A Method and System for Detecting Laryngopharyngeal Reflux Disease Based on Laryngoscopic Images and RFS-Net Algorithm

Abstract

This invention provides a method and system for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm, belonging to the field of intelligent medical image detection technology. The method includes the following steps: acquiring laryngoscopy image samples, preprocessing them to obtain standard images; inputting the standard images into the RFS-Net model to extract spatial and spectral features; training the model based on an improved Focal Loss function combined with the extracted spatial and spectral features to balance the sample distribution; fusing the superficial texture features and deep semantic features of the image samples through a residual attention mechanism (RAM); and performing lesion analysis on patient data based on clinical indicators and the fused features to output detection results. This invention uses the RFS-Net model to extract and fuse multi-dimensional features from laryngoscopy images, covering both spatial and spectral features. Combined with the improved Focal Loss function to optimize model training, it can more accurately identify the characteristics of laryngopharyngeal reflux disease, reducing misdiagnosis and missed diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image detection technology, and in particular to a method and system for detecting laryngopharyngeal reflux disease based on laryngoscope images and the RFS-Net algorithm. Background Technology

[0002] Laryngopharyngeal reflux disease (LDR) is a common upper respiratory tract infection. Its early symptoms are often subtle and easily confused with other throat conditions, making accurate diagnosis crucial for treatment and prognosis. Currently, LDR diagnosis relies heavily on electronic laryngoscopy, which primarily depends on doctors visually interpreting laryngoscopic images. This method has significant drawbacks: firstly, doctors' diagnoses are heavily influenced by subjective experience, and different doctors may interpret the same laryngoscopic image differently, leading to a high rate of misdiagnosis and missed diagnosis; secondly, visual diagnosis is inefficient and cannot meet the needs of large-scale screening.

[0003] With the application of artificial intelligence technology in the medical field, deep learning-based laryngoscopy image analysis technology has gradually emerged. However, existing deep learning detection methods have many shortcomings. In terms of feature extraction, traditional models can often only extract single-level image features, resulting in insufficient expression of the characteristics of laryngopharyngeal reflux disease. Moreover, existing detection systems lack effective integration with clinical indicators, affecting the accuracy and reliability of the final diagnostic results. Therefore, this invention proposes a laryngopharyngeal reflux disease detection method and system based on laryngoscopy images and the RFS-Net algorithm to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a method and system for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm. This system extracts and fuses multi-dimensional features from laryngoscopy images using the RFS-Net model, encompassing both spatial and spectral features. Combined with an improved Focal Loss function to optimize model training, it can more accurately identify the characteristics of laryngopharyngeal reflux disease, reduce misdiagnosis and missed diagnosis, and achieve higher accuracy.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm, comprising the following steps:

[0006] S1: Obtain laryngoscope image samples, perform preprocessing, and obtain standard images;

[0007] S2: Input the standard image into the RFS-Net model to extract spatial and spectral features;

[0008] S3: Based on the improved Focal Loss function, the spatial and spectral features extracted in step S2 are combined to train the model and balance the sample distribution;

[0009] S4: The shallow texture features and deep semantic features of image samples are fused through the residual attention mechanism RAM to obtain fused features;

[0010] S5: Combine clinical indicators and the fusion features from step S4 to perform lesion analysis on patient data and output the test results.

[0011] Further improvements are made in the following aspects: In S1, the preprocessing includes denoising, color correction, image enhancement, and ROI extraction. The ROI extraction is based on the U-Net model to automatically locate the anatomical structure of the larynx. The preprocessing also includes using adaptive histogram equalization (CLAHE) to enhance the contrast of the laryngeal mucosa and extracting the vocal cord vibration frequency features through the Gabor filter bank.

