Artificial intelligence image analysis-based hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system

Abstract

The invention discloses a hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis, and the system receives thyroid ultrasound and CT image data through a Web interface, and carries out the preprocessing through a data preprocessing module, and then obtains a thyroid tumor image through segmentation; the feature extraction and selection module is used for extracting high-throughput radiomics features from the images and further screening out core features to construct radiomics tags; the risk prediction module is used for training a constructed prediction model based on radiomics tags, conventional clinical features, tumor image qualitative features and corresponding classification tags, analyzing key features influencing a prediction result through SHAP, and predicting new data by using the trained prediction model; the risk output module converts the prediction result into a chart for visual output; and the user interface module receives the information output by the risk output module and displays the information to a user in real time through a Web interface. According to the method, the multi-mode image data and the clinical data are integrated, the recurrence risk prediction model driven by hypoxia is constructed, and accurate prediction of the recurrence risk of the PTC patient is achieved.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical image analysis and artificial intelligence, and particularly to a hypoxia-driven recurrence risk prediction system for papillary thyroid cancer based on artificial intelligence image analysis. Background Art

[0002] Papillary thyroid cancer (PTC) is the most common thyroid malignancy. Although the prognosis of PTC is relatively good, some patients still have a recurrence risk, which affects the long-term quality of life of patients. Early prediction of the recurrence risk is of great significance for the formulation of personalized treatment and follow-up strategies. Hypoxia is an important feature of the tumor microenvironment, which significantly promotes the proliferation, invasion and metastasis ability of PTC. However, the detection of PTC hypoxia usually requires invasive methods such as immunohistochemistry, which limits its clinical application. Traditional image analysis methods are difficult to extract deep features related to hypoxia drive from complex medical images. In recent years, the rapid development of artificial intelligence image analysis technology has provided new technical means for extracting high-throughput features from massive image data. Radiomics can reveal the heterogeneity and biological characteristics of tumors by extracting a large number of quantitative features from medical images and combining machine learning algorithms, providing new ideas for recurrence risk prediction.

[0003] However, the existing artificial intelligence image analysis technology still faces the following technical challenges in the prediction of PTC recurrence risk: (1) Limitations in feature extraction: Traditional radiomics methods mainly rely on manually delineating lesions to obtain features, with low segmentation accuracy and efficiency. (2) Lack of deep biological association: Existing methods usually directly predict the recurrence risk without establishing a deep association with hypoxia drive, resulting in insufficient biological interpretability of the model. (3) Single-modal data limitation: Existing systems mostly rely on single-modal data (such as ultrasound or CT), failing to effectively integrate multi-modal image data and clinical data, which limits the prediction performance of the model. (4) Insufficient system interpretability: The prediction results lack interpretability, making it difficult to identify key prediction factors and their contribution degrees, reducing the practicality and credibility of the system. (5) Insufficient clinical usability: Existing systems mostly lack a user-friendly interaction interface, making it difficult to meet the needs of clinical real-time analysis and decision-making. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a hypoxia-driven recurrence risk prediction system for papillary thyroid cancer based on artificial intelligence image analysis. By integrating multi-modal image data (ultrasound and contrast-enhanced CT) and clinical data, a recurrence risk prediction model driven by hypoxia is constructed to achieve precise stratification of the recurrence risk of PTC patients.

[0005] Technical solution: To achieve the above object, the present invention relates to a hypoxia-driven recurrence risk prediction system for papillary thyroid carcinoma based on artificial intelligence image analysis. The system receives thyroid ultrasound and CT image data, as well as corresponding conventional clinical feature data and tumor image qualitative feature data through a Web interface;

[0006] The data preprocessing module on the server side preprocesses the thyroid ultrasound and CT image data and segments the thyroid tumor images;

[0007] The feature extraction and selection module extracts high-throughput radiomics features from the thyroid tumor images and uses the trained lasso regression to screen out the core features related to hypoxia-driven recurrence from the high-throughput radiomics features to construct a radiomics signature;

[0008] The risk prediction module trains the constructed hypoxia-driven papillary thyroid carcinoma recurrence risk prediction model based on the radiomics signature, conventional clinical features, tumor image qualitative features, and corresponding classification labels, and analyzes the key features affecting the prediction results through SHAP. Finally, the trained prediction model is used to predict new data;

[0009] The risk output module visualizes the prediction results into charts for output; the user interface module receives the information output by the risk output module and displays it to the user in real time through the Web interface.

