Training method and system of vocal cord tumor multi-classification analysis model based on voice

Abstract

The invention is suitable for the technical field of machine learning classification, and provides a voice-based vocal cord tumor multi-classification analysis model training method and system, and the method comprises the following steps: voice data collection; processing data; constructing a model; model training and optimization: inputting the initial tensor into a convolution module, and extracting to obtain an initial feature map; the initial feature map is sequentially sent to a multi-stage LHG module to be processed, and an updated feature map is output; and according to the updated feature map, outputting probability distributions corresponding to three categories of vocal cord polyps, vocal cord precancerous lesions and vocal cord malignant tumors. According to the method, the pathological type of the vocal cord tumor can be noninvasively, accurately and automatically analyzed, and the problems that an existing method depends on invasive biopsy, and an existing intelligent algorithm is insufficient in performance in a multi-classification analysis task are solved.

Description

Technical Field

[0001] This invention belongs to the field of machine learning classification technology, and in particular relates to a training method for a multi-classification analysis model of vocal cord tumors based on voice. Background Technology

[0002] The voice is a core medium for human communication, emotional expression, and social participation, and changes in its quality can significantly impact a patient's quality of life. Vocal cord tumors—primarily including benign lesions (such as vocal cord polyps), precancerous lesions (such as hyperkeratosis and dysplasia), and malignant tumors (such as squamous cell carcinoma)—are common organic causes of persistent, progressive hoarseness. Clinically, accurately distinguishing the pathological nature of these lesions is crucial for developing appropriate patient management and treatment strategies.

[0003] Currently, traditional methods for analyzing vocal cord tumors mainly rely on laryngoscopy (such as electronic laryngoscopy) and pathological biopsy. However, these methods have significant limitations: firstly, the interpretation of laryngoscopy results is highly subjective and depends heavily on the physician's expert experience; secondly, pathological biopsy is an invasive procedure that can cause discomfort and potential risks to patients and is not suitable for large-scale initial screening or dynamic monitoring.

[0004] In recent years, with the development of artificial intelligence technology, non-invasive assisted analysis based on voice audio has become an important research direction. Voice is not only a clinical manifestation of vocal cord lesions, but its acoustic characteristics also contain rich pathological information. Researchers are attempting to input various features extracted from patients' audio signals (such as Mel frequency cepstral coefficients (MFCC), Mel spectrograms, etc.) into machine learning or deep learning models (such as support vector machines (SVM) and convolutional neural networks (CNN)) to achieve automatic classification of pathological voices.

[0005] While existing research has achieved high accuracy in binary classification tasks (such as distinguishing between normal and pathological voices), the performance of models still needs improvement in more clinically challenging multi-classification tasks—that is, accurately identifying vocal cord polyps, precancerous lesions, and malignant tumors simultaneously. This is mainly because precancerous vocal cord lesions and early-stage malignant tumors have similar acoustic characteristics, making differentiation difficult. Furthermore, many existing models struggle to effectively and efficiently integrate subtle local anomalies with global acoustic structural contextual information when processing audio spectrograms, resulting in insufficient discriminative power, low computational efficiency, and consequently affecting the accuracy of the final classification and the model's generalization ability.

[0006] Therefore, there is an urgent need in this field for an intelligent analysis model training method that can fully utilize voice audio information, efficiently fuse local and global features of spectrograms, and demonstrate high accuracy and robustness in multi-classification tasks of vocal cord tumors, so as to provide a reliable non-invasive screening tool for clinical use and assist doctors in analysis and identification. Summary of the Invention

[0007] The purpose of this invention is to provide a training method and system for a multi-classification analysis model of vocal cord tumors based on voice, in order to solve the above-mentioned technical problems.

