Multi-part identification method and system for nose and throat endoscope image

Abstract

The invention discloses a multi-part identification method and system for a nose and throat endoscope image, and relates to the technical field of medical image processing, and the method comprises the steps: inputting the nose and throat endoscope image into a lightweight channel feature extraction module, and extracting channel global features and fine granularity dependence features; adding the channel global features and the fine granularity dependency features to obtain channel features; inputting the channel features into a local feature extraction module, and extracting 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 an identification result of each part. According to the application, through the lightweight channel feature extraction module, the local feature extraction module and the mixed loss function, the defect of low classification precision caused by insufficient channel dimension feature modeling, high similarity between nose and throat part classes and seriously unbalanced data classes is optimized.

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 larynx, for example, left and right nasal cavities and left and right nasopharyngeal parts. Without obvious anatomical landmarks, the feature extraction ability 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 an advantage in 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: obtaining a nasopharyngeal laryngoscope image; establishing a classification model for multiple parts of the nasopharyngeal laryngoscope, which includes four stages of feature extraction; inputting the nasopharyngeal laryngoscope image into a 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 dimension reduction compression on the nasopharyngeal laryngoscope image to extract the channel global feature of the nasopharyngeal laryngoscope image; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image to extract the fine-grained dependency feature between the channels of the nasopharyngeal laryngoscope image; and the channel global feature and the fine-grained dependency feature are added to obtain the channel feature; The channel feature is input into a local feature extraction module, 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; The channel fusion features are input into a classifier to obtain recognition results of each part in the nasopharyngeal laryngoscope image. 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.

[0006] In a possible implementation, the compression activation module performs global average pooling and dimension reduction compression on the nasopharyngeal laryngoscope image, including: The spatial dimension of the nasopharyngeal laryngoscope image is compressed by using global average pooling, only the features of the channel dimension are reserved, and first pooling features are obtained; The channel number of the first pooling features is compressed by using 1*1 convolution to obtain compressed features; The compressed features are activated by using a ReLU activation function to obtain restored features; The restored features are normalized by using a Sigmoid activation function to obtain attention weights; The attention weights and the nasopharyngeal laryngoscope image are multiplied channel by channel to obtain channel global features.

[0007] In a possible implementation, the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image, including: Second pooling features of the nasopharyngeal laryngoscope image are extracted by using average pooling; Local dependency weights are obtained by using one-dimensional convolution to simulate local dependency relationships between channels in the second pooling features; The local dependency weights and the nasopharyngeal laryngoscope image are multiplied channel by channel to obtain fine-grained dependency features.

[0008] In a possible implementation, the small receptive field local feature is represented as:

[0009] Wherein, is the small receptive field local feature, is the channel feature, is a 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.

[0010] In a possible implementation, the large receptive field local feature is represented as:

[0011] 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.

[0012] In a possible implementation, the fusion feature is represented as:

[0013] The channel fusion feature is represented as:

[0014] 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.

[0015] 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.

[0016] In another aspect, the embodiments of the present application further provide a multi-part identification system for nasopharyngolaryngoscope images, comprising: 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 a nasopharyngolaryngoscope, and the classification model includes four stages of feature extraction. The light-weight channel feature extraction module includes a parallel compression activation module and an efficient channel attention module; the compression activation module is used for global average pooling and dimension reduction compression on the nasopharyngeal laryngoscope image, so as to extract the channel global feature of the nasopharyngeal laryngoscope image; the efficient channel attention module is used for average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image, so as to extract the fine-grained dependency feature between the channels of the nasopharyngeal laryngoscope image; the channel global feature and the fine-grained dependency feature are added to obtain the channel feature. The local feature extraction module includes a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path is used for extracting small receptive field local features of the channel features by using 3*3 depth separable convolution; the multi-scale receptive field path is used for extracting large receptive field local features of the channel features by 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 1*1 point convolution for channel transformation and compression to obtain channel fusion features. The classifier is used for classifying to obtain the recognition result of each part in the nasopharyngeal laryngoscope image. 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 includes a cross-entropy loss, a label smoothing loss and a focal loss.

[0017] In another aspect, the embodiment of the present application also provides a computer storage medium, which stores a plurality of computer instructions for enabling a computer to execute the above-mentioned method.

[0018] The nasopharyngeal laryngoscope image multi-part recognition method and system provided in the present application have the following advantages: 1. Based on the Mamba structure, combined with the light-weight channel feature extraction (LCA), the local feature extraction (LFE) and the hybrid loss (CL) mechanism, the model is light-weighted, the ability to distinguish high-similarity parts is significantly improved, and the performance is better than that of the traditional CNN or Transformer model in complex endoscope image recognition.

