A method of identifying the recurrent laryngeal nerve and its adjacent anatomical structures in an ultrasound image

Abstract

The application relates to the technical field of medical information processing, and particularly discloses a method for identifying a recurrent laryngeal nerve and adjacent anatomical structures in an ultrasonic image, which comprises the following steps: acquiring a two-dimensional thyroid region ultrasonic image, and finely labeling a recurrent laryngeal nerve, a thyroid true capsule, a thyroid false capsule and connective tissue; uniformly adjusting the size of the labeled image, normalizing the labeled image and enhancing data of the labeled image; performing semantic segmentation on the image to obtain a segmentation image containing a recurrent laryngeal nerve and adjacent anatomical structure regions; extracting deep features of local structures of the segmentation image; extracting high-dimensional geometric structure features of the segmentation image; fusing the extracted features to form a comprehensive feature vector; and inputting the fused comprehensive feature vector into a classification module to automatically classify and identify a recurrent laryngeal nerve and adjacent structures. The application can increase the recognition accuracy and stability in the ultrasonic image.

Description

Technical Field

[0001] This invention relates to the field of medical information processing, and more specifically, to a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images. Background Technology

[0002] Accurate identification of the recurrent laryngeal nerve is crucial for preventing serious complications such as vocal cord paralysis after thyroid surgery. Currently, clinical methods for identifying the recurrent laryngeal nerve mainly include intraoperative visual identification, intraoperative electrophysiological monitoring, and ultrasound evaluation. Intraoperative visual identification is limited by the surgeon's experience, and intraoperative bleeding and surgical manipulation can affect the identification results. Currently, intraoperative electrophysiological monitoring is used to avoid recurrent laryngeal nerve injury during surgery. This involves using a nerve monitoring probe to determine if a structure touched by the probe causes a change in laryngeal position, thus identifying the recurrent laryngeal nerve. However, this method cannot help clinicians predict the anatomical location of the recurrent laryngeal nerve and has limited protective effect. Therefore, from a clinical practice and demand perspective, 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 recurrent laryngeal nerve, a tiny structure, is often difficult to identify in the images. Furthermore, differences in imaging techniques among different operators can lead to 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, increasing the challenge of structural segmentation and localization. Therefore, accurate identification of the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images is of great significance. Summary of the Invention

[0004] This invention provides a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images. This method solves the problem that the recurrent laryngeal nerve and its adjacent anatomical structures in existing ultrasound images have small and weak contrast, which easily leads to inaccurate identification. This method 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 solution:

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

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

[0008] The labeled images are then subjected to uniform size adjustment, normalization, and data augmentation.

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

[0010] The depth features of the local structure of the segmented image are extracted using DenseNet;

[0011] The Grassmann manifold method is introduced to extract the high-dimensional geometric structure 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 integrated feature vector is input into the classification module for automatic classification and recognition of the recurrent laryngeal nerve and its adjacent structures.

[0014] Preferably, the detailed annotation of the recurrent laryngeal nerve, true capsule, pseudo capsule, and connective tissue by an experienced sonographer includes:

[0015] The two hyperechoic lines on the dorsal side of the thyroid parenchyma, from the ventral side to the dorsal side, are labeled as the true capsule of the thyroid gland and the false capsule of the thyroid gland, respectively.

[0016] The oval, hyperechoic, crisscross-shaped structure marked on the dorsal side of the thyroid pseudocapsule is the recurrent laryngeal nerve.

[0017] The irregular, hyperechoic, and non-sieve-like structures on the dorsal side of the thyroid pseudocapsule are identified as connective tissue.