[0008] This invention is implemented as follows: a training method for a multi-classification analysis model of vocal cord tumors based on voice, comprising the following steps:

[0009] Voice data collection: Collecting audio data of the voices of patients with pathologically confirmed vocal cord lesions;

[0010] Data processing: The voice audio data is preprocessed to generate a Mel spectrogram, and the Mel spectrogram is then subjected to data augmentation and segmentation to map and form an initial tensor;

[0011] Model Construction: A local high-order graph neural network model is constructed; the local high-order graph neural network model includes convolutional modules and multi-stage stacked LHG modules;

[0012] Model training and optimization: The initial tensor is input into the convolution module to extract the initial feature map; the initial feature map is then sequentially fed into a multi-stage LHG module for processing, and the updated feature map is output; based on the updated feature map, the probability distributions corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors are output; based on the probability distributions and the real pathological labels, the loss function is calculated as the optimization objective, and the parameters of the local high-order graph neural network model are iteratively updated through the backpropagation algorithm to complete the model training.

[0013] Furthermore, the step of preprocessing the voice audio data to generate a Mel spectrogram specifically includes:

[0014] The voice audio data is framed, windowed, and subjected to discrete Fourier transform to convert it from a time-domain signal to a frequency-domain signal, and then divided into multiple frequency bands to obtain the Mel spectrogram.

[0015] Furthermore, the data augmentation process includes one or more combinations of random occlusion in the time and frequency domains, adding random Gaussian noise, and random scrolling in the time domain; the segmentation process is as follows: the Mel spectrogram processed by the data augmentation is segmented into image blocks of a preset fixed size, and mapped to feature embedding vectors through linear projection to form an initial tensor.

[0016] Furthermore, in each LHG module, the following two graph structures are constructed simultaneously based on the current initial feature map:

[0017] Local k-Nearest Neighbor Graph: Let the spatial size of the input feature map be H×W. Each spatial position is regarded as a node, and the number of nodes is N = H×W. For each node, k spatially nearest nodes are selected based on their Euclidean distances in the feature map to construct an adjacency relationship, obtaining a local adjacency matrix. Different from conventional graph neural networks that construct graphs only relying on feature similarities, this method explicitly preserves the two-dimensional spatial structure of the spectrogram and enhances the spatial consistency modeling of local abnormal patterns.

[0018] Higher-order Fuzzy C-Means Clustering Graph: The Fuzzy C-Means clustering algorithm is used to cluster all node features into m categories, where m << N, obtaining cluster centers and a membership matrix ; A bipartite graph between nodes and cluster centers is constructed , and the edge weights are determined by the membership ; This structure is used to capture the non-local global acoustic pattern distribution and compensate for the long-distance dependencies that are difficult to model by local graphs.

[0019] The LHG module performs graph convolution on the local k-nearest neighbor graph and the higher-order fuzzy C-means clustering graph respectively through the message passing mechanism of the graph neural network, aggregating information from local neighbor nodes and higher-order cluster centers. The two types of updated features are fused through learnable weights to achieve efficient fusion of local detailed anomalies and global structural contexts, and an updated feature map is output.

[0020] Furthermore, the local higher-order graph neural network model further includes:

[0021] A downsampling layer, used to reduce the spatial resolution of the feature map through convolutional operations with a stride greater than 1 between different stages.

[0022] A classification head, used to perform global average pooling and fully connected layer operations on the updated feature map, and output the probability distributions corresponding to three categories of vocal cord polyps, premalignant lesions of vocal cords, and malignant tumors of vocal cords.

[0023] Furthermore, the loss function is a cross-entropy loss function.

[0024] Another object of the present invention is to provide a vocal cord tumor multi-class analysis model trained by using the training method of the above-mentioned voice-based vocal cord tumor multi-class analysis model.

[0025] Another object of the present invention is to provide a training system for a voice-based vocal cord tumor multi-class analysis model, used to implement the training method of the above-mentioned voice-based vocal cord tumor multi-class analysis model, specifically including:

[0026] The data acquisition module is used for voice data collection: collecting voice audio data from patients with pathologically confirmed vocal cord lesions;

[0027] The data processing module is used to preprocess the voice audio data, generate a Mel spectrogram, and perform data augmentation and segmentation processing on the Mel spectrogram to map it into an initial tensor.

[0028] The model building module is used to build a local high-order graph neural network model; the local high-order graph neural network model includes a convolution module and a multi-stage stacked LHG module;

[0029] The model training and optimization module is used to input the initial tensor into the convolution module to extract the initial feature map; the initial feature map is then sequentially fed into a multi-stage LHG module for processing, and an updated feature map is output; based on the updated feature map, a probability distribution corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors is output; based on the probability distribution and the real pathological labels, a loss function is calculated as the optimization objective, and the parameters of the local high-order graph neural network model are iteratively updated through the backpropagation algorithm to complete the model training.

