Throat reflux disease detection method and system based on laryngoscope image and RFS-Net algorithm

Abstract

The invention relates to a throat reflux disease detection method and system based on a laryngoscope image and an RFS-Net algorithm, and the method comprises the steps: collecting the laryngoscope image, screening a disease type by a professional worker according to the laryngoscope image, and carrying out the RFS evaluation of the collected laryngoscope image, and obtaining a first image; the first image is sent to an FCN-ResNet network for pixel-level region prediction, and a second image is obtained; continuously sending the second image into a DSE-ResNet network for feature quantization to obtain a quantization score of the first image; sending the obtained quantized score of the first image into an RFS-Net model, and performing LPR prediction to output a score; and the throat reflux disease is evaluated according to the score, so that the accuracy of diagnosing the throat reflux disease by machine recognition is improved, and automatic evaluation is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a laryngopharyngeal reflux disease detection method and system based on laryngoscope images and RFS-Net algorithm. Background Art

[0002] Laryngoscope image analysis is an important medical technology used to diagnose laryngopharyngeal reflux disease. However, traditional image analysis methods often rely on manual annotation and subjective scoring, which limits the accuracy and consistency of diagnostic results. In order to overcome these shortcomings, deep learning technology has been gradually introduced into the field of image processing in recent years. The RFS-Net algorithm is specifically designed to deal with the problem of class imbalance in images.

[0003] Based on this background, the present invention proposes a laryngopharyngeal reflux disease detection method and system based on laryngoscope images and RFS-Net algorithm. Summary of the invention

[0004] In order to solve the problem that existing image analysis methods for diagnosing laryngopharyngeal reflux disease often rely on manual annotation and subjective scoring, resulting in limited accuracy and consistency of diagnostic results, the present invention provides a laryngopharyngeal reflux disease detection method and system based on laryngoscope images and RFS-Net algorithm, which realizes accurate and automated evaluation of laryngopharyngeal reflux disease.

[0005] The technical solution of the present invention is as follows:

[0006] A laryngopharyngeal reflux disease detection method based on laryngoscope images and RFS-Net algorithm includes an RFS-Net model and an FCN-ResNet network, wherein the FCN-ResNet network uses ResNet101 and a fully connected convolutional neural network FCN as a backbone, and the method is as follows:

[0007] Step S1, collecting laryngoscope images, professional staff screen disease types according to the laryngoscope images, and perform RFS evaluation on the collected laryngoscope images to obtain a first image;

[0008] Step S2, sending the first image to the FCN-ResNet network for pixel-level region prediction to obtain a second image;

[0009] Step S3, sending the second image to the DSE-ResNet network for feature quantization to obtain a quantization score of the first image;

[0010] Step S4: Send the obtained quantization score of the first image into the RFS-Net model to perform LPR prediction and output the score;

[0011] Step S5: Evaluate laryngopharyngeal reflux disease according to the score.

[0012] Furthermore, step S1 specifically includes: obtaining the collected laryngoscope images, and screening the laryngoscope images that do not meet the requirements;

[0013] The screened laryngoscope images are subjected to RFS scoring by professional staff, which includes the scoring of 8 marker items, and the LPR marker corresponding to the laryngoscope image is found to obtain the first image, and the first image is preprocessed.

[0014] Furthermore, the preprocessing is to adjust the pixels of the laryngoscope image to a preset pixel size.

[0015] Furthermore, step S2 specifically includes: sending the first image into a ResNet network for feature extraction, and then sending the first image into an FCN network for region extraction through semantic segmentation, and performing pixel-level segmentation;

[0016] An adaptive pixel class balanced cross entropy loss function is used in the FCN-ResNet network to perform model training to obtain the second image.

[0017] Furthermore, the adaptive pixel class balanced cross entropy loss function formula is as follows:

[0018] Where α is a constant parameter, P represents the total number of pixels in the batch; m y Represents the number of pixels of category y in a batch; y represents the disease category; It means that the predicted scores of the calculated pixel for all categories j are exponentially calculated and summed; Indicates the predicted probability of calculating category y; K y represents a factor related to category y, specifically the importance or frequency of the disease category; C represents the total number of predicted disease categories; It represents the t-th power of the probability that the calculated pixel prediction category is y.