[0010] Among them, the conventional clinical features include the patient's age, gender, whether there is thyroid function abnormality, whether there is nodular goiter, and whether there is Hashimoto's thyroiditis;

[0011] The tumor image qualitative features include tumor size, location, shape, internal components, echo features, degree of contact with the capsule, presence of calcification, and presence of cystic change.

[0012] Among them, the data preprocessing module preprocesses the thyroid ultrasound and CT image data including format conversion, image resampling, and intensity normalization processing; among them, the resampling process is:

[0013] I resampled (x,y,z) = Interpolate(I original ,T resample ),

[0014] In the formula, I original is the original image, T resample is the resampling transformation matrix, Interpolate(*) is the interpolation function, and I resampled (x,y,z) is the pixel value of the resampled image at the point (x,y,z);

[0015] The normalization process is:

[0016]

[0017] Wherein, I min and I max are respectively the minimum and maximum pixel values of the resampled image, and I normalized (x, y, z) is the pixel value of the normalized image at the point (x, y, z).

[0018] Among them, the method for the data preprocessing module to segment the thyroid tumor image is as follows: the pre-trained nnU-Net network model is used to automatically segment the thyroid tumor area from the preprocessed thyroid ultrasound and CT images to obtain the segmented thyroid tumor image; the training data includes obtaining thyroid ultrasound and CT images and the corresponding tumor annotation data, and the preprocessed thyroid ultrasound and CT images and the corresponding tumor annotation data are input into the nnU-Net network model for training to obtain the trained nnU-Net network model.

[0019] Among them, the nnU-Net network model adopts a four-layer encoding and four-layer decoding structure, uses a 3×3×3 convolution kernel for feature extraction and reconstruction, the encoder gradually extracts high-level features through downsampling, and the decoder gradually restores the spatial resolution through upsampling;

[0020] Among them, the feature map of the l-th layer of the encoder is calculated as:

[0021]

[0022] Among them, and are respectively the convolution kernel weight and bias of the l-th layer, σ is the activation function, and * is the convolution operation;

[0023] The feature map of the l-th layer of the decoder is calculated as:

[0024]

[0025] Among them, Upsample is the upsampling operation, and are respectively the convolution kernel weight and bias of the l-th layer, σ is the activation function, and * is the convolution operation;

[0026] When training the nnU-Net network model, first define a Dice loss function to measure the prediction error of the network model, and select an optimizer to update the weights of the network model according to the gradient of the loss function. Through multiple iterations of training, the optimizer continuously adjusts the weights to make the value of the loss function gradually decrease. The formula of the Dice loss function is:

[0027]

[0028] In the formula, p i is the predicted value, and g i is the true value;

[0029] The initial learning rate of the network model is 1e-4, and a cyclic learning rate scheduler is used to dynamically adjust the learning rate. The formula of the cyclic learning rate scheduler is:

[0030]

[0031] In the formula, LR min and LR max are the minimum and maximum learning rates respectively, t is the current training step, and T is the total number of steps.

[0032] Among them, the feature extraction and selection module extracts high-throughput radiomics features from thyroid tumor images, including feature extraction and feature transformation. Among them, feature extraction extracts three types of radiomics features of thyroid tumor images through the PyRadiomics database: shape features, first-order features, and texture features. Shape features include tumor volume, surface area, and surface volume ratio. First-order features include mean, variance, skewness, kurtosis, and percentile. Texture features include gray-level co-occurrence matrix analysis of microstructure arrangement, gray-level dependence matrix detection of local heterogeneity, gray-level run-length matrix evaluation of texture directionality, and gray-level zone size matrix quantification of lesion spatial distribution;

[0033] Feature transformation performs the following transformations on the extracted original features in sequence: exponential transformation, gradient transformation, logarithmic transformation, square transformation, square root transformation, log-σ transformation, and wavelet transformation.