[0030] Furthermore, the system also includes a model evaluation module, which uses accuracy, recall, F1 score, AUC value and confusion matrix as evaluation metrics to evaluate the performance of the trained model.

[0031] Furthermore, the system also includes a model deployment module for deploying the trained model to assist in the classification and prediction of pathological types for new, unknown voice audio.

[0032] The training method for a vocal cord tumor multi-classification analysis model based on voice provided by this invention constructs a deep learning model specifically for three-classification tasks and innovatively introduces a local high-order graph neural network architecture to simultaneously capture subtle local pathological traces and global systemic abnormalities in the spectrogram. This enhances the model's ability to analyze complex cases such as precancerous lesions and malignant tumors. It can automatically analyze the pathological types of vocal cord tumors (including vocal cord polyps, precancerous lesions, and malignant tumors) non-invasively and accurately, overcoming the problems of existing methods relying on invasive biopsies and the insufficient performance of existing intelligent algorithms in multi-classification analysis tasks. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the training method for a multi-classification analysis model of vocal cord tumors based on voice, as provided in an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of a voice-based multi-classification analysis model for vocal cord tumors provided in an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0036] like Figure 1 As shown, in one embodiment of the present invention, a multi-classification analysis model for vocal cord tumors based on voice is provided, the construction and training method of which specifically includes the following steps:

[0037] S1. Voice data collection: Collecting voice audio data of patients with pathologically confirmed vocal cord lesions; specifically, the voice audio data includes steady vowels produced by the patient at a comfortable loudness and continuous sentences read by the patient.

[0038] S2. Data Processing: The voice audio data is preprocessed to generate a Mel-spectrogram, and the Mel-spectrogram is subjected to data augmentation and segmentation processing to map and form an initial tensor.

[0039] The preprocessing methods include: framing, windowing, and performing discrete Fourier transform on the voice audio data to convert it from a time-domain signal to a frequency-domain signal, and dividing it into multiple frequency bands to obtain a Mel spectrogram. Data augmentation includes one or more combinations of random occlusion in the time and frequency domains, adding random Gaussian noise, and random scrolling in the time domain. The segmentation method involves dividing the data-augmented Mel spectrogram into image patches of a preset fixed size, and mapping them to feature embedding vectors through linear projection to form an initial tensor.

[0040] S3. Model Construction: Construct a Local-Higher Order Graph Neural Network (LHGNN) model; the LHGNN model includes convolutional modules and multi-stage stacked LHG modules;

[0041] Specifically, such as Figure 2 As shown, the local high-order graph neural network model used in this embodiment of the invention adopts a pyramid-shaped hierarchical structure, including:

[0042] (1) Convolutional module, namely the convolutional backbone network (Stem Block): It consists of multiple convolutional layers and is used to initially extract basic local visual features from the input Mel spectrogram embedding to generate an initial feature map.

[0043] (2)Multi-stage Stacked Local-Higher Order Graph (LHG) Module (abbreviated as LHG module): The initial feature map is fed into multiple stages for processing, and each stage is stacked by multiple LHG modules. In each of the LHG modules, the following two graph structures are constructed simultaneously based on the current initial feature map:

[0044] ① Local k-Nearest Neighbors (k-NN) graph: Let the spatial size of the input feature map be H×W. Each spatial position is regarded as a node, and the number of nodes is N = H×W. For each node, based on its Euclidean distance in the feature map, the k nearest nodes in space are selected to construct an adjacency relationship, obtaining the local adjacency matrix A local . Different from conventional graph neural networks that construct graphs only relying on feature similarity, this method explicitly retains the two-dimensional spatial structure of the spectrogram, enhances the spatial consistency modeling of local abnormal patterns, and is used to capture subtle abnormalities and local patterns in adjacent regions of the spectrogram, obtaining the information of local neighbor nodes.