[0019] 2. In view of the problems of uneven distribution of nasopharyngeal laryngeal part data and difficulty in feature extraction, a fine feature modeling and loss function design strategy is proposed, so that the model has higher recognition accuracy and stronger generalization ability on multiple data sets, and can better meet the needs of clinical real-time diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on these drawings also belong to the protection scope of the present application.

[0021] Figure 1 A flowchart of a multi-part identification method of a nasopharyngeal laryngoscope image provided by an embodiment of the present application.

[0022] Figure 2 A comparison chart of parameter quantity and F1 score on the data set NIDB of the identification method provided by the embodiment of the present application and multiple comparison methods.

[0023] Figure 3 A comparison chart of classification performance on the data set NIDB of the identification method provided by the embodiment of the present application and multiple comparison methods.

[0024] Figure 4 A confusion matrix result chart on the data set NIDB of the identification method provided by the embodiment of the present application and multiple comparison methods.

[0025] Figure 5 A comparison chart of parameter quantity and F1 score on the data set NIDI of the identification method provided by the embodiment of the present application and multiple comparison methods.

[0026] Figure 6 A comparison chart of classification performance on the data set NIDI of the identification method provided by the embodiment of the present application and multiple comparison methods.

[0027] Figure 7 A confusion matrix result chart on the data set NIDI of the identification method provided by the embodiment of the present application and multiple comparison methods. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0029] Figure 1 A flowchart of a multi-part identification method of a nasopharyngeal laryngoscope image provided by an embodiment of the present application. The embodiment of the present application provides a multi-part identification method of a nasopharyngeal laryngoscope image, which comprises: S100, acquiring a nasopharyngeal laryngoscope image.

[0030] Exemplarily, after acquiring the nasopharyngeal laryngoscope image, preprocessing is performed first, which includes removing the black border area, rotation and size adjustment.

[0031] Specifically, the original endoscope 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 present application embodiment, and the model automatically adjusts the size of all images to a uniform 224x224x3, which is more suitable for subsequent part recognition and classification tasks.

[0032] S110, a nasopharyngeal laryngoscope multi-part classification model is established, which includes four stages of feature extraction.

[0033] Exemplarily, the input of the classification model is a 224x224x3 nasopharyngeal laryngoscope image, which is divided into a tensor size of 56x56xC1 through a patch partition module, and then outputted without changing the tensor size through the first stage of L1 ENT-Mamba module feature extraction and transformation; then outputted as a feature tensor of 28x28xC2 through the second stage of downsampling and L2 ENT-Mamba module feature extraction and transformation; then outputted as a feature tensor of 14x14xC3 through the third stage of downsampling and L3 ENT-Mamba module feature extraction and transformation; and finally outputted as a feature tensor of 7x7xC4 through the fourth stage of downsampling and L4 ENT-Mamba modules, to generate global core semantic features for nasopharyngeal laryngoscope classification. The ENT-Mamba module includes a lightweight channel feature extraction module for improving the feature selection ability and representation ability 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.

[0034] S120, inputting the nasopharyngeal laryngoscope image 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 dimension reduction compression on the nasopharyngeal laryngoscope image to extract the channel global feature of the nasopharyngeal laryngoscope image; the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal laryngoscope image to extract the fine-grained dependency feature between channels of the nasopharyngeal laryngoscope image; and the channel global feature and the fine-grained dependency feature are added to obtain the channel feature.

[0035] Exemplarily, in view of the problems of insufficient Efficient V Mamba channel dimension modeling and feature fusion, an embodiment of the present application proposes a lightweight channel feature extraction module (Lightweight Channel Attention, LCA), which dynamically weights the key of each channel and can improve the feature selection ability and representation ability of the model in the channel dimension. The module consists of two parts: a compression activation module (Squeeze-and-Excitation, SE) and an efficient channel attention module (Efficient Channel Attention, ECA). Among them, SE uses global pooling and twice dimension reduction operation to strengthen global channel dependence modeling, and ECA efficiently captures the local interaction characteristics between channels through adaptive convolution kernel.