[0034] Among them, the lasso regression training data in the feature extraction and selection module includes a set of high-throughput radiomics features and corresponding classification labels. The classification label is the hypoxia-driven recurrence risk grade confirmed by pathology and clinical follow-up. In the training process, by adding an L1 regularization penalty term to the loss function, the weak correlation feature coefficients in the high-throughput radiomics features are compressed to zero, and the core features closely related to hypoxia-driven recurrence are screened out.

[0035] Among them, the prediction model for the recurrence risk of oxygen-driven papillary thyroid cancer in the risk prediction module includes a random forest-based model. During the training process, the recursive feature elimination algorithm is used to recursively remove the least important features from the input features, and then the remaining training set is used to train the model. The training process of the prediction model is as follows: The recursive feature elimination algorithm is used to recursively screen features, including training a random forest model with all current features. The random forest model uses the bootstrap sampling method to draw a sample subset from the training set with replacement to construct decision trees. When each decision tree grows, during the node splitting process, from a randomly selected subset of features, the optimal splitting point is selected based on the minimum Gini impurity criterion. The average reduction in Gini impurity caused by each feature in all decision trees is calculated as the feature importance score. According to the feature importance score, the least important features are removed, and this process is repeated until the number of features reaches a preset value. The prediction results of all decision trees are integrated by the majority voting method to obtain the final prediction, including the final class prediction and the related probability.

[0036] Among them, the prediction model optimizes the hyperparameters of the random forest model through grid search and 5-fold cross-validation to find an optimal combination of hyperparameters.

[0037] Among them, the key prediction factors of the prediction model are analyzed using game theory-based SHAP. By calculating the Shapley value to quantify the contribution degree of each feature used for making predictions to the prediction result of the prediction model, the interpretability of the prediction model is enhanced.

[0038] Beneficial effects: The present invention has the following advantages:

[0039] 1. By automatically segmenting thyroid ultrasound and CT image data, the limitations of manually delineating lesions are overcome, and the accuracy and efficiency of segmentation are improved;

[0040] 2. By screening the core features closely related to hypoxia-driven recurrence, an imaging genomics signature is established, thus linking recurrence risk prediction with the deep biological mechanism of hypoxia drive and improving the biological interpretability of the prediction results;

[0041] 3. By integrating multi-modal image data such as thyroid ultrasound and CT and clinical data, the present invention can more comprehensively capture information related to recurrence risk, thereby improving the prediction performance;

[0042] 4. By calculating the Shapley value to quantify the contribution degree of each feature used for making predictions to the prediction result of the prediction model, the practicality and credibility of the system are enhanced;

[0043] 5. By providing a user-friendly Web interface, users can obtain recurrence risk prediction results in real time, meeting the needs of clinical real-time analysis and decision-making. Brief Description of the Drawings

[0044] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0045] Figure 2 It is a schematic diagram of the segmentation of thyroid ultrasound and CT images by the nnU-Net network;

[0046] Figure 3 It is a schematic diagram of the SHAP interpretable analysis of the present invention;

[0047] Figure 4 It is a schematic diagram of the Web interactive interface of the present invention. Detailed Embodiment

[0048] The technical solution of the present invention will be described in detail below in conjunction with the embodiments and the drawings.

[0049] As Figure 1 shown, the present invention provides a hypoxia-driven thyroid papillary carcinoma recurrence risk prediction system based on artificial intelligence image analysis, including a data preprocessing module, a feature extraction and selection module, a risk prediction module, a risk output module, and a user interface module.

[0050] Among them, the user inputs thyroid ultrasound and CT image data, as well as corresponding conventional clinical feature data and tumor image qualitative feature data through the Web interface; the data preprocessing module on the server side preprocesses the thyroid ultrasound and CT image data and segments the thyroid tumor images; the feature extraction and selection module extracts high-throughput radiomics features from the thyroid tumor images and uses the trained lasso regression to screen out the core features related to hypoxia-driven recurrence from the high-throughput radiomics features to construct a radiomics label; the risk prediction module trains the constructed hypoxia-driven thyroid papillary carcinoma recurrence risk prediction model based on the input conventional clinical features, tumor image qualitative features, radiomics label, and corresponding classification labels, and analyzes the key features affecting the prediction results through SHAP, and finally uses the trained prediction model to predict new data; the risk output module visualizes the prediction results into charts for output; the user interface module receives the information output by the risk output module and displays it to the user in real time through the Web interface.