[0045] ② Higher-order Fuzzy C-Means clustering graph: The fuzzy C-means clustering algorithm is used to cluster all node features into m categories (m << N), obtaining the cluster centers {c1, …, c m} and the membership matrix U ∈ R^(N×m). A bipartite graph G hign =(V, C, E) is constructed between the nodes and the cluster centers, and the edge weights are determined by the membership u ij . This structure is used to capture the non-local global acoustic pattern distribution, compensate for the long-distance dependencies that are difficult to model by local graphs, and obtain the information of higher-order cluster centers.

[0046] This LHG module synchronously aggregates the information from local neighbor nodes and higher-order cluster centers through the message passing mechanism of the graph neural network: perform graph convolution on the local k-NN graph and the higher-order Fuzzy C-Means clustering graph respectively to aggregate the information from local neighbor nodes and higher-order cluster centers; the two types of updated features are fused through learnable weights to achieve the efficient fusion of local detailed abnormalities and global structural contexts, and output the updated feature map.

[0047] (3)Downsampling layer: Between different stages, the spatial resolution of the feature map is reduced through convolutional operations with a stride greater than 1, reducing the computational complexity while expanding the receptive field.

[0048] (4)Classification head: Finally, the feature map obtained after processing through multiple stages undergoes operations such as global average pooling and fully connected layers, and outputs the probability distributions corresponding to the three categories of vocal cord polyps, premalignant lesions of the vocal cord, and malignant tumors of the vocal cord.

[0049] S4. Model Training and Optimization: The initial tensor is input into the convolution module to extract the initial feature map; the initial feature map is sequentially fed into a multi-stage LHG module for processing, and the updated feature map is output; based on the updated feature map, the probability distributions corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors are output; based on the probability distributions and the real pathological labels, the loss function is calculated as the optimization objective, and the parameters of the local high-order graph neural network model are iteratively updated through the backpropagation algorithm to complete the model training;

[0050] Specifically, the labeled dataset is divided into training and test sets in a proportional ratio (e.g., 8:2). Model training uses the cross-entropy loss function as the optimization objective to measure the difference between the model's predictions and the actual pathological labels. The optimization process uses the adaptive moment estimation optimizer (Adam) with appropriate hyperparameters such as learning rate and weight decay. During training, all parameters of the LHGNN model are iteratively updated using the backpropagation algorithm until the model's performance on the test set (e.g., accuracy, AUC value) reaches its optimal and stabilizes.

[0051] S5. Model Evaluation: The performance of the trained model is evaluated using accuracy, recall, F1 score, AUC value and confusion matrix as evaluation metrics.

[0052] S6. Model Implementation and Deployment: Deploy the trained model to assist in the classification and prediction of pathological types for new, unknown voice audio.

[0053] In this embodiment of the invention, a completely non-invasive and objective intelligent analysis tool—a multi-classification analysis model for vocal cord tumors—was developed by analyzing conventionally recorded voice audio data. Based on deep learning technology, it achieves non-invasive and automatic identification of three pathological types: vocal cord polyps, precancerous lesions, and malignant tumors of the vocal cords, thereby assisting in the early identification of vocal cord tumors. To achieve multi-classification analysis and identification of vocal cord tumor pathology, this embodiment of the invention constructs a deep learning model specifically for the three-classification task. Through an innovative LHGNN architecture, it simultaneously captures subtle local pathological traces and global systemic anomalies in Mel spectrograms, effectively fusing local details and global contextual information to extract highly discriminative pathological acoustic features, thereby improving the model's ability to analyze complex cases such as precancerous lesions and malignant tumors. Ultimately, this embodiment of the invention aims to overcome the shortcomings of existing technologies, such as invasiveness, subjectivity, insufficient feature fusion, and poor multi-classification performance, achieving efficient and accurate non-invasive auxiliary analysis of vocal cord tumor pathological types.