[0018] Preferably, the process of uniformly resizing, normalizing, and augmenting the labeled images includes:

[0019] After reading the original ultrasound images, they were uniformly adjusted to a size of 224×224, and the pixel data was normalized. Then, the images were rotated at 20°, 50°, and 90° to simulate the angle 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 step of using a DeepLabV3+ network to perform semantic segmentation of the image includes:

[0021] The DeepLabV3+ network structure is used to decode and segment the input image. The network structure includes: ResNet101 backbone feature extraction network, Spatial Pyramid Pooling (ASPP) with dilated features, low-level feature guidance branches, and a decoder.

[0022] Preferably, the step of decoding and segmenting the input image using the DeepLabV3+ network structure includes:

[0023] First, the input image undergoes feature extraction via a backbone network. This backbone network uses a pre-trained ResNet101. The input image first passes through a 7×7 convolution, batch normalization, ReLU activation function, and max pooling layer to complete preliminary feature extraction and downsampling. Then, it passes through four residual blocks sequentially 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 for further 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, which sequentially passes the low-level features extracted from the layer 1 residual block of the backbone network through convolutional layers, batch normalization, and ReLU to form a concise and efficient edge feature representation.

[0026] The multi-scale features output by ASPP are upsampled to make their spatial dimensions consistent with those of the low-level features, and then the two are concatenated along the channel dimension to form a fused feature map.

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

[0028] In the decoder section, the fused feature map is passed sequentially through two convolutional layers, batch normalization and 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, thereby achieving accurate identification and labeling of the recurrent laryngeal nerve and its adjacent anatomical structures in the image.

[0030] Preferably, the step of extracting the depth 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 convolutional 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 convolutional blocks of the DenseNet network compress and map the high-dimensional feature maps output by the densely connected blocks.

[0034] Preferably, the extraction of high-dimensional geometric structural features of the image using the Grassmann manifold method includes:

[0035] By capturing the inherent low-dimensional structural features of the input image data and mapping them to the Grassmann manifold space for analysis, the input data X is first subjected to singular value decomposition: The principal component orientations are obtained by retaining the previous ones. Main directions Constructing input samples in Projection of a dimensional subspace, for Each column vector Construct the inner product matrix Finally, the upper triangular part of this inner product matrix is ​​flattened into a vector, which is used as the final Grassmann feature.

[0036] Preferably, the fusion of features extracted from DenseNet and Grassmann manifold includes:

[0037] First, the two features are flattened into one-dimensional vectors. Then, 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 integrated feature vector into the classification module for automatic classification and recognition of the recurrent laryngeal nerve and its adjacent structures includes:

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

[0040] This 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 with the advantages of Grassmann manifolds in modeling image geometry, enabling more accurate localization 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 in existing ultrasound images due to their small and weak contrast, thus improving the accuracy and stability of identification of the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images. Attached Figure Description

[0041] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below.

[0042] Figure 1 This invention provides a flowchart illustrating a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images.

[0043] Figure 2 This is a schematic diagram of a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images, provided by an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the process of labeling ultrasound images by an ultrasound physician, as provided in an embodiment of the present invention.

[0045] Figure 4 This is a schematic diagram of the image processing flow provided in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram 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 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 Implementation

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

[0051] To address the problem of inaccurate identification of the recurrent laryngeal nerve and adjacent anatomical structures in current ultrasound images due to their small and weak contrast, this invention provides a method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images. This method solves the problem of inaccurate identification caused by the small and weak contrast in current ultrasound images, and improves the accuracy and stability of identifying the recurrent laryngeal nerve and its 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 ultrasound images includes:

[0053] S1: Obtain two-dimensional ultrasound images of the thyroid region, and have an experienced sonographer perform detailed annotations of the recurrent laryngeal nerve, true capsule, false capsule, and connective tissue.

[0054] S2: Perform uniform size adjustment, normalization, and data augmentation on the labeled images.

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

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

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

[0058] S6: The features extracted from DenseNet and Grassmann manifold are fused to form a comprehensive feature vector.

[0059] S7: Input the fused integrated feature vector into the classification module for automatic classification and recognition of the recurrent laryngeal nerve and its adjacent structures.

