A multi-part identification method and system for nasopharyngolaryngoscope images

Abstract

The application discloses a multi-part identification method and system for nasopharyngeal laryngoscope images, and relates to the technical field of medical image processing. The method comprises the following steps: inputting the nasopharyngeal laryngoscope image into a lightweight channel feature extraction module to extract channel global features and fine-grained dependent features; adding the channel global features and the fine-grained dependent features to obtain channel features; inputting the channel features into a local feature extraction module to extract small receptive field local features and large receptive field local features; adding the small receptive field local features and the large receptive field local features to obtain fusion features, and performing channel transformation and compression on the fusion features to obtain channel fusion features; and inputting the channel fusion features into a classifier to obtain identification results of each part. The lightweight channel feature extraction module, the local feature extraction module and the hybrid loss function optimize the defects of low classification accuracy caused by insufficient channel dimension feature modeling, high similarity between nasopharyngeal laryngeal part classes and serious imbalance of data categories.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a multi-part identification method and system for nasopharyngeal laryngoscope images. BACKGROUND

[0002] In the field of nasopharyngeal laryngoscopy, the existing part classification method still faces many challenges, mainly because there is high inter-class similarity between each part of the nasopharynx and throat, for example, left and right nasal cavities and left and right nasopharyngeal parts. Without obvious anatomical landmarks, the feature extraction capability of the deep learning model will be severely limited. The existence of these conditions puts higher standards on the classification model, which requires the model to have strong feature extraction and processing capabilities.

[0003] At present, the mainstream method mainly adopts CNN (Convolutional Neural Network) or Transformer model. CNN has excellent local information learning ability, but it lacks in capturing global information. Transformer has the advantage of capturing global information due to its long-range dependency modeling capability, but its computational efficiency is low and the inference speed is slow. The existence of these problems leads to the fact that although CNN and Transformer have certain feature extraction capabilities, they still cannot balance accuracy and efficiency in complex scenarios. SUMMARY

[0004] The embodiments of the present application provide a multi-part identification method and system for nasopharyngeal laryngoscope images, to solve the problem that the CNN and Transformer model in the prior art cannot balance accuracy and efficiency in the complex scenario of nasopharyngeal laryngoscope images.

[0005] In one aspect, the embodiments of the present application provide a multi-part identification method for nasopharyngeal laryngoscope images, comprising:

[0006] obtaining a nasopharyngeal laryngoscope image;

[0007] establishing a classification model for multiple parts of the nasopharyngeal laryngoscope, the classification model including four stages of feature extraction;

[0008] inputting the nasopharyngeal laryngoscope image into a lightweight channel feature extraction module, the lightweight channel feature extraction module including a parallel compression activation module and an efficient channel attention module; the compression activation module performs global average pooling and dimension reduction compression on the nasopharyngeal laryngoscope image to extract channel global features of the nasopharyngeal laryngoscope image; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image to extract fine-grained dependency features between channels of the nasopharyngeal laryngoscope image; and the channel global features and the fine-grained dependency features are added to obtain channel features;

[0009] The channel feature is input into a local feature extraction module, and the local feature extraction module includes a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path adopts a 3*3 depth separable convolution to extract small receptive field local features of the channel feature; the multi-scale receptive field path adopts a 5*5 depth separable convolution to extract large receptive field local features of the channel feature; the small receptive field local features and the large receptive field local features are added to obtain fusion features, and the fusion features are subjected to 1*1 point convolution for channel transformation and compression to obtain channel fusion features;

[0010] The channel fusion features are input into a classifier to obtain recognition results of each part in the nasopharyngeal laryngoscope image.

[0011] In the training process, the lightweight channel feature extraction module and the local feature extraction module adopt a hybrid loss function, and the hybrid loss function includes a cross-entropy loss, a label smoothing loss and a focal loss.

[0012] In a possible implementation, the compression activation module performs global average pooling and dimension reduction compression on the nasopharyngeal laryngoscope image, including:

[0013] The global average pooling is used to compress the spatial dimension of the nasopharyngeal laryngoscope image, and only the features of the channel dimension are reserved to obtain first pooling features;

[0014] The 1*1 convolution is used to compress the number of channels of the first pooling features to obtain compressed features;

[0015] The ReLU activation function is used to activate the compressed features to obtain restored features;

[0016] The Sigmoid activation function is used to normalize the restored features to obtain attention weights;

[0017] The attention weights and the nasopharyngeal laryngoscope image are multiplied channel by channel to obtain channel global features.

