Risk prediction method for thyroid diseases

Abstract

The invention relates to the technical field of disease risk prediction, and discloses a risk prediction method for thyroid diseases, which comprises the following steps: acquiring thyroid ultrasound image data and preprocessing the thyroid ultrasound image data, the preprocessing comprising denoising, standardization and image enhancement; extracting feature information in the preprocessed thyroid image data, and performing feature extraction by using a convolutional neural network; obtaining gene data related to the thyroid disease, standardizing the gene data, removing redundant gene data, and selecting gene information related to the thyroid disease; and fusing the image data and the gene data. According to the method, the technical scheme of fusing the image data and the gene data is adopted, two types of data characteristics are effectively combined through mutual information maximization and weighted summation, and the technical effect of remarkably improving the disease risk prediction accuracy is achieved. According to the method, the defect that information of a single data source is incomplete is overcome, and a more comprehensive decision basis is provided.

Description

Technical Field

[0001] The present invention relates to the technical field of disease risk prediction, and in particular to a method for predicting the risk of thyroid disease. Background Art

[0002] Thyroid disease is a common endocrine disease worldwide, among which the diagnosis and risk assessment of thyroid nodules and thyroid cancer are particularly critical. Early detection of high-risk patients is of great significance for formulating treatment plans, optimizing resource allocation, and improving patient prognosis. However, the clinical manifestations of thyroid diseases are complex and diverse, especially in the early stages of the disease, when benign and malignant nodules may not have obvious differences in imaging and pathology, which poses a great challenge to diagnosis. In recent years, with the rapid development of medical imaging technology and genetic testing technology, more and more studies have attempted to combine the two types of data in order to improve the accuracy of diagnosis. However, the existing technical solutions still have many shortcomings in practical applications and it is difficult to fully meet clinical needs.

[0003] Existing risk prediction models are usually based on a single data source, such as ultrasound imaging data. Although imaging technology can provide epigenetic features of thyroid nodules, such as morphology, size, and echo, such methods ignore the molecular biological mechanisms of the disease, especially the potential role of gene mutation and expression. In clinical practice, the morphological features of many early thyroid cancers are similar to those of benign nodules, and image analysis alone is prone to misdiagnosis or missed diagnosis. In addition, some studies have attempted to use genetic data alone for risk assessment, but its ability to predict the disease is one-sided when it is not combined with imaging information, making it difficult to promote its use in clinical practice. Moreover, in practical applications, ultrasound imaging data is usually affected by equipment resolution, acquisition conditions, and individual differences among patients, and may contain a large amount of noise, which directly interferes with the extraction and analysis of image features. Existing technologies for processing image data mostly remain at the basic denoising or enhancement level, and fail to design more advanced processing methods based on the characteristics of medical images. At the same time, due to its high-dimensional characteristics and complexity, genetic data often contains a large number of redundant features, and directly inputting them into the model will increase the computational burden and reduce the efficiency of the model. However, the current technical solutions do not adequately handle the standardization, redundancy removal and feature selection of gene data, resulting in the failure to fully explore the key information hidden in high-dimensional data. Therefore, the present invention proposes a risk prediction method for thyroid disease to address the deficiencies of the prior art. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a risk prediction method for thyroid disease, aiming to solve the technical problems in the prior art of insufficient data utilization, insufficient prediction accuracy and lack of clinical guidance of the results.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for predicting the risk of thyroid disease, comprising the following steps:

[0006] Acquire thyroid ultrasound image data and preprocess them, including denoising, standardization and image enhancement;

[0007] Extract feature information from preprocessed thyroid imaging data and use convolutional neural network for feature extraction;

[0008] Obtain and standardize genetic data related to thyroid disease, remove redundant genetic data, and select genetic information related to thyroid disease;

[0009] The image data and the gene data are fused, wherein the fusion includes feature association based on maximization of mutual information and synthesizing fusion features by weighted summation;

[0010] Using a deep neural network to train the fusion features to obtain a thyroid disease risk prediction model;

[0011] According to the thyroid disease risk prediction model, risk prediction is performed on the samples to be predicted, and the prediction results are output.

[0012] Preferably, the preprocessing of the image data further includes removing noise using Gaussian filtering and median filtering.

[0013] Preferably, the preprocessing of the gene data includes Z-score standardization of the gene expression data and selecting genes associated with thyroid disease through genome-wide association studies.

[0014] Preferably, the fusion of the image data and the gene data includes calculating the mutual information between the image data and the gene data to maximize the amount of shared information between the image data and the gene data.