[0036] Specifically, the compression activation module performs global average pooling and dimension reduction compression on the nasopharyngeal laryngoscope image, including: 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 the first pooling feature :

[0037] Among them, is the feature value at the position b of the c channel of the h nasopharyngeal laryngoscope image, w and H are the height and width of the nasopharyngeal laryngoscope image, respectively. W

[0038] The channel number of the first pooling feature is compressed from C to using a 1x1 convolution to obtain the compressed feature :

[0039] Among them, is the compressed channel number, C is the channel number before compression, is the dimension reduction weight matrix.

[0040] The compressed feature is activated using a ReLU activation function, and then the channel number is restored to C to obtain the restored feature :

[0041] Among them, is the restoration weight matrix.​

[0042] The reduced features are normalized using a sigmoid activation function to obtain attention weights

[0043] wherein, is a sigmoid activation function, , z is a parameter participating in mathematical operation.

[0044] The attention weights are multiplied with the nasopharyngolaryngoscope image channel by channel to obtain channel global features

[0045] Through the SE module, the global pooling operation is used to extract the global features 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 the key features, and suppress irrelevant information, significantly improving the expression ability of the model.

[0046] Further, the efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngolaryngoscope image, including: The second pooling feature of the nasopharyngolaryngoscope image is extracted using average pooling The local dependence weight is obtained by using one-dimensional convolution to simulate the local dependence relationship between channels in the second pooling feature

[0047] wherein, K is the size of the convolution kernel, the value of which is determined dynamically by the number of channels C , represents rounding down, is the convolution weight.

[0048] The local dependence weight is multiplied with the nasopharyngolaryngoscope image channel by channel to obtain fine-grained dependence features

[0049] The ECA module uses local one-dimensional convolution to accurately capture the fine-grained dependence relationship between channels, effectively improving the feature expression level of the channel dimension. Its one-dimensional convolution design highlights the local channel interaction characteristics, complementing the modeling of spatial dependence by the SS2D module.

[0050] ​​​​​After obtaining the channel global feature and the fine-grained dependency feature, the channel feature can be obtained by adding the two features .

[0051] The LCA module compensates for the deficiency of the existing model in modeling channel dependency, and improves the modeling capability of multi-scale features and high-resolution details.

[0052] In S130, the channel feature is 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 feature. The multi-scale receptive field path uses a 5x5 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. The fusion features are subjected to 1x1 point convolution for channel transformation and compression to obtain channel fusion features.

[0053] Exemplarily, to solve the problem that the model has low accuracy when processing highly similar categories, the embodiment of the present application 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, achieve multi-scale feature fusion, and enable the model to capture subtle differences between similar categories, thereby improving classification accuracy.

[0054] Specifically, the small receptive field local feature is represented as:

[0055] wherein, is the small receptive field local feature, is the channel feature, is the depth convolution weight, represents a convolution operation using a convolution kernel with a size of 3x3, is a learnable bias vector with a dimension and an output channel number equal to each other, BN 1 represents batch normalization processing, is a ReLU activation function processing, .

[0056] The above convolution operation combines weight sharing and sparse connection, which not only efficiently extracts key details such as edges and textures of the input feature, 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.

[0057] Further, the large receptive field local feature is represented as:

[0058] 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 with a size of 5x5 is adopted, is a dimension and an output channel number is a learnable bias vector with the same length, BN 2 represents a batch normalization process, is a ReLU activation function process.

[0059] Further, the fused feature is represented as:

[0060] The channel fused feature is represented as:

[0061] wherein, is a channel fused feature, is a dot convolution weight, is a dimension and an output channel number is a learnable bias vector with the same length, represents a batch normalization process.

[0062] The features extracted by the standard convolution path and the multi-scale receptive field path may cover different spatial patterns or context information, and thus the channel fusion fuses the features from different paths at the channel dimension level by means of a 1x1 dot convolution, integrates the information contained in each path, and generates a unified feature representation. Meanwhile, the channel fusion achieves channel dimension reduction by means of the 1x1 dot convolution, converts the high-dimensional features generated by the multi-scale path into a lower dimension, so as to reduce the computational complexity.

[0063] S140, inputting the channel fused feature into a classifier to obtain a recognition result of each part in the nasopharyngolaryngoscope image.

[0064] Exemplarily, the classifier can adopt a fully connected layer.

[0065] 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.

[0066] 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, and 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.