[0019] Furthermore, step S3 is specifically: performing channel compression on the second image to compress the information into a real vector, that is, z=(z1, z2, ..., z c ), compressed using global average pooling, which is calculated as follows:

[0020]

[0021] where u c (i, j) represents the eigenvalue in feature channel c, H and W represent the length and width of the image respectively; z c With F sq (u c ) is the calculation result, which is used to represent the global average pooling compression;

[0022] The activation function operation is performed on the compressed vector, which can be implemented through two fully connected layers, and the calculation method is as follows:

[0023] s=F cx (z,w)=σ(g(z,w))=σ(w2δ(w1z))

[0024] Where σ is the s-type activation function, δ is the rectified linear unit (ReLU) activation function, w1, w2 are the weight matrices after dimensionality reduction of the second image, and the quantized score of the first image is obtained by re-weighted calculation.

[0025] Furthermore, step 4 is specifically as follows: for the first image, training the parameters of LR on the augmented data of the training data set by adding random Gaussian noise;

[0026] For the quantized score of the first image, the corresponding RFS score vector x i =(S1,S2,...,S8), the final RFS score is expected to be the regression of g(·), expressed as: g(x i )=w0+w1S1+w2S2+...+w8S8

[0027] Wherein, w = (w1, w2, ..., w8) is the coefficient vector, w0 is the intercept, and the RFS score predicted by the model is obtained through the sensitivity formula and the specificity formula. S1 to S8 are the vectors of quantitative scores for the eight disease categories output by the quantitative score network.

[0028] Further, for the output of LPR prediction, the result for the correctly classified LPR positive image is set as the true positive TP, and the result for the incorrectly classified LPR positive image is set as the false positive FP, and the correct prediction and incorrect prediction of LPR negative prediction will be the true negative TN and false negative FN respectively;

[0029] The sensitivity formula is as follows:

[0030] Sensitivity=TP / (TP+FN)

[0031] The specificity formula is as follows:

[0032] Specificity = TN / (TN + FP)

[0033] A deep learning-based laryngopharyngeal reflux disease diagnosis system includes a data acquisition module, a pixel-level segmentation module, a feature quantification module, a linear prediction module, and an automated evaluation module. The data acquisition module is used to collect laryngoscope images. Professional staff screen disease types based on the laryngoscope images and perform RFS evaluation on the collected laryngoscope images to obtain a first image.

[0034] The pixel-level segmentation module is used to send the first image into the FCN-ResNet network for pixel-level segmentation to obtain a second image;

[0035] The feature quantization module is used to continue to send the second image to the DSE-ResNet network for feature quantization to obtain a quantization score of the first image;

[0036] The linear prediction module is used to send the obtained quantized score of the first image into the RFS-Net model to perform LPR prediction and output the score;

[0037] The automated assessment module is used to assess laryngopharyngeal reflux disease according to the scores.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) The design of the present invention overcomes the problem that traditional image processing methods cannot automatically handle category imbalance, and enhances the prediction accuracy and reliability of the system.

[0040] (2) In the embodiments of the present invention, the parameters of the processing algorithm can be automatically adjusted according to the different features of the laryngoscope image to achieve targeted optimization.

[0041] (3) The present invention uses a deep learning model to automatically process image data. Through pixel-level prediction and feature quantification, it can improve the accuracy and consistency of diagnosis when processing a large number of images. It is combined with the RFS-Net algorithm to address the problem of class imbalance in images, thereby achieving accurate and automated assessment of laryngopharyngeal reflux disease, breaking through the limitations of traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The overall flow chart of fully automatic quantification of laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm provided in an embodiment of the present invention;

[0043] Figure 2 It is a flow chart of the internal details of the channel attention technology network (DSE-ResNet) provided in an embodiment of the present invention;

[0044] Figure 3This is a comparison chart of the performance measurement indicators of four machine learning models for fully automatic quantification of laryngopharyngeal reflux disease among physicians, interns and RFS-Net in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose and technical solution of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] See also Figure 1 A laryngopharyngeal reflux disease detection method based on laryngoscope images and RFS-Net algorithm includes an RFS-Net model and an FCN-ResNet network. The FCN-ResNet network is based on

[0048] ResNet101 and fully connected convolutional neural network FCN are the backbone, and the method is:

[0049] Step S1, collecting laryngoscope images, professional staff screen disease types according to the laryngoscope images, and perform RFS evaluation on the collected laryngoscope images to obtain a first image;

[0050] Step S2, sending the first image to the FCN-ResNet network for pixel-level region prediction to obtain a second image;

[0051] Step S3, sending the second image to the DSE-ResNet network for feature quantization to obtain a quantization score of the first image;

[0052] Step S4: Send the obtained quantization score of the first image into the RFS-Net model to perform LPR prediction and output the score;

[0053] Step S5: Evaluate laryngopharyngeal reflux disease according to the score.