[0060] Specifically, the DeepLabV3+ network, through multi-scale feature extraction and decoding modules, effectively improves the boundary clarity and target region integrity of segmented images. The DenseNet structure, utilizing a dense connection mechanism, enhances feature transfer, allowing for more complete representation of minute structures such as the recurrent laryngeal nerve in the feature map. The Grassmann manifold module further supplements the geometric information missing in traditional feature extraction. This method, based on DeepLabV3+, DenseNet, and the Grassmann manifold, analyzes ultrasound images, enabling automatic identification and classification of 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 the Grassmann manifold for image geometry, achieving more accurate localization 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 involves the following steps:

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

[0063] Step 2: Preprocess the labeled images by adjusting them to a uniform size (e.g., 224×224), normalizing them, and augmenting the data using methods such as rotation and translation.

[0064] Step 3: Image segmentation is performed using the DeepLabV3+ method to obtain segmented images containing the recurrent laryngeal nerve and adjacent anatomical structures. This method uses Spatial Pyramid Pooling with Holes (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 feature transfer efficiency through dense connections, enabling better extraction of local detail features.

[0066] Step 5: Input the segmented image into the Grassmann manifold feature extraction module to extract high-dimensional geometric features. Low-dimensional representations are extracted from the image through singular value decomposition, thereby capturing the geometric structure information of the image.

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

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

[0069] Furthermore, such as Figure 3 As shown, experienced ultrasound physicians can annotate ultrasound images, and the annotations include:

[0070] (1) Mark the dorsal capsule of the thyroid gland: The two hyperechoic lines on the dorsal side of the thyroid parenchyma from the ventral side to the dorsal side are marked as the true capsule of the thyroid gland and the false capsule of the thyroid gland, respectively.

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

[0072] (3) The irregular shape, hyperechoic and sieveless structure on the dorsal side of the thyroid pseudocapsule is labeled as connective tissue.

[0073] like Figure 4 As shown, to ensure the consistency of model input and enhance the model's generalization ability, preprocessing of the labeled images is necessary. First, the original ultrasound images are read and uniformly resized to 224×224 pixels; then the pixel values ​​are... Range normalized to The range, the formula is Subsequently, data augmentation strategies were applied, including... , and The image is rotated to simulate the angular changes of a clinical ultrasound probe, while horizontal flipping increases the geometric changes in the image, thereby expanding the training sample.

[0074] Furthermore, the processed ultrasound images are input into the DeepLabV3+ image segmentation module. This network structure decodes and segments the input images, enabling the identification of the recurrent laryngeal nerve and its adjacent anatomical structures while preserving high-resolution information. The network structure is as follows: Figure 5 As shown, it includes the backbone feature extraction network ResNet101, the dilated spatial pyramid pooling (ASPP) network, low-level feature guidance branches, and a decoder. Its specific workflow is as follows:

[0075] First, the input image undergoes feature extraction via a backbone network. This backbone network uses a pre-trained ResNet101, which contains several residual layers and extracts multi-level features of the image through layer-by-layer convolution. The input image first passes through a 7×7 convolution, batch normalization, ReLU activation function, and max pooling layer to complete initial feature extraction and downsampling; then it passes through four residual blocks (layer1~layer4) sequentially to obtain feature information from shallow to deep layers. Among them, the output of layer1 retains more edge information and is called low-level features; while the features output of layer4 contain high-level semantic information and are called high-level features.

[0076] Then, the high-level features from layer 4 are fed into ASPP for further multi-scale feature extraction. ASPP uses dilated convolutions with different dilation rates to process the input at multiple scales, effectively improving the model's adaptability to different target 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 obtain a representation with rich contextual information.

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

[0078] Next, the features output by ASPP are upsampled to match the spatial dimensions of the low-level features; then, the two are concatenated along 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 section, the fused feature map is passed sequentially through two convolutional layers, batch normalization and correction layer ReLU, and a third convolutional layer to further extract the nonlinear relationships between the fused features and improve the boundary recognition capability.