[0054] Experimental Case: Medical records from patients who visited the Department of Otolaryngology at the Second Hospital of Jilin University between January 2019 and December 2024 were retrospectively reviewed. Patients with vocal cord tumors who were pathologically diagnosed with vocal cord polyps, precancerous lesions, or vocal cord cancer were screened. These patients underwent voice assessment and electronic laryngoscopy at their initial visit, and all underwent histopathological examination during subsequent treatment. Patients under 18 years of age and those who had previously undergone vocal cord surgery or head and neck radiotherapy were excluded. All abnormal voices were classified according to pathological results: vocal cord polyps, precancerous lesions, and vocal cord cancer. Precancerous lesions included those with pathological reports of hyperkeratosis, hyperplasia, and mild to severe dysplasia; vocal cord cancer included patients with pathological reports of squamous cell carcinoma.

[0055] Voice samples were collected in a dedicated voice examination room with background noise controlled below 45 dB HL. The audio acquisition device was a Steinberg UR22C, with the microphone positioned 12-14 cm from the lips at a 90-degree angle. Praat (version 5.4.22) was used to analyze the digitally recorded speech samples at a sampling frequency of 50,000 Hz. Participants intermittently uttered five vowels and read a continuous sentence at a comfortable loudness (approximately 55-65 dB HL), with vowel durations exceeding 10 seconds. This trial was approved by the Ethics Review Committee of the Second Hospital of Jilin University (Approval No.: SB2025-241) and conducted in accordance with relevant guidelines and regulations. Due to the retrospective nature of this study, informed consent was waived by the Ethics Committee.

[0056] The patients were divided into training and validation sets in an 8:2 ratio. Audio was extracted in 10-second intervals with a 2-second overlap. The audio signal was framed, windowed, and subjected to Discrete Fourier Transform to convert the time-domain signal to the frequency-domain signal, and divided into 128 frequency bands. Using 10 seconds as the baseline, the extracted time-domain dimension was approximately 1024, resulting in a 1×1024×128 feature vector. The feature vector was randomly occluded in both the time and frequency domains (frequency mask of 24, time mask of 256), and random noise (range 0-1) was added. Random forward or backward scrolling of 10 units was performed in the time domain to extract features from different angles. The Mel-spectrum was divided into 16×16 patches and linearly projected into the embedding space. Specifically, a 16×16 convolution kernel was used, with a stride of 10 in both the time and frequency domains, overlapping sampling, an overlap coefficient of 6, and 768 embedded features. By employing 2D convolution and linear projection, the Mel-spectrum is transformed from a feature of 1×1024×128 to a feature of 1×1212×768.

[0057] The Local High-Order Graph Neural Network (LHGNN) model used in this experimental example employs a pyramid-shaped hierarchical structure, mainly including:

[0058] I. Convolutional Module, i.e., Convolutional Backbone Network: Composed of four 3×3 convolutional layers (with strides of 2, 1, 2, and 1 respectively), used to initially extract basic local visual features from the input Mel spectrogram embedding (sized normalized to 1024×128 frames) to generate an initial feature map. .

[0059] II. Multi-stage LHG module stacking: Feature maps are processed in four stages, each consisting of multiple stacked LHG modules. The number of repetitions for stages 1-4 is [2, 2, 6, 2], and the number of channels increases progressively with each stage [80, 160, 320, 640]. Each LHG module simultaneously constructs two graph structures:

[0060] Local k-nearest neighbor graph: Constructed based on the spatial proximity relationships of nodes (image patches) in the feature map, used to capture subtle anomalies and local patterns in adjacent regions of the spectrogram.

[0061] Higher-order fuzzy C-means clustering graph: Using the fuzzy C-means clustering algorithm, feature nodes are divided into multiple clusters, and a relationship graph between nodes and cluster centers is constructed. This structure is used to model non-local, global structural associations and pattern imbalances in spectrograms.

[0062] The LHG module uses the message passing mechanism of graph neural networks to synchronously aggregate information from local neighbor nodes and higher-order cluster centers, thereby achieving efficient fusion of local detail anomalies and global structural context.

[0063] 3. Downsampling layer: Between different stages, the spatial resolution of the feature map is reduced by a 3×3 convolution operation with a stride of 2, which expands the receptive field while reducing computational complexity.

[0064] IV. Classification Head: Finally, the features obtained after four stages of processing are processed through global average pooling, fully connected layers, and other operations to output the probability distributions corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors.

