryngeal nerve, true thyroid capsule, false thyroid capsule, and connective tissue.

[0054] S2: Perform uniform resizing, normalization, and data augmentation on the annotated images.

[0055] S3: The DeepLabV3+ network is used to perform semantic segmentation on the image to obtain a segmented image containing the recurrent laryngeal nerve and adjacent anatomical structures.

[0056] S4: Using DenseNet to extract deep features of the local structure of the segmented image.

[0057] S5: Introduce the Grassmann manifold method to extract high-dimensional geometric structure features of the segmented image.

[0058] S6: Fuse the features extracted from DenseNet and Grassmann manifold to form a comprehensive feature vector.

[0059] S7: The fused comprehensive feature vector is input into the classification module for automatic classification and identification of the recurrent laryngeal nerve and its adjacent structures.

[0060] Specifically, the DeepLabV3+ network, through multi-scale feature extraction and decoding modules, can effectively improve the boundary clarity and target area integrity of the segmented image; the DenseNet structure uses a dense connection mechanism to enhance feature transfer, allowing small structures such as the recurrent laryngeal nerve to be more fully expressed in the feature map; and the Grassmann manifold module further supplements the geometric information missing in traditional feature extraction. This method analyzes ultrasound images based on DeepLabV3+, DenseNet, and Grassmann manifolds, and can automatically identify and classify the recurrent laryngeal nerve and its adjacent anatomical structures. This method fully combines the image segmentation and feature extraction capabilities of deep learning with the modeling advantages of Grassmann manifolds for image geometric structure, and can achieve more accurate positioning and classification in ultrasound images with complex backgrounds.

[0061] like Figure 2 As shown, in one embodiment, the automatic identification of the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image comprises the following specific steps:

[0062] Step 1: Obtain a two-dimensional ultrasound image, and have an experienced ultrasound physician annotate the recurrent laryngeal nerve and its adjacent anatomical structures in the ultrasound image.

[0063] Step 2: Preprocess the labeled images, adjust them to a uniform size (e.g. 224×224), normalize them, and amplify the data using rotation, translation, and other methods.

[0064] Step 3: Use DeepLabV3+ to perform image segmentation, obtaining a segmented image that includes the recurrent laryngeal nerve and adjacent anatomical structures. This method uses atrous spatial pyramid pooling (ASPP) for multi-scale feature extraction and enhances segmentation accuracy through a custom decoder.

[0065] Step 4: Input the segmented image into DenseNet for deep feature extraction. DenseNet improves the efficiency of feature transfer through dense connections and can better extract local detail features.

[0066] Step 5: The segmented image is fed into the Grassmann manifold feature extraction module to extract high-dimensional geometric features. Singular value decomposition is used to extract low-dimensional representations from the image, thereby capturing the image's geometric structure information.

[0067] Step 6: Fuse the features extracted from DenseNet and Grassmann manifold to form a comprehensive feature vector. The fused features can contain the image's detailed information and high-dimensional geometric information.

[0068] Step 7: Input the fused features into the classification module for multi-classification tasks.

[0069] Further, if Figure 3 As shown in the figure, ultrasound images can be annotated by experienced ultrasound physicians. The annotation contents include:

[0070] (1) Labeling the dorsal thyroid capsule: The two layers of high echo lines on the dorsal side of the thyroid parenchyma from ventral to dorsal are labeled as the true thyroid capsule and the false thyroid capsule respectively.

[0071] (2) The oval, hyperechoic, sieve-like structure on the dorsal side of the thyroid pseudocapsule is labeled the recurrent laryngeal nerve.

[0072] (3) Label the irregular, high-echoic, non-sieve-like structure on the dorsal side of the thyroid pseudocapsule as connective tissue.

[0073] like Figure 4 As shown in Figure 2, in order to ensure the consistency of the model input and enhance the generalization ability of the model, the labeled image needs to be preprocessed. First, the original ultrasound image is read and uniformly adjusted to a size of 224×224 pixels; then the pixel values are converted from Range normalized to Range, the formula is ; Data augmentation strategies are then applied, including 、 and Image rotation is performed to simulate the angle change of clinical ultrasound probes, and horizontal flipping is performed to increase the geometric changes of the image to expand the training samples.