[0051] The implementation process of the above modules is as follows:

[0052] 1. The thyroid ultrasound and CT image data obtained by the data preprocessing module support common medical image formats such as DICOM and NIFTI. If it is in other formats such as DICOM, the image data can be further converted into the NIFTY format for subsequent processing.

[0053] The preprocessing includes the steps:

[0054] (1) Resample the thyroid ultrasound and CT image data with a resampling resolution of 1.0 mm and isotropic. The resampling process is as follows:

[0055] I resampled (x,y,z) = Interpolate(I original ,T resample ),

[0056] where I original is the original image, T resample is the resampling transformation matrix, Interpolate(*) is the interpolation function, and L resampled (x,y,z) is the pixel value of the resampled image at the point (x,y,z).

[0057] (2) Normalize the intensity of the resampled image by scaling the pixel values of the resampled image to the range [0,1] to eliminate image intensity differences. The normalization process is as follows:

[0058]

[0059] where I min and I max are the minimum and maximum pixel values of the resampled image respectively, and I normalized (x,y,z) is the pixel value of the normalized image at the point (x,y,z).

[0060] (3) Use the pre-trained nnU-Net network model to automatically segment the thyroid tumor region from the preprocessed thyroid ultrasound and CT images, and obtain the segmented thyroid tumor image. As Figure 2 shown, the input of the model is the preprocessed original thyroid image (showing the complete thyroid tissue structure and surrounding anatomical relationships), and the output is the corresponding thyroid tumor segmentation mask (the green area represents the automatically segmented tumor). The training data includes obtaining 1000 cases of thyroid ultrasound and CT images and the corresponding tumor annotation data. Perform the above preprocessing on the thyroid ultrasound and CT images, input the preprocessed images and the corresponding annotation data into the nnU-Net network model, and train the model to obtain the trained nnU-Net network model.

[0061] As Figure 2 shown, the nnU-Net model adopts a four-layer encoding and four-layer decoding structure, and uses a 3×3×3 convolutional kernel for feature extraction and reconstruction. The encoder part gradually extracts high-level features through downsampling, and the decoder part gradually restores the spatial resolution through upsampling.

[0062] The feature map of the l-th layer of the encoder is calculated as:

[0063]

[0064] Among them, and are the convolution kernel weights and biases of the l-th layer respectively, l is the activation function, and * is the convolution operation.

[0065] The feature map of the l-th layer of the decoder is calculated as

[0066]

[0067] Among them, Upsample is the upsampling operation, and are the convolution kernel weights and biases of the l-th layer respectively, σ is the activation function, and * is the convolution operation.

[0068] When training the nnU-Net network model, first define a Dice loss function to measure the prediction error of the network model, and select an optimizer (such as Adam) to update the weights of the network model according to the gradient of the loss function. Through multiple iterations of training, the optimizer continuously adjusts the weights, making the value of the loss function gradually decrease, thereby improving the prediction performance of the model. The formula of the Dice loss function is as follows:

[0069]

[0070] Among them, p i is the predicted value, and g i is the true value.

[0071] The initial learning rate of the model is 1e-4, and a cyclic learning rate scheduler is used to dynamically adjust the learning rate to reduce overfitting and enhance the generalization ability of the model. The formula of the cyclic learning rate scheduler is as follows:

[0072]

[0073] Among them, LR min and LR max are the minimum and maximum learning rates respectively, t is the current training step, and T is the total number of steps.

[0074] II. The feature extraction and selection module includes functions of feature extraction, feature transformation, and feature selection.