[0065] The labeled dataset was divided into training and test sets in an 8:2 ratio. The model training employed cross-entropy loss as the optimization objective to measure the difference between the model's predictions and the actual pathological labels. The optimization process used an adaptive moment estimation optimizer (Adam) with an initial learning rate of 1e-5, L2 norm regularization weight decay of 5e-7, an exponential decay rate of 0.95 for first-order moment estimation, and an exponential decay rate of 0.999 for second-order moment estimation. Model evaluation metrics included accuracy (ACC), recall, F1 score, and AUC. During training, all parameters of the LHGNN model were iteratively updated using backpropagation until the model's performance on the test set reached its optimal and stabilized.

[0066] The implementation results of the above model are as follows: The final dataset, obtained through balanced sampling, contains 4745 audio samples (1603 vocal cord polyps, 1598 precancerous lesions, and 1544 vocal cord cancers). The test set achieved its best performance after 461 training rounds: overall accuracy of 92.94%, F1 score of 0.9289, recall of 0.9281, and AUC of 0.9877. These data demonstrate that the model possesses extremely high comprehensive discriminative ability and reliability in the multi-class classification task of vocal cord tumor pathology. The multi-class confusion matrix shows that the model's accuracy in analyzing vocal cord polyps, precancerous lesions, and malignant tumors is 91.35%, 90.99%, and 94.05%, respectively. This indicates that the model not only has good overall performance but also demonstrates robust and excellent performance in distinguishing between precancerous lesions and malignant tumors, which is the most difficult and critical aspect of clinical differentiation.

[0067] In summary, this invention, through specific data processing procedures, LHGNN model architecture, and training methods, constructs a model training system capable of deeply mining pathological features from vocal audio and achieving high-precision multi-classification analysis, providing a complete solution for non-invasive auxiliary analysis of vocal cord tumors. This model achieves breakthrough performance (92.94% accuracy and 0.987 AUC) when facing the challenging task of multi-classification of vocal cord lesions, providing a powerful auxiliary tool for the early screening and analysis of vocal cord tumors.

[0068] In another embodiment of the present invention, a training system for a voice-based multi-classification analysis model of vocal cord tumors is also provided to implement the above method, specifically including:

[0069] The data acquisition module is used for recording or collecting voice data: collecting voice audio data from patients with pathologically confirmed vocal cord lesions;

[0070] The data processing module is used to preprocess the voice audio data, generate a Mel spectrogram, and perform data augmentation and segmentation processing on the Mel spectrogram to map it into an initial tensor.

[0071] The model building module is used to build a local high-order graph neural network model; the local high-order graph neural network model includes a convolution module and a multi-stage stacked LHG module;

[0072] The model training and optimization module is used to input the initial tensor into the convolution module to extract the initial feature map; the initial feature map is then sequentially fed into a multi-stage LHG module for processing, and an updated feature map is output; based on the updated feature map, a probability distribution corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors is output; based on the probability distribution and the real pathological labels, a loss function is calculated as the optimization objective, and the parameters of the local high-order graph neural network model are iteratively updated through the backpropagation algorithm to complete the model training;

[0073] In a preferred embodiment of the present invention, the system further includes a model evaluation module, which is used to evaluate the performance of the trained model using accuracy, recall, F1 score, AUC value and confusion matrix as evaluation metrics.

[0074] In a preferred embodiment of the present invention, the system further includes a model deployment module for deploying the trained model for auxiliary classification and prediction of pathological types of new unknown voice audio.

[0075] It should be noted that each of the above modules can be implemented as a computer program, which can run on a computer device. The computer device's memory can store the computer program that makes up each module, enabling the processor to execute each step of the above method.

[0076] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0077] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods.

[0078] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A training method for a multi-classification analysis model of vocal cord tumors based on voice, characterized in that, Includes the following steps: Voice data collection: Collecting audio data of the voices of patients with pathologically confirmed vocal cord lesions; Data processing: The voice audio data is preprocessed to generate a Mel spectrogram, and the Mel spectrogram is then subjected to data augmentation and segmentation to map and form an initial tensor; Model Construction: A local high-order graph neural network model is constructed; the local high-order graph neural network model includes convolutional modules and multi-stage stacked LHG modules; Model training and optimization: The initial tensor is input into the convolution module to extract the initial feature map; the initial feature map is then sequentially fed into a multi-stage LHG module for processing, and the updated feature map is output; based on the updated feature map, the probability distributions corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors are output; based on the probability distributions and the real pathological labels, the loss function is calculated as the optimization objective, and the parameters of the local high-order graph neural network model are iteratively updated through the backpropagation algorithm to complete the model training.