[0067] Label smoothing loss is a regularization technique that aims to mitigate the model's overconfidence in the training data. By adjusting the target label from a completely deterministic single-point distribution (one-hot) to a smoother probability distribution, label smoothing loss can improve the model's generalization ability in the case of poor data quality. Label smoothing smoothes the distribution of the true label. The specific formula is as follows:

[0068] in, is the smoothed target distribution, is the label smoothing factor, which is used to control the smoothness of the target distribution. is the number of categories, is the Kronecker delta function, when The value is 1 when , otherwise it is 0. Label smoothing loss The calculation formula is:

[0069] in, is the sample size, Is the model for the category The predicted probability of .

[0070] focal loss It is a loss function designed specifically for class imbalance problems. Its main goal is to reduce the contribution of easy-to-classify samples to the total loss, allowing the model to focus more on learning difficult-to-classify samples. Its formula is:

[0071] in, is the balancing factor, Different weights can be given to samples of different categories, thereby effectively dealing with the problem of category imbalance. is the focus adjustment factor, which is used to control the weight of difficult-to-classify samples. This can make the loss weight of difficult-to-classify samples higher, thereby reducing the impact of easy-to-classify samples on the loss.

[0072] The hybrid loss function proposed in the embodiment of the present application is a weighted combination of the above three losses, and its expression is:

[0073] in, are weight coefficients, respectively, indicating the importance of different loss terms. In the class-balanced data set, the cross-entropy loss and the label smoothing loss are given higher weights to obtain better training results. In the class-unbalanced data set, the weights of the label smoothing loss and the focal loss are increased to effectively alleviate the adverse effects of data class imbalance on model performance.

[0074] The embodiment of the application further provides a multi-part identification system of a nasopharyngolaryngoscope image, comprising: An image acquisition module is configured to acquire a nasopharyngolaryngoscope image. The 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 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 a 3*3 depth separable convolution. The multi-scale receptive field path is configured to extract large receptive field local features of the channel features by using a 5*5 depth separable convolution. 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 1*1 point convolution for channel transformation and compression to obtain channel fusion features. A classifier is configured to classify to obtain identification results of each part in the nasopharyngolaryngoscope image. The lightweight channel feature extraction module and the local feature extraction module use a hybrid loss function in the training process. The hybrid loss function comprises a cross-entropy loss, a label smoothing loss and a focal loss.

[0075] The embodiment of the application further provides a computer storage medium, which stores a plurality of computer instructions for causing a computer to execute the method described above.

[0076] Experimental verification The application evaluates the role of multi-site classification and recognition tasks on the self-built balanced data set NIDB and unbalanced data set NIDI data set (real nasopharyngeal endoscopy data set) of nasopharyngeal endoscope. The NIDB data set is a balanced data set, a total of 16000 images, including 20 key sites of nasopharyngeal cavity such as left nasal cavity, right nasal cavity, left nasopharynx and right nasopharynx, each site has 800 images, and the data distribution is uniform. The NIDI data set has a class imbalance condition, a total of 20291 images, and the image proportion of common parts in nasopharyngeal examination is high, and the image of some parts is less, such as 7167 throat panoramic images and 224 left tonsil images, showing an uneven data distribution phenomenon.

[0077] Figures 2 to 4 It can be seen that the multi-site recognition method of nasopharyngeal endoscopy image of the application reaches 95.337 in F1-Score evaluation index of balanced data set NIDB, and Parameter evaluation index reaches 24.536M; Figures 5 to 7 It can be seen that in the unbalanced data set NIDI, the F1-Score evaluation index reaches 92.425, and all show high comprehensive performance, which is better than other existing methods. The nasopharyngeal endoscopy site recognition method of the application is suitable for application in actual diagnosis scene, and provides a new idea for intelligent auxiliary development of clinical decision.

[0078] Although the preferred embodiments of the application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all changes and modifications falling within the scope of the application.

[0079] Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims of the application and their equivalent technologies, the application also intends to include these modifications and variations.

Claims

1. A multi-site recognition method for nasopharyngeal endoscope images, characterized in that: include: Obtain nasopharyngeal endoscopic images; Establish a classification model for multiple parts of nasopharyngeal endoscopes, which includes four stages of feature extraction; Inputting the nasopharyngeal endoscope 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 dimensionality reduction compression on the nasopharyngeal endoscope image to extract the channel global features of the nasopharyngeal endoscope image; The efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal endoscope image to extract fine-grained dependency features between channels of the nasopharyngeal endoscope 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, the local feature extraction module includes a parallel standard convolution path and a multi-scale receptive field path; the standard convolution path uses 3×3 depthwise separable convolution to extract small receptive field local features of the channel features; the multi-scale receptive field path uses 5×5 depthwise 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 a fused feature, and the fused feature is subjected to channel transformation and compression through 1×1 point convolution to obtain a channel fused feature; Inputting the channel fusion features into a classifier to obtain recognition results of various parts in the nasopharyngeal endoscope image; The lightweight channel feature extraction module and the local feature extraction module adopt a hybrid loss function during training, and the hybrid loss function includes cross entropy loss, label smoothing loss and focal loss.