[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, thereby achieving accurate identification and labeling of the recurrent laryngeal nerve and its adjacent anatomical structures in the image.

[0081] Step 4: The segmented image is processed through the DenseNet module for deep feature extraction. This module uses an improved, lightweight DenseNet network structure, such as... Figure 6 As shown, it includes an initial convolutional block, a densely connected block, and an output convolutional block. This design aims to efficiently extract local edge features from ultrasound images, enhance the fusion and transfer capabilities of features at different levels, and improve the model's ability to discriminate target regions. The specific processing flow is as follows:

[0082] First, the initial convolutional block, as the initial feature extraction layer of the network, receives the input image. A 5×5 convolutional kernel is used to expand the number of channels to 64, and batch normalization, ReLU, and Dropout are combined to enable the model to extract lower-level edge and texture features in the early stages.

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

[0084] Finally, the output convolutional block compresses and maps the high-dimensional feature map output by the densely connected block. First, a convolutional layer reduces the number of channels from 112 to 64. After batch normalization and activation, another convolutional layer compresses it to the final number of channels, providing compact and expressive deep features for subsequent modules.

[0085] Step 5: The segmented image is processed by the Grassmann manifold feature extraction module to extract high-dimensional geometric features. 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 As shown.

[0086] Specifically, Grassmann manifold The Grassmann manifold feature extraction module first processes the input data to represent the set of all k-dimensional subspaces in an n-dimensional linear space. Perform singular value decomposition: This yields the principal component orientations. By retaining the original... Main directions Constructing input samples in The projection of a dimensional subspace, i.e., for Each column vector Construct the inner product matrix This is the embedding of high-dimensional image features into a k-dimensional Grassmann manifold. The process involves flattening the upper triangular portion of this inner product matrix into a vector, which is then used as the final Grassmann feature. .

[0087] Step 6: The features extracted through the DenseNet feature extraction module and the Grassmann manifold feature extraction module are fused using the feature fusion module. Specifically, the two feature types are first flattened into one-dimensional vectors, denoted as... and Next, there are two feature fusion strategies, such as Figure 8 As shown.

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

[0089] The second is alignment and fusion. First, the shorter feature vector is zero-paddinged to make its length match that of the longer vector, denoted as... Then, the following strategy is applied to further integrate the features and enhance the correlation and complementarity between them:

[0090] (1) Weighted sum: ;

[0091] (2) External product: ;

[0092] (3) Attention mechanism: ,in , , and These 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 class output probability, and take the one with the highest probability score as the final classification class.

[0094] Specifically, the classification module is pre-trained before use. In the pre-training, labeled segmented images and ultrasound images are used as training samples. 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. Then, the loss value corresponding to the pre-training process is smoothed to determine the maximum and minimum learning rates for regular use, thereby improving the performance of the image classification module.

[0095] Therefore, this 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 advantages of Grassmann manifolds in modeling image geometry, enabling more accurate localization 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 in existing ultrasound images due to their small and weak contrast, thus improving the accuracy and stability of identification of the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images.

[0096] The structure, features, and effects of the present invention have been described in detail above with reference to the embodiments shown in the figures. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, shall be within the protection scope of the present invention as long as they do not exceed the spirit covered by the specification and figures.