2. The training method for the vocal cord tumor multi-classification analysis model based on voice according to claim 1, characterized in that, The step of preprocessing the voice audio data to generate a Mel spectrogram specifically includes: The voice audio data is framed, windowed, and subjected to discrete Fourier transform to convert it from a time-domain signal to a frequency-domain signal, and then divided into multiple frequency bands to obtain the Mel spectrogram.

3. The training method for the vocal cord tumor multi-classification analysis model based on voice according to claim 1, characterized in that, The data augmentation process includes one or more combinations of random occlusion in the time and frequency domains, adding random Gaussian noise, and random scrolling in the time domain; the segmentation process is as follows: the Mel spectrogram processed by the data augmentation is segmented into image blocks of a preset fixed size, and mapped into feature embedding vectors through linear projection to form an initial tensor.

4. The training method for the vocal cord tumor multi-classification analysis model based on voice according to claim 1, characterized in that, In each LHG module, the following two graph structures are constructed simultaneously based on the current initial feature map: Local k-nearest neighbor graph: Let the spatial size of the input feature map be H×W, and each spatial location be considered a node, with the number of nodes being N=H×W; for each node, select the k nearest nodes in the feature map based on its Euclidean distance to construct an adjacency relationship, resulting in a local adjacency matrix. ; High-order fuzzy C-means clustering graph: All node features are clustered into m categories using the fuzzy C-means clustering algorithm, where m << N, and the cluster centers are obtained. and the membership matrix ; A bipartite graph between nodes and cluster centers is constructed , and the edge weights are determined by the membership . The LHG module performs graph convolution on the local k-nearest neighbor graph and the high-order fuzzy C-means clustering graph through the message passing mechanism of the graph neural network, aggregating information from local neighbor nodes and high-order cluster centers. The two types of updated features are fused through learnable weights to achieve efficient fusion of local detail anomalies and global structural context, and output an updated feature map.

5. The training method for the vocal cord tumor multi-classification analysis model based on voice according to claim 4, characterized in that, The local high-order graph neural network model also includes: Downsampling layers are used to reduce the spatial resolution of feature maps between different stages by using convolutional operations with a stride greater than 1; The classification head is used to process the updated feature map through global average pooling and fully connected layer operations, and output the probability distribution corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors.

6. The training method for the vocal cord tumor multi-classification analysis model based on voice according to claim 1, characterized in that, The loss function is the cross-entropy loss function.

7. A vocal cord tumor multi-classification analysis model obtained by training the vocal cord tumor multi-classification analysis model based on the voice as described in any one of claims 1-6.

8. A training system for a voice-based multi-classification analysis model of vocal cord tumors, used to implement the training method for the voice-based multi-classification analysis model of vocal cord tumors according to any one of claims 1-6, characterized in that, include: The data acquisition module is used for voice data collection: collecting voice audio data from patients with pathologically confirmed vocal cord lesions; The data processing module is used to preprocess the voice audio data, generate a Mel spectrogram, and perform data augmentation and segmentation processing on the Mel spectrogram to map it into an initial tensor. The model building module is used to build a local high-order graph neural network model; the local high-order graph neural network model includes a convolution module and a multi-stage stacked LHG module; The model training and optimization module is used to input the initial tensor into the convolution module to extract the initial feature map; the initial feature map is then sequentially fed into a multi-stage LHG module for processing, and an updated feature map is output; based on the updated feature map, a probability distribution corresponding to the three categories of vocal cord polyps, vocal cord precancerous lesions, and vocal cord malignant tumors is output; based on the probability distribution and the real pathological labels, a loss function is calculated as the optimization objective, and the parameters of the local high-order graph neural network model are iteratively updated through the backpropagation algorithm to complete the model training.

Images