[0054] Among them, RFS assessment is the reflux finding score, which is mainly based on 24-hour dual-channel pH probe monitoring and subjective assessment by doctors based on a clinical scale.

[0055] After obtaining the LPR prediction score, the final score is judged using linear weighting based on the scores of each subcategory.

[0056] The present invention will be further described below in conjunction with a specific embodiment:

[0057] Step S1: Acquire the collected laryngoscope images and screen the laryngoscope images that do not meet the requirements; Have professional staff perform RFS scoring on the screened laryngoscope images, which includes the scoring of 8 marker items, and find the LPR marker corresponding to the laryngoscope image to obtain the first image, and pre-process the first image.

[0058] Among them, 8 signs are important for distinguishing reflux disease, which are: false vocal groove; disappearance of laryngeal ventricle; erythema / hemorrhage; vocal cord edema; diffuse edema; posterior symphysis hyperplasia; granuloma; and thick mucus adhesion in the larynx.

[0059] The LPR mark here is a flag that defines whether the image is diagnosed as laryngopharyngeal reflux disease after the doctor makes a preliminary judgment on the captured image; in the model prediction process, this item is the true label value.

[0060] The following is an example:

[0061] A large number of laryngoscope images from a wide range of populations were collected, including 524 subjects aged between 4 and 60 years old. Images with vocal cord leukoplakia, laryngeal cancer, hypopharyngeal cancer and other diseases were excluded. The screened images were manually scored with RFS based on personal experience, and the scores included ventricular occlusion, erythema / congestion, vocal cord edema, diffuse laryngeal edema and posterior commissure hypertrophy. The data set with good scores was divided into training data set, validation data set and test data set.

[0062] The laryngoscope dataset is divided into a training dataset, a validation dataset, and a test dataset.

[0063] There are 1661 samples in the training dataset, including 1181 images diagnosed as LPR, 412 images that are not LPR, and 68 unclassified data due to missing RFS assessment items. Three chief physicians marked the laryngoscopy area and severity score according to the RFS score.

[0064] The validation dataset was collected together with the RFS evaluation results from 6 doctors, including two interns, two attending physicians, and two chief physicians. The consistency of the evaluation results between the doctors and the AI ​​model was evaluated. The total number of subjects was 247, ranging in age from 4 to 60 years old. Sample laryngoscopy images and corresponding RFS scores were provided for each test subject.

[0065] The test dataset involved 93 test subjects, including 50 males and 43 females. The 24h MII-pH gold standard was used to replace the ground truth of the LPR project, replacing the manual RFS evaluation to improve the reliability of label scoring. The performance comparison of 6 doctors and the RFS-Net model was achieved. All of them provided two image resolutions, namely 720×576 and 1980×1080 pixels.

[0066] Before entering step S2, the raw data is preprocessed; first, all image data inputs for training or inference should be resized to a shape of 244×244 pixels. Second, flipped data augmentation is also randomly applied to image training.

[0067] Step S2: sending the first image to the ResNet network for feature extraction, and then sending it to the FCN network for region extraction through semantic segmentation and pixel-level segmentation;

[0068] An adaptive pixel class balanced cross entropy loss function is used in the FCN-ResNet network to perform model training to obtain the second image.

[0069] Among them, the ResNet network is responsible for extracting deep features and providing rich contextual information for the FCN network;

[0070] The FCN network uses these features to perform pixel-level segmentation, classifying each pixel in the image into different semantic categories, achieving region extraction and detailed segmentation.

[0071] Step S21: The test data set obtained in step S1 is fed into a ResNet101 and fully connected convolutional neural network (FCN) as the backbone, and semantic segmentation is applied to extract regions. Since the performance of FCN-ResNet is unstable when processing unbalanced medical data sets, pixel-level segmentation is performed. Therefore, the training performance of the model is improved by the reweighting method.