[0074] Furthermore, the processed ultrasound image is input into the DeepLabV3+ image segmentation module. The network structure decodes and segments the input image, and can recognize the recurrent laryngeal nerve and its adjacent anatomical structures while retaining high-resolution information. Figure 5 As shown in the figure, it includes the backbone feature extraction network ResNet101, the void space pyramid pooling ASPP, the low-level feature guidance branch and the decoder. The specific process is as follows:

[0075] First, the input image passes through the backbone network for feature extraction. This backbone network uses a pre-trained ResNet101, which consists of several layers of residual architecture. It extracts multi-level image features through layer-by-layer convolution. The input image undergoes initial feature extraction and downsampling through a 7×7 convolution, batch normalization, ReLU activation, and max pooling layer. It then passes through four residual blocks (layers 1 through 4) to obtain feature information from shallow to deep layers. The output of layer 1 retains more edge information, known as low-level features; the features output by layer 4 contain high-level semantic information, known as high-level features.

[0076] The high-level features from layer 4 are then fed into ASPP for further multi-scale feature extraction. ASPP processes input at multiple scales using dilated convolutions with varying dilation rates, effectively improving the model's adaptability to varying object sizes and enhancing the spatial receptive field of the features. The multi-scale features output by ASPP are then fused through 1×1 convolutions to produce a representation rich in contextual information.

[0077] Next, to balance image detail and high-level semantics, DeepLabV3+ introduces a low-level feature guidance branch. Specifically, low-level features extracted from layer 1 of the backbone network are sequentially passed through convolutional layers, batch normalization, and a ReLU rectification layer to form a concise and efficient edge feature representation.

[0078] The features output by ASPP are then upsampled to keep their spatial size consistent with the low-level features, and the two are concatenated in the channel dimension to form a fused feature map. This fusion strategy effectively integrates the local edge structure and global semantic information of the image.

[0079] In the decoder part, the fused feature map passes through two convolutional layers, batch normalization and correction layer ReLU and the third convolutional layer in sequence to further extract the nonlinear relationship between the fused features and improve the boundary recognition ability.

[0080] Finally, the decoded output is upsampled back to the original image size to obtain a semantic segmentation result of the same size as the input image, enabling accurate identification and labeling of the recurrent laryngeal nerve and its adjacent anatomical structures in the image.

[0081] Step 4: The segmented image is passed through the DenseNet module for deep feature extraction. The network structure of this module adopts the improved lightweight DenseNet network structure, such as Figure 6 As shown in the figure, it includes the initial convolution block, dense connection block, and output convolution block. This design aims to efficiently extract local edge features in ultrasound images, enhance the fusion and transfer capabilities of features at different levels, and improve the model's ability to discriminate target areas. The specific processing flow is as follows:

[0082] First, the initial convolution block receives the input image as the initial feature extraction layer of the network, uses a 5×5 convolution kernel to expand the number of channels to 64, and combines batch normalization, ReLU, and Dropout to enable the model to extract lower-level edge and texture features in the early stage.

[0083] Next, we enter the core densely connected block, which consists of multiple densely connected convolutional layers. Each layer consists of a convolutional block with a kernel size of 5×5. The number of output channels is controlled by the growth rate, for example, set to 16. The output of each layer is concatenated with the outputs of all previous layers along the channel dimension to form the tensor input to the next layer. Taking three convolutional layers as an example, the final output channel number of the densely connected block is 64 + 16 × 3 = 112.

[0084] Finally, the output convolution block compresses and maps the high-dimensional feature maps output by the densely connected blocks. A convolution layer first reduces the 112-dimensional channel count to 64. After batch normalization and activation, another convolution layer compresses it to the final channel count, providing compact and expressive deep features for subsequent modules.