[0015] Preferably, the weighted summation of the fusion features is achieved by the following formula:

[0016] Z=w X ·X+w Y ·Y

[0017] Among them: Z is the fusion feature; X is the feature set of image data; Y is the feature set of gene data; w X and w Y is the weight of image data and gene data.

[0018] Preferably, the deep neural network includes multiple hidden layers, each hidden layer uses a ReLU activation function, and the output layer uses a Softmax or linear activation function depending on the task type.

[0019] Preferably, the training process uses the Adam optimizer, and the update rule is:

[0020]

[0021] Where: θ t is the network parameter of step t; and are the first moment estimate and the second moment estimate of the gradient respectively; η is the learning rate; ∈ is a constant to prevent division by zero.

[0022] Preferably, the model is evaluated using evaluation indicators such as accuracy, AUC value and F1-score.

[0023] Preferably, the risk prediction results are used to provide clinical decision support based on the output of the model, including providing personalized diagnosis and treatment recommendations for patients.

[0024] A risk prediction system for thyroid disease, comprising:

[0025] An image data acquisition module, used for acquiring thyroid ultrasound image data;

[0026] A gene data acquisition module, used to acquire gene data related to thyroid disease;

[0027] Data processing module, used to pre-process image data and gene data and extract features;

[0028] A fusion module is used to fuse the image data features with the gene data features;

[0029] The prediction module is used to train a deep neural network based on the fusion features to perform risk prediction;

[0030] Output module, used to output prediction results.

[0031] The present invention provides a method for predicting the risk of thyroid disease. It has the following beneficial effects:

[0032] 1. The present invention adopts a technical solution of fusing image data with gene data, effectively combining the characteristics of the two types of data by maximizing mutual information and weighted summation, and achieves a technical effect of significantly improving the accuracy of disease risk prediction. Compared with the prediction scheme in the prior art that only relies on a single dimension of image data or gene data, the present invention overcomes the deficiency of incomplete information from a single data source and provides a more comprehensive basis for decision-making.

[0033] 2. The present invention achieves high efficiency and standardization of data preprocessing by introducing image processing technologies such as Gaussian filtering and median filtering, as well as methods for standardizing gene data and removing redundant genes, and achieves the technical effect of optimizing data quality and improving the reliability of subsequent model input. Compared with the technical solutions in the prior art that do not fully process noise and redundant data and lead to reduced model performance, the present invention solves the problems of large data noise interference and insufficient feature extraction, and significantly improves the utilization rate of data.

[0034] 3. The present invention adopts a technical solution of training fusion features with a deep neural network, combined with multi-layer convolution and activation function optimization strategies, to achieve the technical effect of capturing complex feature relationships and improving model prediction performance. Compared with the existing solution of using shallow machine learning models that are difficult to handle nonlinear features, the present invention solves the problem of the traditional method's insufficient ability to model complex data, making risk prediction more accurate and reliable.

[0035] 4. The present invention achieves the technical effect of providing targeted decision support for clinicians through the visualization output of prediction results and the generation of personalized diagnosis and treatment recommendations. Compared with the existing solutions with single output results and lack of clinical practical value, the present invention solves the problem that prediction results are difficult to convert into specific diagnosis and treatment recommendations, and provides patients with more targeted treatment directions and risk management solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0037] Figure 2 It is a schematic diagram of the system architecture of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Please refer to the attached Figure 1 The present invention provides a method for predicting the risk of thyroid disease, comprising the following steps:

[0040] S1. Acquire thyroid ultrasound image data and preprocess them, including denoising, standardization and image enhancement;

[0041] Specifically, first of all, the acquired thyroid ultrasound image data is one of the important inputs for thyroid disease risk prediction. These image data are usually taken by ultrasound equipment and contain information such as the morphology, size, and echo characteristics of thyroid nodules. Ultrasound images can provide morphological characteristics of thyroid nodules, which are crucial for determining whether the nodules are benign or malignant. After acquiring these image data, the preprocessing stage is entered to ensure that the data meets the requirements of subsequent processing and analysis.

[0042] After acquiring the thyroid ultrasound image data, preprocessing steps such as denoising, standardization, and image enhancement are first performed to ensure the quality of the data and optimize the effect of subsequent feature extraction.

[0043] Denoising: There may be noise interference in the acquired ultrasound images, which affects the image quality and may cause errors in subsequent analysis. Therefore, the image needs to be denoised first. In this embodiment, two common denoising methods, Gaussian filtering and median filtering, are used. Gaussian filtering blurs the image by applying a Gaussian kernel function to reduce high-frequency noise. Specifically, the formula for Gaussian filtering is as follows:

[0044]

[0045] Where: I(x,y) is the pixel value of the original image; I ′ (x, y) is the pixel value after denoising; σ is the standard deviation, which controls the diffusion range of the Gaussian kernel and determines the smoothness of the filter; x, y are the coordinates of the current pixel; i, j are the coordinates of the pixel within the coverage range of the Gaussian kernel; 2πσ 2 is the normalization factor, which ensures that the sum of the weights of the Gaussian filter is 1 and avoids changes in image brightness; is the weight value of the Gaussian kernel, along with xi) 2 +(yj) 2 The farther away from the center point, the smaller the weight.