[0072] Step S22: Since the laryngoscope dataset has class imbalance, which is common in medical datasets, we designed an adaptive pixel class balanced cross entropy loss function (APCB) to perform reweighting. Based on the CB theory, we further improved the loss weighting strategy of the laryngoscope dataset and used the adaptive pixel class balanced loss function for segmentation model training. The formula is:

[0073]

[0074] Where α is a constant parameter, P represents the total number of pixels in the batch; m y Represents the number of pixels of category y in a batch; y represents the disease category; It means that the predicted scores of the calculated pixel for all categories j are exponentially calculated and summed; Indicates the predicted probability of calculating category y; K y represents a factor related to category y, specifically the importance or frequency of the disease category; C represents the total number of predicted disease categories; It represents the t-th power of the probability that the calculated pixel prediction category is y.

[0075] Therefore, the loss of model training is the sum of the losses of each category. y For the category y of the batch, the weight of the corresponding term in the loss function can be calculated as:

[0076]

[0077] K y It represents the weight of disease category y in a batch of images as a percentage of all the weights of the batch. Disease category y is one of the total number of predicted disease categories c. There are 8 types of c, and y is one of them.

[0078] where k y Can be defined as:

[0079]

[0080] Here m y represents the number of pixels of category y in the batch, β is a parameter less than 1 and greater than 0, k y It is to achieve the weight of pixel-level category balance. Combined with the above formula reasoning, the final loss function is the sum of the loss values ​​of each category. The formula is as follows:

[0081]

[0082] Where C is the total number of predicted disease categories, L(p,y) is the cross entropy loss of the above-mentioned single disease category, and finally the losses of each category are summed to get the final loss.

[0083] Empirically, the segmentation model training was optimized by stochastic gradient descent with momentum and weight decay, processed with a batch size of 10, and completed after 30 epochs, with the learning rate initialized from 0.01. The batch size of the feature quantization model was increased to 40, while the model training was initialized with a learning rate of 10e-6 and completed within 100 epochs. The learning rate was delayed by 0.5 every 10 epochs. In addition, early stopping techniques were used to avoid overfitting.

[0084] Step S23: For the verification of this module, two evaluation criteria are designed to accurately evaluate the segmentation performance, namely pixel accuracy (PA) and intersection over union (IoU). PA is expressed as the percentage of correctly classified pixels:

[0085]

[0086] where p ij is the pixel predicted to belong to class i and labeled as class j. Therefore, p ii are correctly classified pixels. In addition, the IoU measures the pixel value between the predicted and target masks divided by the pixel value of the pixels involved in both masks, as:

[0087]

[0088] p ji To predict pixels belonging to category j and labeled as category i;

[0089] In addition, the present invention also uses fine-grained PA i , for each segmentation category i, i Assessments in specific areas are:

[0090]

[0091] Step S3: sending the second image to the DSE-ResNet network for feature quantization to obtain a quantization score of the first image;

[0092] Step S31: According to the result of step S2, the DSE-ResNet network is input for further channel weighting. In the feature quantization stage, the LPR item score is given according to the segmented area. For the subjective items in the RFS scale, there are 5 subpaths through the feature quantization model, and the item scores S1-8 are obtained by the 8 individual training models of DSE-ResNet. Its structure is as follows Figure 2 As shown in Figure 2, the feature channels are effectively re-weighted.

[0093] For the input segmentation result feature map, global average pooling is performed to ensure the global receptive field, where its operation formula F sq As shown below:

[0094]

[0095] where z c is to extract the global information of the feature channel, compressed into a real vector, and u c (i, j) represents the eigenvalue in feature channel c, H and W represent the length and width of the image respectively; z c With F sq (uc ) is the calculation result, which is used to represent the global average pooling compression;

[0096] Step S32: by calculating the compressed vector value, the compressed value of each channel is weighted and then activated by the ReLU function, and the activation value is weighted again and activated by the sigmoid function. The formula is as follows:

[0097] s=F cx (z,w)=σ(g(z,w))=σ(w2δ(w1z))

[0098] Where σ is an s-type activation function, δ is a rectified linear unit (ReLU) activation function, w1 and w2 are weight matrices after dimensionality reduction of the second image, and w1 and w2 are weight values ​​of the first weighting and the second weighting, respectively.

[0099] The vector shape of the compressed vector is (B, C), where B is the batch size and C is the number of channels. Another meaning of the vector shape is that it represents the sub-item category.