[0012] A further improvement is that S2 includes the following steps:

[0013] An improved ResNet-50 network was used to extract local texture features from laryngoscope images using 3×3 convolutional kernels as spatial features, including mucosal blood vessel distribution and ulcer edges;

[0014] The wavelet transform algorithm is introduced to decompose the image into low-frequency approximate components and high-frequency detail components, capturing high-frequency pathological signals in the lesion area as spectral features, including edge features of abnormally proliferating tissue;

[0015] Spatial features and spectral features are stitched together channel by channel;

[0016] The cross-attention mechanism (CAM) dynamically adjusts the weights of features in the two branches.

[0017] A further improvement is made in S3, where the improved Focal Loss function is:

[0018] FL(pt) = -a(1-pt) γ log(pt)·e -λ·|pt-0.5|

[0019] Where γ is the adjustment factor; λ is the class balance coefficient; pt is the model predicted probability; e is the base of the natural logarithm, equal to 2.71828; a is the balance factor used to adjust the weights of positive and negative samples in the loss function; |pt-0.5| represents the absolute difference between the predicted probability pt and 0.5; when pt is close to 0.5, this difference is small, and the exponential term e -λ·|pt-0.5| A larger value of pt increases the loss value of the sample; when pt is far from 0.5, the difference is large and the exponential term is small, thus reducing the loss value of the sample; the classification loss weight is dynamically adjusted according to the difficulty of the sample, and the contribution of easy-to-classify samples is suppressed by combining spatial features and spectral features. A class weight parameter is introduced to balance the samples.

[0020] A further improvement is made in S4, where the residual attention mechanism RAM implements feature weighting using the following formula:

[0021]

[0022] Among them, F out and F in These represent output and input features, respectively, used to describe the flow of data; σ represents element-wise multiplication, used to apply attention weights to the input features; σ is the Sigmoid function, used to map the input to the range (0, 1) to generate attention weights; W1 and W2 are learnable weight matrices, used to transform features in the channel dimension and capture the relationship between different channels; δ is the ReLU function, used to introduce non-linearity to increase the model's expressive power; GlobalAvgPool is global average pooling, used to compress features in the spatial dimension into a scalar; b1 and b2 represent bias terms, used to adjust the model's fitting ability; the original feature information is preserved through residual connections, and different feature weights are assigned according to the importance of lesion features using the attention mechanism; channel weights are generated through global average pooling and 1×1 convolution, and a spatial weight map is generated using 3×3 depthwise separable convolution; the original features and attention-weighted features are added element-wise, and the shallow texture features and deep semantic features of the image samples are fused to obtain the fused features.

[0023] Further improvements are made in S5, where clinical indicators include the reflux symptom index (RSI), reflux symptom score (RFS), pH monitoring value, and pepsin level.

[0024] A further improvement is that S5 includes the following steps:

[0025] Input a patient's laryngoscopy image, and output the lesion probability value P1 by combining and comparing features through the RFS-Net model;

[0026] Collect the patient's reflux symptom index (RSI), reflux symptom score (RFS), pH monitoring value, and pepsin level;

[0027] RSI is obtained through a patient self-assessment questionnaire, while RFS is assessed by a physician based on eight signs, ultimately yielding clinical indicators.

[0028] The positive results of clinical indicators are converted into the lesion probability value P2;

[0029] Combine P1 and P2 to output the detection result.

[0030] A laryngopharyngeal reflux disease detection system based on laryngoscope images and RFS-Net algorithm includes a data acquisition module, a feature extraction module, a feature fusion module, a model training module, and a classification decision module.

[0031] The data acquisition module is used to collect and store laryngoscope image samples, and also to acquire patient laryngoscope images and clinical indicator data in real time.

[0032] The feature extraction module is based on the RFS-Net model and is used to extract spatial and spectral features of image samples.

[0033] The model training module uses an improved Focal Loss function combined with spatial and spectral features extracted from the feature extraction module for training, balancing the sample distribution;

[0034] The feature fusion module is used to fuse shallow texture features and deep semantic features of image samples through a residual attention mechanism (RAM) to obtain fused features.