[0046] By choosing a suitable σ value, high-frequency noise in the image can be effectively removed. The median filter method removes salt and pepper noise by taking the median of the neighborhood pixel values, which is particularly common in medical images. Through these denoising methods, the quality of the image is significantly improved, ensuring the accuracy of subsequent analysis.

[0047] Standardization: The standardization process of image data is to ensure that the pixel values ​​of different images are in the same range to avoid the influence caused by different lighting and contrast of the images. In this embodiment, each pixel value of the image is normalized so that its range is converted from the original value to between 0 and 1. The standardized image pixel values ​​follow the same scale, ensuring the balance of each feature during the model training process.

[0048] Image enhancement: The purpose of image enhancement is to enhance the contrast of the image, especially the contrast between thyroid nodules and normal tissues, so that the boundary of the nodules is clearer. Common image enhancement techniques include histogram equalization, brightness adjustment, and contrast stretching. In this embodiment, histogram equalization is used to enhance the image contrast. Histogram equalization redistributes the frequency distribution of pixel values ​​to make the brightness range in the image wider, so that the edge information of the nodule can be displayed more clearly.

[0049] The calculation method of histogram equalization is to equalize the histogram of the original image so that the pixel values ​​of the image are more evenly distributed in the whole range, thereby improving the quality of the image. This method is particularly suitable for images with low contrast and can effectively highlight the key information in the image.

[0050] After preprocessing, the quality of the image data has been improved and feature extraction can be performed. Feature extraction is usually achieved through convolutional neural networks (CNN). CNN can automatically extract key features from images, such as the morphology, boundaries, and echo intensity of nodules. CNN gradually extracts local features of the image through multiple convolutional layers and pooling layers, and then synthesizes global features through fully connected layers to form a high-dimensional representation of the image.

[0051] Through the deep learning model, CNN can automatically identify key features from images without human intervention and use these features for subsequent risk prediction analysis. The advantage of CNN is that it can extract consistent features from different images and can process information with complex structure and diversity.

[0052] After the above preprocessing, the quality and accuracy of the image data are guaranteed, providing clearer input data for subsequent model training. Denoising effectively removes interference information in the image, standardization ensures data consistency, and image enhancement improves the recognizability of nodules. Feature extraction through CNN automatically obtains key information from the image, avoiding the subjective factors in traditional manual feature extraction methods, making subsequent risk prediction more accurate and efficient.

[0053] S2, extracting feature information from the preprocessed thyroid image data and using a convolutional neural network for feature extraction;

[0054] Specifically, in step S1, the thyroid ultrasound image data has been preprocessed, including denoising, standardization and image enhancement, which ensure the quality and clarity of the image data, eliminate the noise that affects the subsequent analysis, and enhance the visibility of thyroid nodules. The preprocessed image data provides a good basis for feature extraction. Next, enter the image feature extraction stage. The goal of this embodiment is to automatically extract meaningful features from the image through a convolutional neural network (CNN) to provide effective input for subsequent risk prediction.

[0055] The preprocessed image data is subjected to feature extraction through a convolutional neural network (CNN). CNN is a deep learning model that can automatically extract local features from images and gradually build complex feature representations in a multi-layer network structure. The basic building blocks of CNN include convolutional layers, pooling layers, activation functions, and fully connected layers.

[0056] Convolutional neural networks extract image features in a hierarchical manner. Each convolutional layer performs weighted processing on the local area of ​​the image through the convolution kernel to extract features at different levels. For example, the first convolutional layer may extract simple features in the image, such as edges and textures, and the second convolutional layer may extract more complex local morphological features. As the layers deepen, the network can extract high-level features of the image. Through multi-level convolution operations, CNN can gradually build a complete feature representation of the image from low-level features to high-level semantic information.

[0057] There are multiple convolutional layers used, including convolutional layers, pooling layers, and fully connected layers. Specifically, the role of the convolutional layer is to filter the input image through the convolution kernel to identify the local features of the image. The output of each convolution operation is usually processed by an activation function (such as ReLU), which can introduce nonlinear characteristics and improve the expression ability of the network. The pooling layer reduces the dimension of the data through dimensionality reduction operations while retaining key information, which helps to reduce the computational burden and enhance the robustness of the model.