[0018] In a possible implementation, the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image, including:

[0019] The average pooling is used to extract second pooling features of the nasopharyngeal laryngoscope image;

[0020] The one-dimensional convolution is used to simulate the local dependency relationship between the channels in the second pooling features to obtain local dependency weights;

[0021] The local dependency weights and the nasopharyngeal laryngoscope image are multiplied channel by channel to obtain fine-grained dependency features.

[0022] In a possible implementation, the small receptive field local features are represented as:

[0023]

[0024] wherein, is a small receptive field local feature, is a channel feature, is a deep convolution weight, represents a convolution operation, and a convolution kernel size of 3x3 is adopted, is a dimension and an output channel number is a learnable bias vector of the same length, BN 1 represents a batch normalization process, is a ReLU activation function process.

[0025] In a possible implementation, the large receptive field local feature is represented as:

[0026]

[0027] wherein, is a large receptive field local feature, is a channel feature, is a multi-scale convolution kernel, represents a convolution operation, and a convolution kernel size of 5x5 is adopted, is a dimension and an output channel number is a learnable bias vector of the same length, BN 2 represents a batch normalization process, is a ReLU activation function process.

[0028] In a possible implementation, the fusion feature is represented as:

[0029]

[0030] The channel fusion feature is represented as:

[0031]

[0032] wherein, is a channel fusion feature, is a point convolution weight, is a dimension and an output channel number is a learnable bias vector of the same length, represents a batch normalization process.

[0033] In a possible implementation, after the nasopharyngolaryngoscope image is acquired, preprocessing is performed, and the preprocessing includes removing a black border area, rotation, and size adjustment.

[0034] In another aspect, the embodiment of the present application also provides a multi-part identification system for nasopharyngolaryngoscope images, comprising:

[0035] An image acquisition module is configured to acquire nasopharyngolaryngoscope images.

[0036] A model establishment module is configured to establish a classification model for multiple parts of the nasopharyngolaryngoscope, which comprises four stages of feature extraction.

[0037] The light-weight channel feature extraction module comprises a parallel compression activation module and an efficient channel attention module; the compression activation module is configured to perform global average pooling and dimension reduction compression on the nasopharyngolaryngoscope images to extract channel global features of the nasopharyngolaryngoscope images; and the efficient channel attention module is configured to perform average pooling and one-dimensional convolution on the nasopharyngolaryngoscope images to extract fine-grained dependency features between channels of the nasopharyngolaryngoscope images; and the channel global features and the fine-grained dependency features are added to obtain channel features.

[0038] The local feature extraction module comprises a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path is configured to extract small receptive field local features of the channel features by using 3*3 depth separable convolution; and the multi-scale receptive field path is configured to extract large receptive field local features of the channel features by using 5*5 depth separable convolution; and the small receptive field local features and the large receptive field local features are added to obtain fusion features, and the fusion features are subjected to 1*1 point convolution for channel transformation and compression to obtain channel fusion features.

[0039] A classifier is configured to classify to obtain identification results of each part in the nasopharyngolaryngoscope images.

[0040] In the training process, the light-weight channel feature extraction module and the local feature extraction module adopt a hybrid loss function, and the hybrid loss function comprises a cross-entropy loss, a label smoothing loss and a focal loss.

[0041] In another aspect, the embodiment of the present application also provides a computer storage medium, which stores a plurality of computer instructions for causing a computer to execute the above-mentioned method.

[0042] The multi-part identification method and system for nasopharyngolaryngoscope images have the following advantages:

[0043] 1. Based on the Mamba structure, combined with light-weight channel feature extraction (LCA), local feature extraction (LFE) and hybrid loss (CL) mechanism, the model is light-weighted while the ability to distinguish high-similarity parts is significantly improved, which is better than the performance of traditional CNN or Transformer model in complex endoscope image recognition.

[0044] 2. In view of the uneven data distribution and the difficulty of feature extraction in the nasopharyngeal and laryngeal parts, a refined feature modeling and loss function design strategy is proposed, which makes the model perform better in recognition accuracy and generalization ability on multiple data sets, and better meets the needs of real-time clinical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0046] Figure 1 A flowchart of a nasopharyngeal and laryngeal endoscope image multi-part recognition method provided by an embodiment of the present application.