[0075] (1) Feature extraction: Three types of radiomics features of thyroid tumor images are extracted through the PyRadiomics database: shape features (such as tumor volume, surface area, surface volume ratio, etc.), which quantitatively characterize the morphological heterogeneity of tumors; first-order features (such as mean, variance, skewness, kurtosis, percentiles, etc.), which reflect the pixel intensity distribution characteristics of the tumor region; texture features, including gray-level co-occurrence matrix (GLCM) to analyze the microstructure arrangement, gray-level dependence matrix (GLDM) to detect local heterogeneity, gray-level run-length matrix (GLRLM) to evaluate texture directionality, and gray-level size zone matrix (GLSZM) to quantify the spatial distribution of lesions.

[0076] (2) Feature transformation: The following transformations are successively performed on the extracted original features: exponential transformation, gradient transformation, logarithmic transformation, square transformation, square root transformation, log-σ transformation, and wavelet transformation to enhance the feature expression ability.

[0077] Finally, 954 radiomics features are obtained from two-dimensional ultrasound images, and 1389 radiomics features are obtained from three-dimensional CT images.

[0078] (3) Feature selection: The trained Lasso Regression is used to screen the high-throughput radiomics features, and the core features closely related to hypoxia-driven recurrence are selected to construct a radiomics signature. The training data includes a set of high-throughput radiomics features and corresponding classification labels, and the classification labels are hypoxia-driven recurrence risk grades (high risk = 1, low risk = 0) confirmed by pathology and clinical follow-up. In the training process, by adding an L1 regularization penalty term to the loss function, the weak correlation feature coefficients in the high-throughput radiomics features are compressed to zero, and the core features closely related to hypoxia-driven recurrence are selected.

[0079] III. The risk prediction module includes a prediction model for the recurrence risk of hypoxia-driven papillary thyroid carcinoma. This prediction model is trained based on conventional clinical features, tumor imaging qualitative features, features within the radiomics signature, and corresponding classification labels (hypoxia-driven recurrence risk grades confirmed by pathology (high risk = 1, low risk = 0)), and the key prediction factors are explained through SHAP analysis. The trained prediction model can comprehensively analyze the conventional clinical features, tumor imaging qualitative features, and features within the radiomics signature of new papillary thyroid carcinoma patients to predict the risk probability of their occurrence of hypoxia-driven recurrence.

[0080] Among them, the conventional clinical features include: the patient's age, gender, whether there is thyroid function abnormality, whether there is nodular goiter, and whether there is Hashimoto's thyroiditis. The tumor imaging qualitative features include: tumor size, location, shape, internal components, echo characteristics, degree of contact with the capsule, whether there is calcification, and whether there is cystic change.

[0081] The prediction model for the recurrence risk of hypoxic-driven papillary thyroid carcinoma includes a random forest-based model. In each training of the model, the recursive feature elimination algorithm is used to recursively remove the least important features from the input features, and then the remaining training set is used to train the model to establish a model that can accurately predict new input data.

[0082] The training process of the prediction model is as follows: The recursive feature elimination algorithm (RFE) is used to recursively screen features, including training a random forest model with all current features. The random forest model uses the Bootstrap method to draw a sample subset from the training set with replacement, constructs 500 decision trees. When each decision tree grows, during the node splitting process, from a randomly selected subset of features (the number of features is set to the square root of the total number of features), the optimal splitting point is selected based on the minimum Gini impurity criterion, and the average reduction in Gini impurity caused by each feature in all decision trees is calculated as the feature importance score. According to the feature importance score, a certain number of the least important features are removed, and this process is repeated until the number of features reaches the preset value. The prediction results of all decision trees are integrated by the majority voting method to obtain the final prediction, including the final class prediction and the related probability.

[0083] Optimization of the prediction model: The hyperparameters of the random forest model are optimized through grid search and 5-fold cross-validation to find an optimal combination of hyperparameters, so that the random forest model has the highest prediction accuracy on unknown data and good generalization ability, that is, the prediction model can not only perform well on the training set, but also maintain a high prediction accuracy, sensitivity, specificity, F1 score, Kappa coefficient, Brier score, etc. on unseen data, as shown in Table 1.

[0084] Table 1 Evaluation of the efficacy of the system of the present invention in predicting the recurrence risk of PTC patients in the training set and validation set

[0085]

[0086] Note: The values in parentheses are the 95% confidence intervals.