[0035] The classification decision module is used to compare the patient's real-time laryngoscope image with fused features and output the detection results in combination with clinical indicators.

[0036] A further improvement is that the data acquisition module has a built-in preprocessing module, which is used to perform noise reduction, color correction, image enhancement, ROI extraction, enhance the contrast of the laryngeal mucosa, and extract the vibration frequency features of the vocal cords.

[0037] A further improvement is that the classification decision module also includes a multilayer perceptron (MLP) for classifying fused features and clinical indicators.

[0038] The beneficial effects of this invention are as follows:

[0039] 1. This invention uses the RFS-Net model to extract and fuse multi-dimensional features from laryngoscopy images, covering spatial and spectral features. Combined with an improved Focal Loss function to optimize model training, it can more accurately identify the characteristics of laryngopharyngeal reflux disease, reduce misdiagnosis and missed diagnosis, and achieve higher accuracy.

[0040] 2. This invention employs a residual attention mechanism (RAM) and an improved Focal Loss function, which makes the model more adaptable to laryngoscope images of different qualities and imbalanced datasets, thereby enhancing the robustness of the model.

[0041] 3. In the decision-making process, this invention combines fusion features with clinical indicators for diagnosis, combines the determined lesion probability values ​​P1 and P2, and outputs the detection results to provide doctors with more comprehensive information, which helps doctors make more scientific and reasonable clinical decisions and improve treatment effects. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a system composition diagram of the present invention. Detailed Implementation

[0044] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0045] Example 1

[0046] according to Figure 1 , 2 As shown, this embodiment proposes a method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net (Recurrent Feature Sharing Network, a deep learning model combining recurrent neural networks and feature sharing mechanisms, mainly used for processing sequence data or multi-task learning problems) algorithm, including the following steps:

[0047] S1: Acquire laryngoscope image samples, perform preprocessing, and obtain standard images. Preprocessing includes denoising, color correction, image enhancement, and ROI (Region of Interest) extraction. ROI extraction is based on the U-Net (a convolutional neural network with a symmetric encoder-decoder structure) model to automatically locate the anatomical structures of the larynx. Preprocessing also includes using Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance the contrast of the laryngeal mucosa and extracting vocal cord vibration frequency features through a Gabor filter bank to improve image quality.

[0048] S2: Input the standard image into the RFS-Net model to extract spatial and spectral features; specifically, this includes the following steps:

[0049] An improved ResNet-50 network was used to extract local texture features from laryngoscope images using 3×3 convolutional kernels as spatial features, including mucosal blood vessel distribution and ulcer edges.

[0050] A wavelet transform algorithm is introduced to decompose the image into low-frequency approximate components and high-frequency detail components, capturing high-frequency pathological signals in the lesion area as spectral features, including edge features of abnormally proliferating tissue.

[0051] Spatial features and spectral features are spliced ​​together channel by channel.

[0052] The cross-attention mechanism (CAM) dynamically adjusts the weights of features in the two branches, improving the dynamic adaptability of features for subsequent use.

[0053] S3: Based on the improved Focal Loss function (cross-entropy loss function), the spatial and spectral features extracted in step S2 are combined to train the model and balance the sample distribution; the improved Focal Loss function is:

[0054] FL(pt) = -a(1-pt) γ log(pt)·e -λ·|pt-0.5|

[0055] Where γ is the adjustment factor; λ is the class balance coefficient; pt is the model predicted probability; e is the base of the natural logarithm, equal to 2.71828; a is the balance factor used to adjust the weights of positive and negative samples in the loss function; |pt-0.5| represents the absolute difference between the predicted probability pt and 0.5; when pt is close to 0.5, this difference is small, and the exponential term e -λ·|pt-0.5| A larger value of pt increases the loss value of the sample; when pt is far from 0.5, the difference is large and the exponential term is small, thus reducing the loss value of the sample; the classification loss weight is dynamically adjusted according to the difficulty of the sample, and the contribution of easy-to-classify samples is suppressed by combining spatial features and spectral features. A class weight parameter is introduced to balance the samples.