[0047] Figure 2 A comparison chart of parameter quantity and F1 score of the recognition method provided by the embodiment of the present application and multiple comparison methods on the data set NIDB.

[0048] Figure 3 A comparison chart of classification performance of the recognition method provided by the embodiment of the present application and multiple comparison methods on the data set NIDB.

[0049] Figure 4 A confusion matrix result chart of the recognition method provided by the embodiment of the present application and multiple comparison methods on the data set NIDB.

[0050] Figure 5 A comparison chart of parameter quantity and F1 score of the recognition method provided by the embodiment of the present application and multiple comparison methods on the data set NIDI.

[0051] Figure 6 A comparison chart of classification performance of the recognition method provided by the embodiment of the present application and multiple comparison methods on the data set NIDI.

[0052] Figure 7 A confusion matrix result chart of the recognition method provided by the embodiment of the present application and multiple comparison methods on the data set NIDI. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] Figure 1 A flowchart of a multi-part identification method for nasopharyngeal endoscopic images is provided for the embodiments of the present application. The embodiments of the present application provide a multi-part identification method for nasopharyngeal endoscopic images, which comprises:

[0055] S100, acquiring a nasopharyngeal endoscopic image.

[0056] Exemplarily, after acquiring the nasopharyngeal endoscopic image, preprocessing is performed, which includes removing the black border area, rotating, and adjusting the size.

[0057] Specifically, the original endoscopic image contains an endoscope area and a black border area, wherein the black border area contains a large amount of irrelevant information. In the data preprocessing stage, the black border area is first cropped to remove irrelevant information. For different shooting methods of different medical institutions, the direction of the image needs to be uniformly rotated to ensure that the directions of all images are consistent. Subsequently, the processed image is uploaded to the model of the embodiments of the present application, which automatically adjusts the size of all images to a uniform 224x224x3, which is more suitable for subsequent part identification and classification tasks.

[0058] S110, establishing a classification model for nasopharyngeal endoscopic multi-parts, which includes four stages of feature extraction.

[0059] Exemplarily, the input of the classification model is a 224x224x3 nasopharyngeal endoscopic image, which is divided into 56x56xC1 by the patch partition module, and then the output tensor size remains unchanged after the first stage of L1 ENT-Mamba module feature extraction and transformation. After the second stage of downsampling and L2 ENT-Mamba module feature extraction and transformation, the output tensor is a 28x28xC2 feature tensor. After the third stage of downsampling and L3 ENT-Mamba module feature extraction and transformation, the output tensor is a 14x14xC3 feature tensor. Finally, after the fourth stage of downsampling and L4 ENT-Mamba modules, the output tensor is a 7x7xC4 feature tensor, generating global core semantic features for nasopharyngeal classification. The ENT-Mamba module includes a lightweight channel feature extraction module for improving the feature selection and representation capabilities of the classification model in the channel dimension, and a local feature extraction module for the classification model to capture subtle differences between similar categories and improve classification accuracy.

[0060] S120, the nasopharyngeal endoscopy image is input into the lightweight channel feature extraction module, which includes a parallel compression activation module and an efficient channel attention module; the compression activation module performs global average pooling and dimensionality reduction compression on the nasopharyngeal endoscopy image to extract the global channel features of the nasopharyngeal endoscopy image; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal endoscopy image to extract the fine-grained dependency features between the channels of the nasopharyngeal endoscopy image; the global channel features and the fine-grained dependency features are added together to obtain the channel features.

[0061] For example, to address the shortcomings of Efficient VMamba channel-dimensional modeling and feature fusion, this application proposes a lightweight channel feature extraction module (Lightweight Channel Attention, LCA). This module dynamically weights the criticality of each channel, improving the model's feature selection and representation capabilities in the channel dimension. This module consists of two parts: a squeeze-and-excitation (SE) module and an efficient channel attention (ECA) module. SE utilizes global pooling and two dimensionality reduction operations to enhance global channel dependency modeling, while ECA efficiently captures local interaction characteristics between channels through adaptive convolutional kernels.