[0058] The first convolutional layer of CNN uses a small-sized convolution kernel (such as 3×3 or 5×5) to perform convolution operations on the input image to generate local feature maps. The next pooling layer uses the maximum pooling operation (MaxPooling) to take the maximum value of each small area, reducing the spatial dimension while retaining the most important features. After that, the network continues to extract more advanced features through multiple convolution and pooling operations, and finally synthesizes the extracted features through the fully connected layer to generate a global representation of the image.

[0059] After the image is input into CNN, the goal of the network is to gradually learn the meaningful features in the image. The weight parameters of each convolution layer (i.e., convolution kernel) are continuously optimized through the back-propagation algorithm during the training process. In this way, after multiple layers of convolution and pooling operations, the network can extract key information about nodules in the image, including the morphology, boundaries, echo intensity and other features of the nodules. These features will be input into subsequent models to help predict disease risks.

[0060] Taking thyroid ultrasound images as an example, after the image preprocessing steps (denoising, standardization and enhancement), it is input into the convolutional neural network. The network first extracts local features such as the contour, size and morphology of the nodule in multiple convolutional layers, and then extracts complex structural information (such as whether the nodule is a malignant mass) at a deeper level. These extracted features can provide key information for subsequent risk prediction.

[0061] Automatically extracting image features through CNN avoids the subjectivity of traditional manual feature extraction methods, and the network can adaptively learn the multi-level features of the image, improving the efficiency and accuracy of feature extraction. Compared with traditional feature extraction methods, CNN can identify more complex image patterns, especially when faced with thyroid nodules with complex or irregular shapes, CNN shows stronger processing capabilities and robustness. This enables subsequent risk prediction models to obtain more accurate feature inputs, thereby improving the accuracy of thyroid disease prediction.

[0062] Through multi-level convolution operations and pooling processing, CNN can efficiently extract and synthesize key features from image data. These features can reflect the morphology, size, echo and other information of nodules, providing strong support for disease risk prediction. At the same time, due to the automation characteristics of CNN, the entire process does not require manual intervention, greatly improving processing efficiency and stability.

[0063] S3, obtaining and standardizing genetic data related to thyroid disease, removing redundant genetic data, and selecting genetic information related to thyroid disease;

[0064] Specifically, in step S2, the image data has been preprocessed and relevant features have been extracted through a convolutional neural network (CNN), which provide valuable information for subsequent risk prediction. The next step is to process genetic data related to thyroid disease. Genetic data plays an important role in the occurrence and development of diseases, especially the occurrence of diseases such as thyroid cancer is often closely related to mutations in certain genes. Therefore, in step S3, it is first necessary to obtain genetic data related to thyroid disease and perform appropriate preprocessing on it to ensure that these data can be effectively used for subsequent analysis.

[0065] After obtaining the genetic data related to thyroid disease, it is first standardized, then redundant genetic data is removed, and finally key genes related to the disease are selected. Through these processing steps, the quality and validity of the genetic data in subsequent analysis are ensured.

[0066] Genetic data generally include gene mutation information related to thyroid disease, such as mutations of BRAF and RAS genes. These mutations play an important role in the occurrence of thyroid cancer. In the present embodiment, after obtaining the genetic data, it is first necessary to perform a standardized process on it. The purpose of standardization is to eliminate the scale differences of different gene expression amounts, ensure that the data of different genes are compared within the same range, and avoid the influence of scale differences on subsequent analysis.

[0067] The Z-score standardization method can be used to process gene data. The Z-score standardization method subtracts the mean of each gene expression value and divides it by the standard deviation, so that the standardized data has zero mean and unit variance. The formula is as follows:

[0068]

[0069] Where: z i is the normalized gene expression value; x i is the original gene expression value; μ is the mean of the gene; σ is the standard deviation of the gene.

[0070] The standardized genetic data has a unified scale, which facilitates subsequent data analysis and model training.

[0071] Through Z-score standardization, the units of gene expression values ​​are changed to standardized units (standard deviation units), which allows different genes to be compared on the same scale and avoids the impact of expression differences on model training. Genetic data processed in this way provide standardized and comparable data for subsequent risk prediction.

[0072] The standardized gene data can directly enter the next step of feature selection. However, in practical applications, gene data often contains a lot of redundant information. In order to improve the effectiveness and computational efficiency of the model, redundant gene data must be removed. The process of removing redundant genes is usually achieved through correlation analysis. For genes with high correlation, one of them can be deleted, retaining the gene features with more information.