[0056] Experiments on the LPR dataset (20% positive samples) show that the improved Focal Loss increases the model's F1-score for the positive class from 0.76 to 0.89, reducing the false negative rate by 42%. Through a confidence-sensitive adjustment term, the model's misclassification rate for noisy samples such as reflections and motion blur in laryngoscope images decreases from 18.3% to 7.1%. The dynamic α mechanism enables the model to quickly focus on positive samples in the early stages of training, reducing the convergence rounds from 120 rounds with traditional Focal Loss to 85 rounds, improving training efficiency by 29%. Parameter settings: γ = 2.0, λ = 3.0, initial rtarget = 0.2, k = 5.0; joint optimization of the loss function and DiceLoss: total loss Ltotal = 0.7·FLimproved + 0.3·LDice, to balance class balance and region overlap.

[0057] S4: Shallow texture features and deep semantic features are fused using the Residual Attention Mechanism (RAM) to obtain fused features; the RAM uses the following formula to weight features:

[0058]

[0059] Among them, Fout and F in These represent output and input features, respectively, used to describe the flow of data; σ represents element-wise multiplication, used to apply attention weights to the input features; σ is the Sigmoid function, used to map the input to the range (0, 1) to generate attention weights; W1 and W2 are learnable weight matrices, used to transform features in the channel dimension and capture the relationship between different channels; δ is the ReLU function, used to introduce non-linearity to increase the model's expressive power; GlobalAvgPool is global average pooling, used to compress features in the spatial dimension into a scalar; b1 and b2 represent bias terms, used to adjust the model's fitting ability; the original feature information is preserved through residual connections, and different feature weights are assigned according to the importance of lesion features using the attention mechanism; channel weights are generated through global average pooling and 1×1 convolution, and a spatial weight map is generated using 3×3 depthwise separable convolution; the original features and attention-weighted features are added element-wise, and the shallow texture features and deep semantic features of the image samples are fused to obtain the fused features.

[0060] Shallow texture feature extraction: A lightweight convolutional network (MobileNetV2) is used to extract shallow texture features from the image.

[0061] Deep semantic feature extraction: Deep semantic features of the image are extracted using ResNet-101. Attention fusion: Feature weights are dynamically adjusted through channel attention (CA) and spatial attention (SA) modules to fuse shallow and deep features.

[0062] S5: Combine clinical indicators with the fusion features from step S4 to perform pathological analysis on patient data and output the test results. Clinical indicators include the reflux symptom index (RSI), reflux symptom score (RFS), pH monitoring value, and pepsin level. Specifically, this includes the following steps:

[0063] Input a patient's laryngoscope image, and output the lesion probability value P1 by combining and comparing features using the RFS-Net model.

[0064] The patient's reflux symptom index (RSI), reflux symptom score (RFS), pH monitoring value, and pepsin level were collected.

[0065] RSI is obtained through a patient self-assessment questionnaire, while RFS is assessed by a physician based on eight signs, ultimately yielding clinical indicators.

[0066] The positive results of clinical indicators are converted into the lesion probability value P2.

[0067] Combine P1 and P2 to output the detection result.