[0062] Specifically, the compression activation module performs global average pooling and dimensionality reduction compression on the nasopharyngeal endoscopy images, including:

[0063] Global average pooling is used to compress the spatial dimensions of nasopharyngeal endoscopy images, retaining only channel-dimensional features, resulting in the first pooling feature. :

[0064]

[0065] in, For from the first b The first nasopharyngeal endoscopy image c The location extracted from each channel ( h , w The eigenvalue at position ) H and W These represent the height and width of the nasopharyngeal endoscopy image, respectively.

[0066] Use a 1×1 convolution to reduce the number of channels in the first pooling feature from... C Compress to To obtain compression features :

[0067]

[0068] wherein, is the number of compressed channels, C is the number of channels before compression, is the dimension reduction weight matrix.

[0069] The compressed features are activated using a ReLU activation function, and then the number of channels is restored to C to obtain the restored features :

[0070]

[0071] wherein, is the restored weight matrix.

[0072] The restored features are normalized using a Sigmoid activation function to obtain the attention weight :

[0073]

[0074] wherein, is a Sigmoid activation function, , z is a generic term for parameters participating in mathematical operations.

[0075] The attention weight and the nasopharyngeal laryngoscope image are multiplied channel by channel to obtain the channel global feature :

[0076]

[0077] Through the SE module, the global pooling operation is used to extract the global feature of the channel, effectively making up for the deficiency of the convolution and SS2D (two-dimensional selective scanning) module in the original model in capturing global dependence. This module can dynamically adjust the channel weight, highlight key features, and suppress irrelevant information, significantly improving the expression ability of the model.

[0078] Further, the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image, including:

[0079] The second pooling feature of the nasopharyngeal laryngoscope image is extracted using average pooling ;

[0080] The local dependence weight is obtained by using one-dimensional convolution to simulate the local dependence relationship between channels in the second pooling feature:

[0081]

[0082] wherein, Kis the size of the convolution kernel, and its value is determined by the number of channels C determined dynamically, denotes rounding down, is the convolution weight.

[0083] The local dependency weight and the nasopharyngolaryngoscope image are multiplied channel by channel to obtain fine-grained dependency features :

[0084]

[0085] The ECA module uses local one-dimensional convolution to accurately capture the fine-grained dependency between channels, effectively improving the feature expression level of the channel dimension. The one-dimensional convolution design of the ECA module highlights the local channel interaction characteristics, which complements the modeling of spatial dependency by the SS2D module.

[0086] After obtaining the channel global features and fine-grained dependency features, the two are added to obtain the channel features .

[0087] The LCA module compensates for the shortcomings of existing models in modeling channel dependency, and improves the modeling capability of multi-scale features and high-resolution details.

[0088] In S130, the channel features are input into a local feature extraction module, which includes a parallel standard convolution path and a multi-scale receptive field path. The standard convolution path uses a 3x3 depth separable convolution to extract small receptive field local features of the channel features. The multi-scale receptive field path uses a 5x5 depth separable convolution to extract large receptive field local features of the channel features. The small receptive field local features and the large receptive field local features are added to obtain fusion features. The fusion features are subjected to 1x1 point convolution for channel transformation and compression to obtain channel fusion features.

[0089] By way of example, to address the problem of low accuracy of the model when processing highly similar categories, the present embodiment proposes a local feature extraction module (Local Feature Extraction, LFE). This module uses multi-path convolution processing and channel fusion strategies to effectively extract local information from input features, achieving multi-scale feature fusion and allowing the model to capture subtle differences between similar categories, thereby improving classification accuracy.

[0090] Specifically, the small receptive field local features are represented as:

[0091]

[0092] wherein, is the small receptive field local feature, is the channel feature, is the depth convolution weight, denotes a convolution operation, and a convolution kernel with a size of 3x3 is used, is a dimension and an output channel number is a learnable bias vector with the same length, BN 1 denotes a batch normalization process, is a ReLU activation function process, .

[0093] The convolution operation described above combines weight sharing and sparse connection, not only efficiently extracts the key details such as edges and textures of input features, but also significantly improves the calculation efficiency. At the same time, the features output by the standard convolution path serve as the basis for multi-scale feature fusion, providing fine-grained feature support for subsequent channel fusion and larger receptive field paths.