2. The multi-site recognition method of nasopharyngeal endoscope images according to claim 1, characterized in that: The compression activation module performs global average pooling and dimensionality reduction compression on the nasopharyngeal endoscope image, including: Using global average pooling to compress the spatial dimension of the nasopharyngeal endoscope image, retaining only the features of the channel dimension, to obtain the first pooling feature; Use 1×1 convolution to compress the number of channels of the first pooled feature to obtain a compressed feature; Use the ReLU activation function to activate the compressed features to obtain restored features; The restored features are normalized using a Sigmoid activation function to obtain an attention weight; The attention weight and the nasopharyngeal endoscope image are multiplied channel by channel to obtain the channel global feature.

3. The multi-site recognition method of nasopharyngeal endoscope images according to claim 1, characterized in that: The efficient channel attention module performs average pooling and one-dimensional convolution on the nasopharyngeal endoscope image, including: Extracting second pooled features of the nasopharyngeal endoscope image using average pooling; Using one-dimensional convolution to simulate the local dependency relationship between channels in the second pooling feature to obtain local dependency weights; The local dependency weight and the nasopharyngeal endoscope image are multiplied channel by channel to obtain the fine-grained dependency feature.

4. The multi-site recognition method for nasopharyngeal endoscope images according to claim 1, characterized in that: The local feature of the small receptive field is expressed as: in, is the local feature of the small receptive field, is the channel characteristic, is the depth convolution weight, Represents the convolution operation, using a convolution kernel size of 3×3. is the dimension and number of output channels learnable bias vectors of equal length, BN 1 means batch normalization processing, Processed by the ReLU activation function.

5. The multi-site recognition method of nasopharyngeal endoscope images according to claim 4, characterized in that: The large receptive field local feature is expressed as: in, is the large receptive field local feature, is the channel characteristic, is a multi-scale convolution kernel, Represents the convolution operation, using a convolution kernel size of 5×5. is the dimension and number of output channels learnable bias vectors of equal length, BN 2 represents batch normalization processing, Processed by the ReLU activation function.

6. The multi-site recognition method of nasopharyngeal endoscope images according to claim 5, characterized in that: The fusion features Expressed as: The channel fusion feature is expressed as: in, is the channel fusion feature, is the point convolution weight, is the dimension and number of output channels learnable bias vectors of equal length, Indicates batch normalization processing.

7. The multi-site recognition method for nasopharyngeal endoscope images according to claim 1, characterized in that: After the nasopharyngeal endoscope image is acquired, it is first pre-processed, and the pre-processing includes removing black edge areas, rotating and resizing.

8. A system using the multi-site recognition method for nasopharyngeal endoscope images according to any one of claims 1 to 7, characterized in that: include: An image acquisition module, used for acquiring nasopharyngeal endoscope images; A model building module is used to build a classification model for multiple parts of nasopharyngeal endoscopes, and the classification model includes four stages of feature extraction; A lightweight channel feature extraction module, comprising a parallel compression activation module and an efficient channel attention module; the compression activation module is used to perform global average pooling and dimensionality reduction compression on the nasopharyngeal endoscope image to extract the channel global features of the nasopharyngeal endoscope image; The efficient channel attention module is used to perform average pooling and one-dimensional convolution on the nasopharyngeal endoscope image to extract fine-grained dependency features between channels of the nasopharyngeal endoscope image; the channel global features and the fine-grained dependency features are added to obtain channel features; A 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 the small receptive field local features of the channel features using 3×3 depth-separable convolution; the multi-scale receptive field path is used to extract the 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 a fused feature, and the fused feature is subjected to channel transformation and compression through 1×1 point convolution to obtain a channel fused feature; A classifier, configured to classify and obtain recognition results of various parts in the nasopharyngeal endoscope image; The lightweight channel feature extraction module and the local feature extraction module adopt a hybrid loss function during training, and the hybrid loss function includes cross entropy loss, label smoothing loss and focal loss.

9. A computer storage medium, characterized in that The computer storage medium stores a plurality of computer instructions, and the plurality of computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 7.

Images