[0068] Laryngoscopy image features were analyzed using the RFS-Net model to output the probability value P1 of LPR lesions. Simultaneously, the patient's Regurgitation Symptom Index (RSI) and Regurgitation Sign Score (RFS) were collected. The RSI was obtained through a patient self-assessment questionnaire (containing 9 symptoms, such as hoarseness or speech impairment, frequent throat clearing, excessive sputum or postnasal drip, difficulty swallowing food / liquids / pills, coughing after meals or lying down, shortness of breath or a feeling of suffocation, bothersome cough, a feeling of a foreign body in the throat (globus hystericus), heartburn or chest pain / acid reflux; the specific score was determined based on the patient's actual condition, with each item scored from 0 to 5 points, and a total score ≥13 points considered positive). The RFS was assessed by the physician based on 8 physical signs (such as vocal cord edema, erythema, false vocal cord groove, disappearance of the laryngeal ventricle, diffuse edema, and posterior commissure enlargement). The criteria for diagnosis include granulomas, laryngeal mucus adhesions, etc. (specifically determined based on the type of disease, each item is scored from 0-4, with a total score ≥7 considered positive). A professional assessment is then conducted, and the positive results of RSI and RFS are ultimately converted into a probability value P2. Weighted voting: Image feature weight ω1 = 0.6, clinical indicator weight ω2 = 0.4, and the comprehensive diagnostic probability is calculated using the formula P = ω1 × P1 + ω2 × P2. When P ≥ 0.5, it is considered LPR positive; otherwise, it is negative. Dynamic adjustment: When the initial P1 value output by the model is in the range [0.3, 0.7], the clinical indicator verification process is triggered. At this time, the weights are dynamically adjusted to ω1 = 0.4 and ω2 = 0.6 to strengthen the decision support of clinical indicators for ambiguous cases.

[0069] Example 2

[0070] according to Figure 1 , 2 As shown, this embodiment proposes a method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm, including the following steps:

[0071] Data Acquisition and Preprocessing:

[0072] We collected 3,000 laryngoscopy images (1,000 normal, 1,000 suspected, and 1,000 confirmed cases), and simultaneously recorded patients' pH monitoring values, pepsin levels, and symptom scores.

[0073] The image is enhanced using AHE, denoised using NLM, and normalized.

[0074] Model building and training:

[0075] The RFS-Net model is constructed, including a spatial feature extraction network (ResNet-50) and a spectral feature extraction network (STFT+CNN).

[0076] The model was trained using an improved Focal Loss function.

[0077] Feature fusion and classification:

[0078] Shallow texture features (extracted via MobileNetV2) and deep semantic features (extracted via ResNet-101) are fused using the RAM mechanism.

[0079] Image features are concatenated with clinical indicators and then input into a multilayer perceptron (MLP) for classification.

[0080] Experimental results:

[0081] On the test set, the accuracy of the method of this invention is 92.5%, which is significantly higher than that of traditional CNN methods (78.3%) and RF-Net (85.2%).

[0082] Example 3

[0083] according to Figure 1 , 2 As shown, this embodiment proposes a method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm, including the following steps:

[0084] Image preprocessing:

[0085] Input the original laryngoscope image (resolution 1920×1080) and remove noise using non-local means denoising (search window 7×7, similarity threshold 0.1);

[0086] The color space is standardized to the DICOM preset range using histogram matching;

[0087] The laryngeal mucosa region is segmented using U-Net++ (encoder: EfficientNet-B0), and the ROI mask is output.

[0088] Feature extraction:

[0089] Spatial branch: ResNet-50 stage 3 output feature map (size 28×28×512);

[0090] Spectrum branch: Perform STFT (window length 256, overlap rate 50%) on the ROI, generate a spectrogram, and input it into MobileNetV3-Large.

[0091] Feature fusion:

[0092] Shallow features (ResNet-50 stage 1) and deep features (stage 4) are input into RAM respectively;

[0093] The RAM output feature maps are then concatenated after being reduced to 256 dimensions using 1×1 convolution.

[0094] Model training:

[0095] The initial learning rate is 0.001, and a cosine annealing strategy is used.

[0096] The Focal Loss parameters were improved to α = 0.25, γ = 2, and λ = 3.

[0097] Train for 100 rounds on the LPR dataset (containing 1200 positive samples and 4800 negative samples).

[0098] Clinical decision fusion:

[0099] When the model outputs a probability p∈[0.3,0.7], the RSI questionnaire (≥13 points) and RFS score (≥7 points) are activated for verification.

[0100] The final detection result is a weighted sum of image probability and clinical indicators (weight ratio 3:2).