[0094] Further, the large receptive field local feature is represented as:

[0095]

[0096] wherein, is a large receptive field local feature, is a channel feature, is a multi-scale convolution kernel, denotes a convolution operation, and a convolution kernel with a size of 5x5 is used, is a dimension and an output channel number is a learnable bias vector with the same length, BN 2 denotes a batch normalization process, is a ReLU activation function process.

[0097] Further, the fused feature is represented as:

[0098]

[0099] The channel fusion feature is represented as:

[0100]

[0101] wherein, is a channel fusion feature, is a dot convolution weight, is a dimension and an output channel number is a learnable bias vector with the same length, denotes a batch normalization process.

[0102] The features extracted by the standard convolution path and the multi-scale receptive field path may cover different spatial patterns or context information, so the channel fusion fuses the features from different paths at the channel dimension level by means of a 1*1 point convolution, integrates the information contained in each path, and generates a unified feature representation. Meanwhile, the 1*1 point convolution is used to achieve channel dimension reduction, so as to convert the high-dimensional features generated by the multi-scale path into a lower dimension, thereby reducing the computational complexity.

[0103] S140, input the channel fusion features into a classifier to obtain the recognition results of each part in the nasopharyngolaryngoscope image.

[0104] Exemplarily, the classifier can adopt a fully connected layer.

[0105] Further, the lightweight channel feature extraction module and the local feature extraction module adopt a hybrid loss function in the training process, and the hybrid loss function includes a cross-entropy loss, a label smoothing loss and a focal loss.

[0106] In the embodiments of the present application, the cross-entropy loss, the label smoothing loss and the focal loss are combined to realize multi-objective adjustment of model optimization. The hybrid loss function can effectively cope with the class imbalance problem and can also reduce the influence of inter-class similarity on the model.

[0107] The label smoothing loss is a kind of regularization technique, which functions to alleviate the overconfident condition of the model for the training data. By adjusting the target label from a completely determined one-hot distribution to a more smooth probability distribution, the label smoothing loss can improve the generalization ability of the model in the case of poor data quality. Label smoothing is a smoothing process for the distribution of the true label, and the specific formula is as follows:

[0108]

[0109] wherein, is the smoothed target distribution, is a label smoothing factor for controlling the smoothing degree of the target distribution, is the number of classes, is a Kronecker delta function, which takes a value of 1 when , otherwise 0. The calculation formula of the label smoothing loss is as follows:

[0110]

[0111] wherein, is the number of samples, is the prediction probability of the model for the class .

[0112] The focal loss is a loss function designed specifically for the problem of class imbalance, the main goal is to reduce the contribution of easy-to-classify samples to the total loss, so that the model focuses more on learning difficult-to-classify samples. Its formula is:

[0113]

[0114] wherein, is a balance factor, through different weights can be given to samples of different classes, thus effectively dealing with the problem of class imbalance, is a focal adjustment factor, used to control the weight of difficult-to-classify samples, through the loss weight of difficult-to-classify samples can be made higher, so as to reduce the influence of easy-to-classify samples on the loss.

[0115] The hybrid loss function proposed in the embodiments of the present application is a combination of the weights of the above three losses, and its expression is:

[0116]

[0117] wherein, is a weight coefficient, respectively representing the degree of attention to different loss terms. In a class-balanced data set, cross-entropy loss and label smoothing loss are given higher weights to obtain better training results. In a class-unbalanced data set, the weights of label smoothing loss and focal loss are increased, so as to effectively alleviate the adverse effects of data class imbalance on model performance.

[0118] The embodiments of the present application also provide a multi-part identification system for nasopharyngolaryngoscope images, comprising:

[0119] An image acquisition module is configured to acquire nasopharyngolaryngoscope images.

[0120] A lightweight channel feature extraction module includes a parallel compression activation module and an efficient channel attention module. The compression activation module is configured to perform global average pooling and dimension reduction compression on the nasopharyngolaryngoscope images to extract channel global features of the nasopharyngolaryngoscope images. The efficient channel attention module is configured to perform average pooling and one-dimensional convolution on the nasopharyngolaryngoscope images to extract fine-grained dependency features between channels of the nasopharyngolaryngoscope images. The channel global features and the fine-grained dependency features are added to obtain channel features.