[0073] After performing correlation analysis on the gene data, if the correlation coefficient of two genes exceeds a certain threshold (for example, 0.9), it can be considered that the two genes have high redundancy and one of them is removed. This process can effectively reduce the data dimension and retain more key information.

[0074] The selection of genes can be based on existing medical literature or genome-wide association studies (GWAS). Through these studies, some genes that are closely related to the risk of thyroid disease have been identified. Common genes associated with thyroid cancer include BRAF, RAS, etc. For the selection of these genes, they can be selected according to their role and expression characteristics in thyroid disease, without having to screen all genes. In this way, the gene selection process is more efficient and targeted.

[0075] Through standardization, removal of redundant genes, and gene selection, the genetic data is effectively processed during the preprocessing process to ensure its quality and availability. Standardization eliminates the interference of differences in gene expression, allowing data to be compared on a uniform scale; removing redundant genes reduces the dimension of the data, reduces the complexity of subsequent models, and improves computational efficiency; gene selection further ensures the validity of relevant genetic data and avoids the influence of irrelevant information. After these processing steps, the genetic data can effectively provide clear and reliable input data for the risk prediction model of thyroid disease.

[0076] S4, fusing the image data and the gene data, wherein the fusion includes performing feature association based on maximizing mutual information and synthesizing fusion features by weighted summation;

[0077] Specifically, in step S3, the gene data has been standardized, redundancy removed, and relevant genes selected to ensure its effectiveness in subsequent risk prediction. The image data has been preprocessed and relevant features have been extracted through a convolutional neural network (CNN). Next, the core goal of step S4 is to fuse the image data and gene data to provide more comprehensive and effective feature input for subsequent risk prediction.

[0078] After the image features and gene features are extracted, the two types of feature data are first fused to maximize the shared information between them and reduce redundancy. To achieve this goal, this embodiment uses a mutual information maximization method and a weighted summation method to effectively combine the two types of features.

[0079] Image data and gene data provide information for the risk prediction model from different dimensions. Image data provides visual information such as the morphology, size, and echo characteristics of thyroid nodules, while gene data reveals potential genetic risk factors. The information from these two data sources complement each other, and the combination can provide a more complete disease risk assessment. Therefore, how to effectively integrate these two types of information is a key link in the present invention.

[0080] First, the correlation between image data and gene data is evaluated by calculating the mutual information between them. Mutual information is an information theory metric that measures the key information shared between two random variables. Specifically, the mutual information I(X,Y) of image data X and gene data Y is calculated by the following formula:

[0081] I(X,Y)=H(X)+H(Y)-H(X,Y)

[0082] Where: I(X,Y) is the mutual information between image data and gene data; H(X) and H(Y) are the entropy (amount of information) of image data and gene data respectively; H(X,Y) is their joint entropy.

[0083] By maximizing mutual information, it is possible to ensure that the shared information between image data and genetic data is preserved while avoiding interference from redundant information.

[0084] The operation of maximizing mutual information can effectively identify the useful information contained in both imaging data and genetic data, and remove the noise and irrelevant parts. In this way, the model can focus more on the effective connection between the two types of data when fusing them, rather than duplicate information. Through such a data fusion process, the sensitivity and accuracy of the risk prediction model to disease risk can be improved.

[0085] After calculating mutual information and optimizing feature fusion, the features of image data and gene data are combined using a weighted summation method. The weighted summation method assigns different weights w to image features and gene features according to their importance in risk prediction. X and w Y , and finally form the fusion feature Z, which is expressed as follows:

[0086] Z=w X ·X+w Y ·Y

[0087] Among them: Z represents the fused features; X is the feature set of image data processed by CNN; Y is the feature set of gene data after standardization; w X and w Y is the weight of image data and gene data.

[0088] Weight coefficient w X and w Y During the model training process, dynamic adjustments will be made based on the characteristics of the data set, so that the influence of imaging data and genetic data is reasonably balanced. This adjustment process enables the fusion features to better reflect the relative contributions of different data sources in predicting thyroid disease risk.

[0089] By maximizing mutual information and weighted summation, image data and gene data are effectively fused, and valuable information is retained to the maximum extent in the process. Maximizing mutual information ensures the optimization of shared information between the two types of data, while weighted summation ensures the balance and accuracy of the fused features. This data fusion method not only effectively reduces the interference of redundant information, but also further enhances the generalization ability of the prediction model.

[0090] Through this method, the fused feature Z is input into the deep neural network for further training and optimization, and finally a model for thyroid disease risk prediction is generated. This feature fusion method enables the model to not only comprehensively consider the visual features of the image data, but also combine the genetic risk of the genetic data, comprehensively improving the accuracy of risk prediction.