[0101] Example 4

[0102] according to Figure 1 , 2 As shown, this embodiment proposes a laryngopharyngeal reflux disease detection system based on laryngoscope images and the RFS-Net algorithm, including a data acquisition module, a feature extraction module, a feature fusion module, a model training module, and a classification decision module.

[0103] The data acquisition module is used to collect and store laryngoscope image samples, and also to acquire patient laryngoscope images and clinical indicator data in real time.

[0104] The feature extraction module is based on the RFS-Net model and is used to extract spatial and spectral features of image samples.

[0105] The model training module uses an improved Focal Loss function combined with spatial and spectral features extracted from the feature extraction module for training, thereby balancing the sample distribution.

[0106] The feature fusion module is used to fuse shallow texture features and deep semantic features of image samples through a residual attention mechanism (RAM) to obtain fused features.

[0107] The classification decision module is used to compare the patient's real-time laryngoscope image with fused features and output the detection results in combination with clinical indicators.

[0108] The data acquisition module has a built-in preprocessing module, which is used to perform noise reduction, color correction, image enhancement, ROI extraction, enhance the contrast of the laryngeal mucosa, and extract the vocal cord vibration frequency features.

[0109] The classification decision module also includes a multilayer perceptron (MLP) for classifying fused features and clinical indicators.

[0110] This invention utilizes the RFS-Net model for multi-dimensional feature extraction and fusion of laryngoscopy images, encompassing spatial and spectral features. Combined with an improved Focal Loss function to optimize model training, it can more accurately identify the characteristics of laryngopharyngeal reflux disease, reducing misdiagnosis and missed diagnosis, and achieving higher accuracy. Furthermore, this invention employs a residual attention mechanism (RAM) and an improved Focal Loss function, making the model more adaptable to laryngoscopy images of varying quality and imbalanced datasets, thus enhancing its robustness. Simultaneously, during the decision-making process, diagnosis is performed by combining fused features with clinical indicators. The determined lesion probability values ​​P1 and P2 are combined to output the detection results, providing doctors with more comprehensive information, helping them make more scientific and rational clinical decisions, and improving treatment outcomes.

[0111] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm, characterized in that, Includes the following steps: S1: Obtain laryngoscope image samples, perform preprocessing, and obtain standard images; S2: Construct the RFS-Net model, including a spatial feature extraction network and a spectral feature extraction network. Input the standard image into the RFS-Net model to extract spatial and spectral features, including the following steps: An improved ResNet-50 network was used to extract local texture features from laryngoscope images using 3×3 convolutional kernels as spatial features, including mucosal blood vessel distribution and ulcer edges; The wavelet transform algorithm is introduced to decompose the image into low-frequency approximate components and high-frequency detail components, capturing high-frequency pathological signals in the lesion area as spectral features, including edge features of abnormally proliferating tissue; Spatial features and spectral features are stitched together channel by channel; The weights of the two branches are dynamically adjusted using the cross-attention mechanism (CAM). S3: Based on the improved Focal Loss function, the spatial and spectral features extracted in step S2 are combined to train the model and balance the sample distribution; In S3, the improved Focal Loss function is: ； in, As a regulating factor; This is the category balance coefficient; Predict probabilities for the model; The base of the natural logarithm is 2.71828; This is a balancing factor used to adjust the weights of positive and negative samples in the loss function; Indicates the predicted probability The absolute difference between 0 and 0.5; when When the value is close to 0.5, the difference is small, and the exponential term... A larger value increases the loss value of that sample; when When the value is far from 0.5, the difference is large and the exponential term is small, thus reducing the loss value of the sample; the classification loss weight is dynamically adjusted according to the difficulty of the sample, and the contribution of easy-to-classify samples is suppressed by combining spatial features and spectral features, and the class weight parameter is introduced to balance the samples. S4: The shallow texture features and deep semantic features of image samples are fused through the residual attention mechanism RAM to obtain fused features; S5: Combine clinical indicators and the fusion features from step S4 to perform lesion analysis on patient data and output the test results.