[0121] The local feature extraction module includes a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path is used to extract small receptive field local features of channel features by using 3*3 deep separable convolution; the multi-scale receptive field path is used to extract large receptive field local features of channel features by using 5*5 deep separable convolution; the small receptive field local features and the large receptive field local features are added to obtain fusion features, and the fusion features are subjected to 1*1 point convolution for channel transformation and compression to obtain channel fusion features;

[0122] The classifier is used for classifying to obtain recognition results of each part in the nasopharyngolaryngoscope image.

[0123] The mixed loss function includes a cross-entropy loss, a label smoothing loss and a focal loss.

[0124] The application also provides a computer storage medium, which stores a plurality of computer instructions for enabling a computer to execute the method.

[0125] Experimental verification

[0126] The application evaluates the role of the multi-part classification and recognition task on the nasopharyngolaryngoscope self-built balanced dataset NIDB and unbalanced dataset NIDI dataset (real nasopharyngolaryngoscope dataset). The NIDB dataset is a balanced dataset, which has a total of 16000 images, including 20 key parts of the nasopharynx and larynx such as the left nasal cavity, the right nasal cavity, the left nasopharynx and the right nasopharynx, and each part has 800 images, and the data is uniformly distributed. The NIDI dataset has a class imbalance condition, and a total of 20291 images are recorded, and the image proportion of the common parts in the nasopharyngeal examination is high, and the image of some parts is less, such as 7167 throat panoramic images and 224 left tonsil images, which presents an uneven data distribution phenomenon.

[0127] Figures 2 to 4 It can be seen that the multi-part recognition method of the nasopharyngolaryngoscope image of the application achieves an F1-Score evaluation index of 95.337 and a Parameter evaluation index of 24.536M on the balanced dataset NIDB; Figures 5 to 7 It can be seen that the F1-Score evaluation index of the unbalanced dataset NIDI reaches 92.425, which shows high comprehensive performance, and is better than other existing methods. The nasopharyngolaryngoscope part recognition method of the application is suitable for application in actual diagnosis scenes, and provides a new idea for the intelligent auxiliary development of clinical decision-making.

[0128] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, it is intended that such additions and modifications be included within the scope of the application. It is the following claims, including any amendments thereto, which define the scope of the application.

[0129] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A multi-part identification method for nasopharyngolaryngoscope images, characterized in that, The method comprises the steps of: obtaining a nasopharyngolaryngoscope image; establishing a nasopharyngolaryngoscope multi-site classification model, which comprises four stages of feature extraction; inputting the nasopharyngolaryngoscope image into a lightweight channel feature extraction module, which comprises a parallel compression activation module and an efficient channel attention module; the compression activation module performs global average pooling and dimension reduction compression on the nasopharyngolaryngoscope image to extract channel global features of the nasopharyngolaryngoscope image; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngolaryngoscope image to extract fine-grained dependency features between channels of the nasopharyngolaryngoscope image; the channel global features and the fine-grained dependency features are added to obtain channel features; inputting the channel features into a local feature extraction module, which comprises a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path adopts 3×3 depth separable convolution to extract small receptive field local features of the channel features; the multi-scale receptive field path adopts 5×5 depth separable convolution to extract large receptive field local features of the channel features; the small receptive field local features and the large receptive field local features are added to obtain fusion features, and the fusion features are subjected to 1×1 point convolution for channel transformation and compression to obtain channel fusion features; inputting the channel fusion features into a classifier to obtain recognition results of each part in the nasopharyngolaryngoscope image; wherein the lightweight channel feature extraction module and the local feature extraction module use a hybrid loss function in the training process, and the hybrid loss function comprises a cross-entropy loss, a label smoothing loss and a focal loss; wherein the compression activation module performs global average pooling and dimension reduction compression on the nasopharyngolaryngoscope image, comprising: using global average pooling to compress the spatial dimension of the nasopharyngolaryngoscope image, only retaining the features of the channel dimension to obtain first pooling features; using 1×1 convolution to compress the number of channels of the first pooling features to obtain compressed features; using a ReLU activation function to activate the compressed features to obtain restored features; using a Sigmoid activation function to normalize the restored features to obtain attention weights; multiplying the attention weights and the nasopharyngolaryngoscope image channel by channel to obtain the channel global features; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngolaryngoscope image, comprising: using average pooling to extract second pooling features of the nasopharyngolaryngoscope image; using one-dimensional convolution to simulate the local dependency relationship between channels in the second pooling features to obtain local dependency weights; multiplying the local dependency weights and the nasopharyngolaryngoscope image channel by channel to obtain the fine-grained dependency features.