[0091] S5. Using a deep neural network to train the fusion features to obtain a thyroid disease risk prediction model;

[0092] Specifically, in the above steps, the imaging data and gene data have been preprocessed, feature extracted and fused to generate feature inputs that can be used for risk prediction. These feature inputs provide key information on thyroid disease risk to a certain extent and are ready for model training. Next, the key task of step S5 is to use these fused features to train a deep neural network (DNN) to establish a thyroid disease risk prediction model.

[0093] The feature data fused in step S4 is trained through a deep neural network (DNN). The training process is designed to enable the model to learn potential disease risk patterns from the data. A deep neural network is a typical nonlinear model that extracts complex features from the data layer by layer through multiple hidden layers, and finally generates a prediction of the risk through the output layer. The neurons and weights of each layer are gradually optimized through training, so that the model can adapt to different data characteristics and provide more accurate risk predictions.

[0094] The structure of a deep neural network includes multiple hidden layers, and each hidden layer uses the ReLU activation function. ReLU (Rect ifi ed Li near Un it) is a commonly used activation function with the advantages of simple calculation and fast convergence speed. In each layer, the input data will be matrix multiplied with the weight matrix of the current layer to produce a weighted sum, which is then activated by the ReLU function to obtain the output of the layer. The formula of the ReLU function is as follows:

[0095] f(x)=max(0,x)

[0096] Where: x is the weighted sum of the input of this layer; f(x) is the output result.

[0097] The advantage of ReLU is that it can introduce nonlinear factors, allowing neural networks to learn complex patterns.

[0098] The output layer of the network is designed according to the task type. If the prediction task is a classification task, the output layer usually uses a Softmax activation function to generate a probability value for each category; if the task is a regression problem, a linear activation function is used to output a continuous risk score.

[0099] During training, the fusion feature input is passed through the neural network layer by layer, and weighted calculations are performed between neurons in each layer. Each layer of the network calculates the result through forward propagation, and uses the backpropagation algorithm to adjust the weight according to the error between the actual output and the predicted output. The specific error calculation is usually implemented through a loss function, and common loss functions include the cross entropy loss function and the mean square error loss function.

[0100] If it is a classification task, the loss function is the cross entropy loss function, and the formula is as follows:

[0101]

[0102] Where: y i is the true label; is the probability value of the label predicted by the model; N is the total number of samples.

[0103] The cross entropy loss function is used to measure the difference between the predicted value and the true value. A smaller cross entropy value indicates that the model's prediction is more accurate.

[0104] If it is a regression task, the mean square error (MSE) loss function is used, and the formula is:

[0105]

[0106] Where: y i is the actual value; is the predicted value; N is the number of samples.

[0107] The mean square error loss function calculates the average difference between the predicted value and the actual value. A smaller mean square error indicates a more accurate prediction model.

[0108] During the training process, each weight value of the neural network is updated through an optimization algorithm. The goal of the optimization algorithm is to minimize the loss function, that is, to gradually adjust the weights so that the network's prediction results are closer and closer to the actual results. The commonly used optimization algorithm is the Adam optimizer, which combines the gradient descent method with adaptive learning rate adjustment to accelerate convergence and improve the stability of the model during training. The update rules of the Adam optimizer are as follows:

[0109]

[0110] Where: θ t is the network parameter of step t; and are the first moment estimate and the second moment estimate of the gradient respectively; η is the learning rate; ∈ is a constant to prevent division by zero.

[0111] The Adam optimizer can adaptively adjust the learning rate during the training process, making the network weight update smoother and thus improving the training efficiency.

[0112] Throughout the training process, the purpose of weight updating is to minimize the loss function so that the network can gradually reduce the prediction error on each training sample. After multiple iterations, the network eventually learns the optimal parameters for thyroid disease risk prediction, allowing the model to make accurate predictions when faced with unseen data.

[0113] Through deep neural network training, the model can gradually learn the rules of thyroid disease risk prediction from the input features. Because deep neural networks have powerful nonlinear mapping capabilities, they can process complex input data and automatically extract valuable information from image features and gene features. This process greatly reduces the need for manual feature engineering and can demonstrate strong predictive capabilities on complex data sets.

[0114] S6. Perform risk prediction on the samples to be predicted based on the thyroid disease risk prediction model and output the prediction results.

[0115] Specifically, in the above steps, the fused imaging data and genetic data were trained through a deep neural network to establish a risk prediction model for thyroid disease. At this point, the model has learned to identify potential disease risk patterns from the input data. Next, the task of step S6 is to evaluate the trained model and use it to predict the risk of new patient data.