[0085] Step 5: The segmented image is passed through the Grassmann manifold feature extraction module for high-dimensional geometric feature extraction. This module captures the inherent low-dimensional structural features of the input image data and maps them to the Grassmann manifold space for analysis, such as Figure 7 shown.

[0086] Specifically, the Grassmann manifold Represents the set of k-dimensional subspaces in all n-dimensional linear spaces. The Grassmann manifold feature extraction module first performs Perform singular value decomposition: , and obtain the direction of its principal component. Main direction , construct the input sample in The projection of the dimensional subspace, that is, Each column vector of , construct the inner product matrix This is the embedding of high-dimensional image features into the k-dimensional Grassmann manifold Finally, the upper triangular part of the inner product matrix is flattened into a vector as the final Grassmann feature .

[0087] Step 6: The features extracted by the DenseNet feature extraction module and the Grassmann manifold feature extraction module are fused through the feature fusion module. Specifically, the two features are first flattened and converted into a one-dimensional vector, which is recorded as and , then there are two feature fusion strategies, such as Figure 8 shown.

[0088] One is to directly concatenate to form a higher-dimensional joint feature vector: .

[0089] The second is alignment fusion. First, zero-fill the shorter feature vector to make it consistent with the length of the longer vector, which is recorded as , and then apply the following strategies to further integrate and enhance the correlation and complementarity between different features:

[0090] (1) Weighted sum: ;

[0091] (2) Outer product: ;

[0092] (3) Attention mechanism: ,in , , and are trainable weights.

[0093] Step 7: Pass the fused features through the fully connected layer and softmax of the classification module to obtain the final category output probability, and take the category with the highest probability score as the final classification category.

[0094] Specifically, the classification module is pre-trained before use. Labeled segmentation images and ultrasound images are used as training samples in the pre-training. The learning rate and loss value corresponding to each training cycle are obtained during training to construct a functional relationship between the learning rate and the loss value. The corresponding loss value in the pre-training process is then smoothed, and the maximum and minimum learning rates for common use are determined, thereby improving the performance of the image classification module.

[0095] As can be seen, the present invention provides a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images. It fully combines the image segmentation and feature extraction capabilities of DeepLabV3+ and DenseNet, as well as the modeling advantages of Grassmann manifolds for image geometry, and can achieve more accurate positioning and classification in ultrasound images with complex backgrounds. It solves the problem of small and weak contrast in existing ultrasound images of the recurrent laryngeal nerve and its adjacent anatomical structures, which easily leads to inaccurate identification. It can improve the accuracy and stability of the recognition of the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images.

[0096] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image, characterized in that: include: Two-dimensional ultrasound images of the thyroid region were obtained, and experienced ultrasound physicians carefully marked the recurrent laryngeal nerve, true thyroid capsule, false thyroid capsule, and connective tissue. Perform uniform resizing, normalization, and data augmentation on the annotated images; The DeepLabV3+ network was used to perform semantic segmentation on the image to obtain a segmented image containing the recurrent laryngeal nerve and adjacent anatomical structures. Using DenseNet to extract deep features of the local structure of the segmented image; A Grassmann manifold method is introduced to extract high-dimensional geometric structural features of the segmented image; The features extracted from DenseNet and Grassmann manifold are fused to form a comprehensive feature vector; The fused comprehensive feature vector is input into the classification module for automatic classification and identification of the recurrent laryngeal nerve and its adjacent structures.

2. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 1, wherein: Experienced ultrasound physicians will carefully mark the recurrent laryngeal nerve, true thyroid capsule, false thyroid capsule, and connective tissue, including: The two layers of high echo lines on the dorsal side of the thyroid parenchyma from ventral to dorsal are marked as the true thyroid capsule and the false thyroid capsule respectively; The oval, hyperechoic, cribriform structure on the dorsal side of the thyroid pseudocapsule is labeled as the recurrent laryngeal nerve; The irregular, hyperechoic, non-cribriform structure on the dorsal side of the thyroid pseudocapsule is labeled as connective tissue.

3. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 2, wherein: The uniform resizing, normalization and data augmentation of the labeled images include: After reading, the original ultrasound images were uniformly resized to 224×224, and the pixel data was normalized. The images were then rotated by 20°, 50°, and 90° to simulate the angular changes of clinical ultrasound probes. At the same time, the geometric changes of the images were increased by translation and horizontal flipping, thereby expanding the training samples.

4. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 3, wherein: The DeepLabV3+ network is used to perform semantic segmentation on the image, including: The DeepLabV3+ network structure is used to decode and segment the input image. The network structure includes: a backbone feature extraction network ResNet101, a void spatial pyramid pooling ASPP, a low-level feature guidance branch and a decoder.

5. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 4, wherein: The DeepLabV3+ network structure is used to decode and segment the input image, including: First, the input image is passed through the backbone network for feature extraction. The backbone network uses the pre-trained ResNet101. The input image first undergoes 7×7 convolution, batch normalization, ReLU activation function and maximum pooling layer to complete preliminary feature extraction and downsampling. Then it passes through four residual blocks in sequence to obtain feature information from shallow to deep layers. Then, the high-level features from the layer4 residual block are fed into the hole-space pyramid pooling (ASPP) to continue multi-scale feature extraction. The multi-scale features output by ASPP are then fused through 1×1 convolution to obtain a representation with rich contextual information. A low-level feature guidance branch is introduced to pass the low-level features extracted from the layer 1 residual block of the backbone network through the convolution layer, batch normalization and ReLU in sequence to form a concise and efficient edge feature representation; The multi-scale features output by ASPP are upsampled to keep their spatial size consistent with the low-level features, and the two are concatenated in the channel dimension to form a fused feature map.

6. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 5, wherein: The DeepLabV3+ network structure is used to decode and segment the input image, and further includes: In the decoder, the fused feature map is sequentially passed through two convolutional layers, batch normalization, ReLU, and a third convolutional layer to further extract the nonlinear relationship between the fused features and improve the boundary recognition capability. Finally, the decoded output is upsampled back to the original image size to obtain a semantic segmentation result of the same size as the input image, enabling accurate identification and labeling of the recurrent laryngeal nerve and its adjacent anatomical structures in the image.

7. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 6, characterized in that: The extracting the deep features of the local structure of the segmented image by using DenseNet includes: The initial convolutional block of the DenseNet network is used as the initial feature extraction layer of the network to receive the input image. A 5×5 convolution kernel is used to expand the number of channels to 64, and batch normalization, ReLU, and Dropout are combined to extract edge and texture features. Next, the number of output channels is controlled by the densely connected blocks of the DenseNet network; Finally, the output convolution block of the DenseNet network compresses and maps the high-dimensional feature map output by the densely connected block.

8. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 7, wherein: The introduction of the Grassmann manifold method to extract high-dimensional geometric structural features of the image includes: By capturing the inherent low-dimensional structural features of the input image data and mapping it to the Grassmann manifold space for analysis, the input data X is first subjected to singular value decomposition: , get the direction of its principal component; by retaining the previous Main direction , construct the input sample in The projection of the dimensional subspace to Each column vector of , construct the inner product matrix , and finally flatten the upper triangular part of the inner product matrix into a vector as the final Grassmann feature.

9. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 8, wherein: The features extracted from DenseNet and Grassmann manifold are fused, including: First, the two features are flattened and converted into one-dimensional vectors. Next, the two features are directly concatenated to form a higher-dimensional joint feature vector to preserve the diversity of the original information, or the two features are aligned and fused to enhance the correlation and complementarity between different features.

10. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image according to claim 9, wherein: The step of inputting the fused comprehensive feature vector into a classification module for automatic classification and identification of the recurrent laryngeal nerve and its adjacent structures includes: The fused comprehensive features are input into the fully connected layer, and the multi-category probabilities are output through the softmax function, and the category with the highest probability score is taken as the final classification category.

Images