[0087] Interpretation of the prediction model: As Figure 3As shown, the key predictive factors of the prediction model are explained using game theory-based SHAP analysis, and the contribution of each feature used for making predictions to the prediction result of the prediction model is quantified by calculating the Shapley value. Each point in the figure represents a sample, and its horizontal position indicates the contribution direction of the feature to the prediction result (a positive SHAP value on the right indicates an increased risk, and a negative value on the left indicates a decreased risk), and the color depth reflects the feature value size (red indicates a high value, and blue indicates a low value). The results show that the red dots (high feature values) of the radiomics label and clinical features (Hashimoto's thyroiditis) in terms of SHAP values tend to be distributed more on the right side, indicating that these features are significantly positively correlated with the risk of HIF-1α positive expression. Among them, the red dots of the radiomics label are distributed farthest to the right, indicating that it has the most significant impact on the prediction result. This visualization analysis method not only enhances the interpretability of the prediction model by quantifying the contribution of each feature but also provides clinical decision support, making the application of the model in clinical practice more reliable and practical.

[0088] IV. The risk output module constructs a web-based interactive machine learning calculator that can analyze and visualize the patient's recurrence risk in real time. As Figure 4 shown, through a simple and intuitive interface, this module allows users to input the key clinical and imaging information (such as radiomics label, Hashimoto's thyroiditis, etc.) that has been determined in the SHAP analysis. The module can directly output the risk probability of the patient's hypoxia-driven recurrence risk and visually display the recurrence risk probability through a circular pie chart (the proportion of the red area represents the high-risk degree). All analysis results can be generated in real time within 1 second after the user inputs the information and support instant result updates after dynamically adjusting the key clinical and imaging information, effectively assisting clinical decision-making.

Claims

1. A hypoxia-driven recurrence risk prediction system for papillary thyroid carcinoma based on artificial intelligence image analysis, characterized in that, The system receives thyroid ultrasound and CT image data, as well as corresponding conventional clinical feature data and tumor image qualitative feature data through a Web interface; The data preprocessing module on the server side preprocesses the thyroid ultrasound and CT image data and segments the thyroid tumor images; The feature extraction and selection module extracts high-throughput radiomics features from the thyroid tumor images and uses the trained lasso regression to screen out the core features related to hypoxia-driven recurrence from the high-throughput radiomics features to construct a radiomics signature; The risk prediction module trains the constructed hypoxia-driven papillary thyroid carcinoma recurrence risk prediction model based on the radiomics signature, conventional clinical features, tumor image qualitative features, and corresponding classification labels, and analyzes the key features affecting the prediction results through SHAP, and finally uses the trained prediction model to predict new data; The risk output module visualizes the prediction results by converting them into charts for output; The user interface module receives the information output by the risk output module and displays it to the user in real time through a Web interface.

2. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 1, wherein The conventional clinical features include the patient's age, gender, whether there is thyroid function abnormality, whether there is nodular goiter, and whether there is Hashimoto's thyroiditis; The tumor image qualitative features include tumor size, location, shape, internal components, echo features, degree of contact with the capsule, presence of calcification, and presence of cystic change.

3. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 1, wherein The data preprocessing module preprocesses the thyroid ultrasound and CT image data including format conversion, image resampling, and intensity normalization processing; among them, the resampling process is: I resampled (x,y,z) = Interpolate(I original , T resample ), Where, I original is the original image, T resample is the resampling transformation matrix, Interpolate(*) is the interpolation function, and I resampled (x, y, z) is the pixel value of the resampled image at the point (x, y, z); The normalization process is: Where, I min and I max are respectively the minimum and maximum pixel values of the resampled image, and I normalized (x, y, z) is the pixel value of the normalized image at the point (x, y, z).

4. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 1, wherein The method for the data preprocessing module to segment the thyroid tumor images is: using the pre-trained nnU-Net network model to automatically segment the thyroid tumor area from the preprocessed thyroid ultrasound and CT images to obtain the segmented thyroid tumor images; the training data includes obtaining thyroid ultrasound and CT images and corresponding tumor annotation data, and inputting the preprocessed thyroid ultrasound and CT images and corresponding tumor annotation data into the nnU-Net network model for training to obtain the trained nnU-Net network model.

5. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 4, characterized in that The nnU-Net network model adopts a four-layer encoding and four-layer decoding structure, uses a 3×3×3 convolution kernel for feature extraction and reconstruction, the encoder gradually extracts high-level features through downsampling, and the decoder gradually restores the spatial resolution through upsampling; Among them, the feature map of the l-th layer of the encoder is calculated as: where, and are the convolutional kernel weights and biases of the l-th layer respectively, σ is the activation function, and * is the convolution operation; The feature map of the l-th layer of the decoder is calculated as: Among them, Upsample is the upsampling operation, and are the convolutional kernel weights and biases of the l-th layer respectively, σ is the activation function, and * is the convolution operation; When training the nnU-Net network model, first define a Dice loss function to measure the prediction error of the network model, and select an optimizer to update the weights of the network model according to the gradient of the loss function. Through multiple iterations of training, the optimizer continuously adjusts the weights to make the value of the loss function gradually decrease. The Dice loss function formula is: where p i is the predicted value, and g i is the true value; The initial learning rate of this network model is 1e-4, and a cyclic learning rate scheduler is used to dynamically adjust the learning rate. The formula of the cyclic learning rate scheduler is: where LR min and LR max are the minimum and maximum learning rates respectively, t is the current training step, and T is the total number of steps.

6. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 1, wherein The feature extraction and selection module extracts high-throughput radiomics features from thyroid tumor images, including feature extraction and feature transformation. Among them, feature extraction extracts three types of radiomics features from thyroid tumor images through the PyRadiomics database: shape features, first-order features, and texture features. Shape features include tumor volume, surface area, and surface volume ratio. First-order features include mean, variance, skewness, kurtosis, and percentile. Texture features include gray-level co-occurrence matrix analysis of microscopic structure arrangement, gray-level dependence matrix detection of local heterogeneity, gray-level run-length matrix evaluation of texture directionality, and gray-level zone size matrix quantification of lesion spatial distribution. Feature transformation performs the following transformations on the extracted original features in sequence: exponential transformation, gradient transformation, logarithmic transformation, square transformation, square root transformation, logarithmic-σ transformation, and wavelet transformation.

7. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 1, characterized in that The lasso regression training data in the feature extraction and selection module includes a set of high-throughput radiomics features and corresponding classification labels. The classification labels are hypoxic-driven recurrence risk grades confirmed by pathology and clinical follow-up. During the training process, by adding an L1 regularization penalty term to the loss function, the weak correlation feature coefficients in the high-throughput radiomics features are compressed to zero, and the core features closely related to hypoxic-driven recurrence are screened out.

8. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 1, characterized in that The prediction model for the recurrence risk of oxygen-driven papillary thyroid carcinoma in the risk prediction module includes a random forest-based model. During the training process, the recursive feature elimination algorithm is used to recursively remove the least important features for the prediction target from the input features, and then the remaining training set is used to train the model. The training process of the prediction model is as follows: The recursive feature elimination algorithm is used to recursively screen features, including training a random forest model with all current features. The random forest model uses the bootstrap sampling method to draw a sample subset from the training set with replacement to construct decision trees. When each decision tree grows, during the node splitting process, from a randomly selected subset of features, the optimal splitting point is selected based on the minimum Gini impurity criterion, and the average reduction in Gini impurity caused by each feature in all decision trees is calculated as the feature importance score. According to the feature importance score, the least important feature is removed, and this process is repeated until the number of features reaches the preset value. The prediction results of all decision trees are integrated through the majority voting method to obtain the final prediction, including the final class prediction and related probabilities.

9. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 8, characterized in that, The prediction model optimizes the hyperparameters of the random forest model through grid search and 5-fold cross-validation to find an optimal combination of hyperparameters.

10. The hypoxia-driven papillary thyroid carcinoma recurrence risk prediction system based on artificial intelligence image analysis according to claim 8, wherein The key prediction factors of the prediction model are analyzed using game theory-based SHAP. By calculating the Shapley value to quantify the contribution of each feature used for prediction to the prediction result of the prediction model, the interpretability of the prediction model is enhanced.

Images