[0116] Model evaluation and risk prediction are key steps to ensure the effectiveness and accuracy of the established model. First, the performance of the model is evaluated, usually by using multiple evaluation indicators to comprehensively measure the model's prediction effect. Second, after the evaluation is completed, the model is used to predict new data and provide corresponding risk scores or risk classification results.

[0117] The model is evaluated through the following main indicators: accuracy, AUC (area under the curve) value and F1-score. Each indicator has its applicable scenario and can reflect the performance of the model from different dimensions. Specifically, accuracy measures the proportion of the model's correct classification of all samples, AUC value reflects the model's ability to distinguish between positive and negative samples, and F1-score combines the model's precision and recall rate, which is especially suitable for cases of class imbalance.

[0118] Use cross-validation to evaluate the stability and generalization ability of the model. By dividing the data set into multiple subsets, each subset is used as a test set, and the remaining subsets are used as training sets, and multiple rounds of training and evaluation are performed. Cross-validation can help eliminate the bias caused by a single data partition, making model evaluation more accurate and reliable. The average result of cross-validation can reflect the performance of the model in a variety of scenarios.

[0119] After the model training is completed, the model is first evaluated using the test set. The test set refers to the data set that was not involved in the model training during the training process and can be used to measure the performance of the model on unseen data. The evaluation results include but are not limited to accuracy, AUC, and F1-score, which can comprehensively measure the predictive ability of the model. For example, the closer the AUC value is to 1, the stronger the classification ability of the model; and the F1-score value provides a balance between precision and recall, which is particularly important when dealing with imbalanced data.

[0120] If the model is a classification problem (for example, predicting whether a thyroid nodule is malignant), the cross entropy loss function is used to calculate the error between the predicted value and the actual value. The cross entropy loss function formula is as follows:

[0121]

[0122] Where: y i is the true label; is the probability value of the label predicted by the model; N is the total number of samples.

[0123] By minimizing the cross entropy loss function, the model is able to improve its prediction accuracy for classification tasks.

[0124] If the model is a regression problem (for example, predicting the risk score for thyroid disease), the mean square error (MSE) loss function is used to measure the difference between the model's predicted value and the actual value. The formula for MSE is:

[0125]

[0126] Where: y i is the actual value; is the predicted value; N is the number of samples.

[0127] The MSE loss function helps the model minimize the prediction error and improve the accuracy of continuous numerical predictions.

[0128] Model evaluation is not limited to the above indicators. The performance of the classification model can be further evaluated through the confusion matrix. The confusion matrix can show the classification of the prediction results in detail, including true positives, false positives, true negatives, and false negatives, thus helping to further analyze the performance of the model on different categories.

[0129] When using the confusion matrix for evaluation, you can calculate indicators such as precision, recall, and F1-score. Precision reflects the proportion of samples predicted to be positive that are actually positive; recall measures the proportion of samples that are actually positive that are correctly predicted by the model. F1-score, as the harmonic average of precision and recall, provides a more comprehensive evaluation result.

[0130] After the evaluation is completed, the trained model is used to predict the risk of new sample data. The imaging data and genetic data of the new sample are input, and the model will output a risk score or category label. For example, the model outputs a numerical score indicating the risk of malignancy of a thyroid nodule, or outputs a classification result indicating whether the nodule is malignant or benign. Based on the risk score or category output by the model, doctors can make clinical decisions and further determine whether further testing or treatment is needed.

[0131] By using these evaluation indicators, the performance of the model can be comprehensively and objectively evaluated. Especially when faced with complex data, model evaluation not only helps ensure prediction accuracy, but also guides subsequent model optimization and improves the stability and generalization ability of the model. After the model evaluation is completed, based on the evaluation results, doctors can make timely and effective diagnosis and treatment decisions through risk prediction results, further improving the accuracy of early diagnosis of thyroid diseases.

[0132] A risk prediction system for thyroid disease described below and a risk prediction method for thyroid disease described above can be referenced to each other.

[0133] Please refer to the attached Figure 2 , a risk prediction system for thyroid disease, comprising:

[0134] An image data acquisition module, used for acquiring thyroid ultrasound image data;

[0135] A gene data acquisition module, used to acquire gene data related to thyroid disease;

[0136] Data processing module, used to pre-process image data and gene data and extract features;

[0137] A fusion module is used to fuse the image data features with the gene data features;

[0138] The prediction module is used to train a deep neural network based on the fusion features to perform risk prediction;

[0139] Output module, used to output prediction results.

[0140] Specifically,

[0141] Image data acquisition module: used to acquire thyroid ultrasound image data, and can support data format input of different types of ultrasound equipment, such as B-ultrasound, color ultrasound, etc. Extended design: The module further integrates real-time data acquisition function, can be connected to ultrasound equipment, and receive real-time image data through DI COM standard interface. It also supports the import function of historical image data for long-term monitoring of patients' thyroid health status.