[0100] Finally, the quantization score of the first image is obtained through reweighting calculation. The calculation process is: reweight the original image twice, sum the results, and dot product the summation result with the original image; for the value of each channel, dot product it with the input feature map, that is, apply the reweighted value to each pixel of each channel, and then prompt the evaluation accuracy of RFS. The formula is as follows:

[0101] a c =F scale (u c ,s c )=s c u c

[0102] u c is the feature map of category c of the input vector, s c For u c The reweighted vector;

[0103] Step S4: Send the obtained quantized score of the first image to LPR Prediction to obtain a discovery score.

[0104] The quantized input based on the feature quantization in step S3 is used to perform LPRD prediction and implement linear regression (LR) to obtain the final score. The parameters of LR are trained on the augmented data of the training dataset by adding random Gaussian noise. For each image, the corresponding RFS score vector x i =(S1,S2,...,S8), which can be expressed as:

[0105] g(x i )=w0+w1S1+w2S2+...+w8S8

[0106] where w i is the coefficient vector.

[0107] Wherein, w = (w1, w2, ..., w8) is the coefficient vector, w0 is the intercept, and the RFS score predicted by the model is obtained through the sensitivity formula and the specificity formula. S1 to S8 are the vectors of quantitative scores for the eight disease categories output by the quantitative score network.

[0108] For the output of LPR prediction, we can have a true positive (TP) for the correctly classified LPR positive image and a false positive (FP) for the incorrectly classified LPR positive image. Similarly, the correct prediction and incorrect prediction of LPR negative prediction will be true negative (TN) and false negative (FN) respectively. Therefore, the performance can be evaluated by sensitivity and specificity. The sensitivity formula is as follows:

[0109] Sensitivity=TP / (TP+FN)

[0110] The specificity formula is as follows:

[0111] Specificity = TN / (TN + FP)

[0112] In order to better understand the performance of the model, the F1 score is also introduced. The main way in which the F1 score affects the performance of the model is to evaluate the overall performance of the model through the comprehensive precision and recall rate. The F1 score finds a balance between precision and recall rate, avoiding the situation where one of the indicators is overemphasized. The F1 score is expressed as follows:

[0113] F1=2TP / (2TP+FN+FP)

[0114] In addition, the two-sided Mann-Whitney U test was used for consistency analysis, which is also non-parametric. Receiver operating characteristic (ROC) analysis and area under the curve (AUC) analysis were performed using confidence interval calculations. The DeLong's test was used to compare the ROC curves of the two models, with a p-value of 0.05. Figure 3 As shown, in this embodiment, the six medical staff are also ranked according to their experience in the test criteria for the results, starting from the most junior intern. In addition, after the chief physician, the RFS score predicted by the model is also added.

[0115] exist Figure 3 The symbols on the horizontal axis are explained as follows:

[0116] MA, intern a; MB, intern b; DA, attending physician a; DB, attending physician; SA, chief physician a; SB, chief physician b; RFS-Net model.

[0117] It is noteworthy that specific performance drops sharply at intermediate attending physicians, indicating that the correlation between specific performance and clinical experience is unstable. However, RFS-Net's performance is close to the average level of chief physicians and better than the most senior chief physicians with more than 20 years of clinical experience.

[0118] The two-sided Mann-Whitney U test is an existing technique for comparing whether the distributions of two independent samples are significantly different. It is a nonparametric test that does not require the data to conform to a normal distribution. By comparing the results or performance distributions of two models, their consistency or difference can be evaluated.

[0119] DeLong's test is an existing technology used to compare the AUC values ​​of two ROC curves to test whether the performance of the two models is significantly different, and the p value (0.05) is used to judge the significance of the results. If the p value is less than the significance level, it is considered that there is a significant difference in the AUC values ​​of the two models.

[0120] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm, characterized in that: It includes an RFS-Net model and an FCN-ResNet network, wherein the FCN-ResNet network uses ResNet101 and a fully connected convolutional neural network FCN as a backbone, and the method is: Step S1, collecting laryngoscope images, professional staff screen disease types according to the laryngoscope images, and perform RFS evaluation on the collected laryngoscope images to obtain a first image; Step S2, sending the first image to the FCN-ResNet network for pixel-level region prediction to obtain a second image; Step S3, sending the second image to the DSE-ResNet network for feature quantization to obtain a quantization score of the first image; Step S4: Send the obtained quantization score of the first image into the RFS-Net model to perform LPR prediction and output the score; Step S5: Evaluate laryngopharyngeal reflux disease according to the score.

2. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 1, characterized in that: Step S1 specifically includes: acquiring the collected laryngoscope images, and screening the laryngoscope images that do not meet the requirements; The screened laryngoscope images are subjected to RFS scoring by professional staff, which includes the scoring of 8 marker items, and the LPR marker corresponding to the laryngoscope image is found to obtain the first image, and the first image is preprocessed.

3. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 2, characterized in that: The preprocessing is to adjust the pixels of the laryngoscope image to a preset pixel size.

4. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 2, characterized in that: The eight markers are false vocal groove, effacement of laryngeal ventricle, erythema or hemorrhage, vocal cord edema, diffuse edema, posterior symphysis hyperplasia, granuloma and mucus adhesion in the larynx.

5. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 1, characterized in that: Step S2 specifically includes: sending the first image into a ResNet network for feature extraction, and then sending the first image into an FCN network for region extraction through semantic segmentation, and performing pixel-level segmentation; An adaptive pixel class balanced cross entropy loss function is used in the FCN-ResNet network to perform model training to obtain the second image.

6. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 5, characterized in that: The formula of the adaptive pixel class balanced cross entropy loss function is as follows: Where α is a constant parameter, P represents the total number of pixels in the batch; m y Represents the number of pixels of category y in a batch; y represents the disease category; It means that the predicted scores of the calculated pixel for all categories j are exponentially calculated and summed; Indicates the predicted probability of calculating category y; K y represents a factor related to category y, specifically the importance or frequency of the disease category; C represents the total number of predicted disease categories; It represents the t-th power of the probability that the calculated pixel prediction category is y.

7. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 1, characterized in that: Step S3 specifically includes: performing channel compression on the second image to compress the information into a real vector, that is, z=(z1, z2, ..., z c ), compressed using global average pooling, which is calculated as follows: where u c (i, j) represents the eigenvalue in feature channel c; H and W represent the length and width of the image respectively; z c With F sq (u c ) is the calculation result, which is used to represent the global average pooling compression; The activation function operation is performed on the compressed vector, which can be implemented through two fully connected layers, and the calculation method is as follows: s=F cx (z,w)=σ(g(z,w))=σ(w2δ(w1z)) Where σ is the s-type activation function, δ is the rectified linear unit (ReLU) activation function, w1, w2 are the weight matrices after dimensionality reduction of the second image, and the quantized score of the first image is obtained by re-weighted calculation, and w is the width of the image.

8. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 1, characterized in that: Step 4 is specifically as follows: for the first image, training the parameters of LR on the augmented data of the training data set by adding random Gaussian noise; For the quantized score of the first image, the corresponding RFS score vector x i =(S1,S2,...,S8), the final RFS score is expected to be the regression of g(·), expressed as: g(x i )=w0+w1S1+w2S2+...+w8S8 Wherein, w = (w1, w2, ..., w8) is the coefficient vector, w0 is the intercept, and the RFS score predicted by the model is obtained through the sensitivity formula and the specificity formula. S1 to S8 are the vectors of quantitative scores for the eight disease categories output by the quantitative score network.

9. The method for detecting laryngopharyngeal reflux disease based on laryngoscope images and RFS-Net algorithm according to claim 8, characterized in that: For the output of LPR prediction, the result for the correctly classified LPR positive image is set as the true positive TP, and the result for the incorrectly classified LPR positive image is set as the false positive FP. The correct prediction and incorrect prediction of LPR negative prediction will be the true negative TN and false negative FN respectively; The sensitivity formula is as follows: Sensitivity=TP / (TP+FN) The specificity formula is as follows: Specificity = TN / (TN + FP).

10. A deep learning-based laryngopharyngeal reflux disease diagnosis system, characterized in that: It includes a data acquisition module, a pixel-level segmentation module, a feature quantification module, a linear prediction module and an automatic evaluation module. The data acquisition module is used to collect laryngoscope images. Professional staff screen disease types according to the laryngoscope images and perform RFS evaluation on the collected laryngoscope images to obtain a first image. The pixel-level segmentation module is used to send the first image into the FCN-ResNet network for pixel-level segmentation to obtain a second image; The feature quantization module is used to continue to send the second image to the DSE-ResNet network for feature quantization to obtain a quantization score of the first image; The linear prediction module is used to send the obtained quantized score of the first image into the RFS-Net model to perform LPR prediction and output the score; The automated assessment module is used to assess laryngopharyngeal reflux disease according to the scores.

Images