2. The method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm according to claim 1, characterized in that: In S1, the preprocessing includes denoising, color correction, image enhancement, and ROI extraction. The ROI extraction is based on the U-Net model to automatically locate the anatomical structure of the larynx. The preprocessing also includes using adaptive histogram equalization (CLAHE) to enhance the contrast of the laryngeal mucosa and extracting the vocal cord vibration frequency features through Gabor filter banks.

3. The method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm according to claim 1, characterized in that: In S4, the residual attention mechanism RAM implements feature weighting through the following formula: ； in, and These represent output and input features, respectively, used to describe the flow of data; This indicates element-wise multiplication, used to apply attention weights to the input features; The sigmoid function is used to map the input to the range (0, 1) to generate attention weights; , It is a learnable weight matrix used to transform features along the channel dimension and capture the relationships between different channels; This is the ReLU function, used to introduce nonlinearity to increase the expressive power of the model; This is global average pooling, used to compress features across spatial dimensions into a single scalar. and The bias term is used to adjust the model's fitting ability. The original feature information is preserved through residual connections, and different feature weights are assigned according to the importance of lesion features using an attention mechanism. Channel weights are generated through global average pooling and 1×1 convolution, and a spatial weight map is generated using 3×3 depthwise separable convolution. The original features and attention-weighted features are added element by element, and the shallow texture features and deep semantic features of the image samples are fused to obtain the fused features.

4. The method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm according to claim 3, characterized in that: In S5, the clinical indicators include the reflux symptom index (RSI), the reflux symptom score (RFS), pH monitoring value, and pepsin level.

5. The method for detecting laryngopharyngeal reflux disease based on laryngoscopy images and the RFS-Net algorithm according to claim 4, characterized in that: S5 includes the following steps: Input a patient's laryngoscopy image, and output the lesion probability value P1 by combining and comparing features through the RFS-Net model; The patient's reflux symptom index (RSI), reflux symptom score (RFS), pH monitoring value, and pepsin level were collected. RSI is obtained through a patient self-assessment questionnaire, while RFS is assessed by a physician based on N signs to ultimately obtain clinical indicators. The positive results of clinical indicators are converted into the lesion probability value P2; Combine P1 and P2 to output the detection result.

6. A laryngopharyngeal reflux disease detection system based on laryngoscopy images and the RFS-Net algorithm, applied to the laryngopharyngeal reflux disease detection method based on laryngoscopy images and the RFS-Net algorithm described in any one of claims 1-5, characterized in that, It includes a data acquisition module, a feature extraction module, a feature fusion module, a model training module, and a classification decision module; The data acquisition module is used to collect and store laryngoscope image samples, and also to acquire patient laryngoscope images and clinical indicator data in real time. The feature extraction module is based on the RFS-Net model and is used to extract spatial and spectral features of image samples. The model training module uses an improved Focal Loss function combined with spatial and spectral features extracted from the feature extraction module for training, balancing the sample distribution; The feature fusion module is used to fuse shallow texture features and deep semantic features of image samples through a residual attention mechanism (RAM) to obtain fused features. The classification decision module is used to compare the patient's real-time laryngoscope image with fused features and output the detection results in combination with clinical indicators.

7. The laryngopharyngeal reflux disease detection system based on laryngoscopy images and RFS-Net algorithm according to claim 6, characterized in that: The data acquisition module has a built-in preprocessing module, which is used to perform noise reduction, color correction, image enhancement, ROI extraction, enhance the contrast of the laryngeal mucosa, and extract the vocal cord vibration frequency features.

8. The pharyngeal reflux disease detection system based on laryngoscopy images and RFS-Net algorithm according to claim 6, characterized in that: The classification decision module also includes a multilayer perceptron (MLP) for classifying fused features and clinical indicators.

Images