[0142] Gene data acquisition module: The module supports multi-source gene data input, including gene reports provided by external testing agencies or data generated by high-throughput sequencing platforms. It supports parsing of standardized data formats (such as VCF or FASTQ formats) and can perform personalized data annotation in combination with the patient's clinical background information (such as age and gender).

[0143] Data processing module: Image data processing: The module supports a variety of image enhancement technologies, including contrast enhancement, edge enhancement, etc., and integrates noise processing algorithms (such as Gaussian filtering and median filtering). Gene data processing: The module supports the standardization of gene data, the removal of redundant features, and the comparison analysis with public databases (such as C linVar, dbSNP) to extract important gene mutation information. Abnormal data detection: The abnormal data detection function is introduced to mark abnormal values ​​or missing data, and trigger a system alarm to prompt the user to check the data quality.

[0144] Fusion module: The module supports multimodal feature fusion algorithms, including mutual information maximization and weight allocation strategy based on attention mechanism. The fused features can be output in a unified format for direct call by deep neural network models. It provides feature importance analysis function to explain the contribution of each feature in the final prediction and improve system transparency.

[0145] Prediction module: The system has built-in multiple deep neural network architectures (such as CNN and Transformer), which can be flexibly selected according to data characteristics. The module supports both online and offline training modes. The online mode can update model parameters in real time, and the offline mode is used for batch training and verification. The integrated transfer learning function can use the existing medical imaging big data model for fine-tuning, reducing the demand for training data.

[0146] Output module: The output module supports multiple forms of prediction result display, including graphical risk scores, classification results (such as benign / malignant nodules) and detailed probability distribution. The module supports the generation of personalized diagnosis and treatment recommendation reports, including prediction results, key feature analysis, and recommended clinical examinations or treatment plans. Provides a data interface to support the synchronization of prediction results to the hospital's electronic medical record system (EMR) to achieve seamless connection of clinical data.

[0147] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the risk of thyroid disease, characterized in that: The following steps are involved: Acquire thyroid ultrasound image data and preprocess them, including denoising, standardization and image enhancement; Extract feature information from preprocessed thyroid imaging data and use convolutional neural network for feature extraction; Obtain and standardize genetic data related to thyroid disease, remove redundant genetic data, and select genetic information related to thyroid disease; The image data and the gene data are fused, wherein the fusion includes feature association based on maximization of mutual information and synthesizing fusion features by weighted summation; Using a deep neural network to train the fusion features to obtain a thyroid disease risk prediction model; According to the thyroid disease risk prediction model, risk prediction is performed on the samples to be predicted, and the prediction results are output.

2. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The preprocessing of the image data further includes removing noise using Gaussian filtering and median filtering.

3. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The preprocessing of the gene data includes Z-score standardization of the gene expression data and selecting genes related to thyroid diseases through genome-wide association studies.

4. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The fusion of the image data and the gene data includes calculating the mutual information between the image data and the gene data, and maximizing the amount of shared information between the image data and the gene data.

5. A method for predicting the risk of thyroid disease according to claim 4, characterized in that: The weighted summation of the fusion features is achieved by the following formula: Z=w X ·X+w Y ·Y Among them: Z is the fusion feature; X is the feature set of image data; Y is the feature set of gene data; w X and w Y is the weight of image data and gene data.

6. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The deep neural network includes multiple hidden layers, each hidden layer uses a ReLU activation function, and the output layer uses a Softmax or linear activation function depending on the task type.

7. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The training process uses the Adam optimizer, and the update rule is: Where: θ t is the network parameter of step t; and are the first moment estimate and the second moment estimate of the gradient respectively; η is the learning rate; ∈ is a constant to prevent division by zero.

8. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The model is evaluated using accuracy, AUC value, and F1-score.

9. A method for predicting the risk of thyroid disease according to claim 1, characterized in that: The risk prediction results are used to provide clinical decision support based on the output of the model, including providing personalized diagnosis and treatment recommendations for patients.

10. A risk prediction system for thyroid disease, applied to a risk prediction method for thyroid disease as claimed in any one of claims 1 to 9, characterized in that: include: An image data acquisition module, used for acquiring thyroid ultrasound image data; A gene data acquisition module, used to acquire gene data related to thyroid disease; Data processing module, used to pre-process image data and gene data and extract features; A fusion module is used to fuse the image data features with the gene data features; The prediction module is used to train a deep neural network based on the fusion features to perform risk prediction; Output module, used to output prediction results.

Images