Claims

1. A method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images, characterized in that, include: Two-dimensional ultrasound images of the thyroid region were obtained, and experienced ultrasound physicians meticulously annotated the recurrent laryngeal nerve, the true capsule, the false capsule, and the connective tissue. The labeled images are then subjected to uniform size adjustment, normalization, and data augmentation. The DeepLabV3+ network was used to perform semantic segmentation on the image to obtain segmented images containing the recurrent laryngeal nerve and adjacent anatomical structures. The depth features of the local structure of the segmented image are extracted using DenseNet; The Grassmann manifold method is introduced to extract the high-dimensional geometric structure 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 contains detailed information and high-dimensional geometric information of the image. The fused integrated feature vector is input into the classification module to automatically classify and identify the recurrent laryngeal nerve and its adjacent structures, and the one with the highest probability score is taken as the final classification category. The step of extracting deep features of the local structure of the segmented image 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 convolutional 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 convolutional blocks of the DenseNet network compress and map the high-dimensional feature maps output by the densely connected blocks; The method of introducing Grassmann manifolds to extract high-dimensional geometric structure features of images includes: By capturing the inherent low-dimensional structural features of the input image data and mapping them to the Grassmann manifold space for analysis, the input data X is first subjected to singular value decomposition: The principal component directions are obtained; by retaining the first K principal directions... Construct the projection of the input sample onto the K-dimensional subspace, that is, for Each column vector Construct the inner product matrix This is the embedding of high-dimensional geometric features into a k-dimensional Grassmann manifold. The process involves flattening the upper triangular portion of this inner product matrix into a vector, which is then used as the final Grassmann feature.

2. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 1, characterized in that, The recurrent laryngeal nerve, true capsule, pseudo capsule, and connective tissue were meticulously annotated by an experienced sonographer, including: The two hyperechoic lines on the dorsal side of the thyroid parenchyma, from the ventral side to the dorsal side, are labeled as the true capsule of the thyroid gland and the false capsule of the thyroid gland, respectively. The oval, hyperechoic, crisscross-shaped structure marked on the dorsal side of the thyroid pseudocapsule is the recurrent laryngeal nerve. The irregular, hyperechoic, and sieveless structures on the dorsal side of the thyroid pseudocapsule are identified as connective tissue.

3. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 2, characterized in that, The process of uniformly resizing, normalizing, and augmenting the labeled images includes: After reading the original ultrasound images, they were uniformly adjusted to a size of 224×224, and the pixel data was normalized. Then, the images were rotated at 20°, 50°, and 90° to simulate the angle 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, characterized in that, The semantic segmentation of images using the DeepLabV3+ network includes: The DeepLabV3+ network structure is used to decode and segment the input image. The network structure includes: ResNet101 backbone feature extraction network, Spatial Pyramid Pooling (ASPP) with dilated features, low-level feature guidance branches, and a decoder.

5. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 4, characterized in that, The process of decoding and segmenting the input image using the DeepLabV3+ network structure includes: First, the input image undergoes feature extraction via a backbone network. This backbone network uses a pre-trained ResNet101. The input image first passes through a 7×7 convolution, batch normalization, ReLU activation function, and max pooling layer to complete preliminary feature extraction and downsampling. Then, it passes through four residual blocks sequentially 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 for further 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, which sequentially passes the low-level features extracted from the layer 1 residual block of the backbone network through convolutional layers, batch normalization, and ReLU to form a concise and efficient edge feature representation. The multi-scale features output by ASPP are upsampled to make their spatial dimensions consistent with those of the low-level features, and then the two are concatenated along the channel dimension to form a fused feature map.

6. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 5, characterized in that, The method of decoding and segmenting the input image using the DeepLabV3+ network structure also includes: In the decoder section, the fused feature map is passed sequentially through two convolutional layers, batch normalization and 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, thereby achieving 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 fusion of features extracted from DenseNet and Grassmann manifold includes: First, the two features are flattened into one-dimensional vectors. Then, 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.

8. The method for identifying the recurrent laryngeal nerve and its adjacent anatomical structures in ultrasound images according to claim 7, characterized in that, The step of inputting the fused integrated feature vector into the classification module for automatic classification and recognition of the recurrent laryngeal nerve and its adjacent structures includes: The fused integrated features are input into a fully connected layer, and the multi-class probabilities are output through the softmax function. The class with the highest probability score is taken as the final classification class.

Images