2. The multi-part identification method of a nasopharyngolaryngoscope image according to claim 1, characterized in that, The small receptive field local features are represented as: wherein, is the local feature of the small receptive field, is the channel feature, is the deep convolution weight, represents a convolution operation with a convolution kernel size of 3x3, is the dimension and the number of output channels is a learnable bias vector of the same length, BN 1 represents a batch normalization process, is a ReLU activation function process.

3. The method of claim 2, wherein the method further comprises: determining a plurality of positions of the endoscope in the nasal cavity based on the plurality of images. The large receptive field local features are represented as: wherein, is the large receptive field local feature, is the channel feature, is the multi-scale convolution kernel, represents a convolution operation with a convolution kernel size of 5x5, is the dimension and the number of output channels is the equal-length learnable bias vector, BN 2 represents the batch normalization processing, is the ReLU activation function processing.

4. The multi-part identification method of a nasopharyngolaryngoscope image according to claim 3, characterized in that, The fusion features is represented as: The channel fusion features are represented as: wherein, is the channel fusion feature, is the point convolution weight, is the dimension and the number of output channels is a learnable bias vector of the same length, denotes a batch normalization.

5. The method of claim 1, wherein the method further comprises: determining a plurality of regions of interest in the image based on the plurality of anatomical landmarks. After obtaining the nasopharyngolaryngoscope image, preprocessing is performed, which comprises removing black border areas, rotating and adjusting the size.

6. A system for multi-part identification of a nasopharyngolaryngoscope image according to any one of claims 1-5, characterized in that, ​ An image acquisition module is configured to acquire a nasopharyngolaryngoscope image; A model establishment module is configured to establish a classification model for multiple parts of the nasopharyngolaryngoscope image, wherein the classification model comprises four stages of feature extraction; A lightweight channel feature extraction module comprises a parallel compression activation module and an efficient channel attention module; the compression activation module is configured to perform global average pooling and dimension reduction compression on the nasopharyngolaryngoscope image to extract channel global features of the nasopharyngolaryngoscope image; The efficient channel attention module is configured to perform average pooling and one-dimensional convolution on the nasopharyngolaryngoscope image to extract fine-grained dependency features between channels of the nasopharyngolaryngoscope image; The channel global features and the fine-grained dependency features are added to obtain channel features; The compression activation module performs global average pooling and dimension reduction compression on the nasopharyngolaryngoscope image, including: using global average pooling to compress the spatial dimension of the nasopharyngolaryngoscope image, retaining only the features of the channel dimension to obtain first pooling features; using 1×1 convolution to compress the number of channels of the first pooling features to obtain compressed features; using a ReLU activation function to activate the compressed features to obtain restored features; using a Sigmoid activation function to normalize the restored features to obtain attention weights; multiplying the attention weights and the nasopharyngolaryngoscope image channel by channel to obtain the channel global features; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngolaryngoscope image, including: using average pooling to extract second pooling features of the nasopharyngolaryngoscope image; using one-dimensional convolution to simulate the local dependency relationship between channels in the second pooling features to obtain local dependency weights; multiplying the local dependency weights and the nasopharyngolaryngoscope image channel by channel to obtain the fine-grained dependency features; A local feature extraction module comprises a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path is configured to extract small receptive field local features of the channel features using 3×3 depth separable convolution; the multi-scale receptive field path is configured to extract large receptive field local features of the channel features using 5×5 depth separable convolution; the small receptive field local features and the large receptive field local features are added to obtain fusion features, and the fusion features are subjected to channel transformation and compression through 1×1 point convolution to obtain channel fusion features; A classifier is configured to classify to obtain recognition results of each part in the nasopharyngolaryngoscope image. In the training process, the lightweight channel feature extraction module and the local feature extraction module use a hybrid loss function, and the hybrid loss function comprises a cross-entropy loss, a label smoothing loss and a focal loss.

7. A computer storage medium, characterized in that The computer storage medium stores a plurality of computer instructions for causing a computer to execute the method of any one of claims 1-5.

Images