A two-stage clustering diagnosis and automated machine learning prediction method and system

Abstract

The embodiment of the application provides a two-stage clustering diagnosis and automatic machine learning prediction method and model, the method comprises the following steps: feature weighting is performed on 3D space coordinates and SUV in tumor PET images to realize first-stage clustering; voxel-level clustering analysis is performed on tumor regions in the tumor PET images to realize second-stage clustering, and tumor sub-regions with different metabolic characteristics are outlined; based on the tumor sub-regions outlined in the two-stage clustering, features of the whole tumor and different sub-regions are extracted; based on the extracted features, an automatic machine learning prediction algorithm is constructed to predict labels in a classification task and numerical values in a regression task; based on the Kernel SHAP method, the feature importance of the prediction result is analyzed for interpretability. The method and model provided by the application can accurately identify and predict tumor high-risk and low-risk sub-regions, thereby improving the accuracy of individualized radiotherapy plans and improving the clinical prognosis of patients.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical machine learning, in particular to a two-stage clustering diagnosis and automatic machine learning prediction method and model. BACKGROUND

[0002] Tumor heterogeneity is one of the characteristics of malignant tumors, which can cause differences in tumor growth rate, invasion and metastasis, drug sensitivity, prognosis, etc. For example, some tumor regions may be sensitive to radiotherapy, while other regions may exhibit drug resistance. Therefore, accurately delineating the sub-regions of the tumor and dividing the tumor into different biological property regions is the basis for realizing personalized radiotherapy, which can guide more precise targeted radiotherapy strategies. In addition, the standard uptake value (SUV) is a semi-quantitative index commonly used in tumor diagnosis by Positron Emission Computed Tomography (PET). The SUV value of the tumor before radiotherapy has a significant correlation with the radiotherapy effect and survival rate of patients, and the change in SUV value during and after radiotherapy can also provide important information about the tumor's response to radiotherapy. Therefore, by analyzing the dynamic changes in SUV value, the radiotherapy effect of patients can be more accurately predicted, thereby guiding personalized radiotherapy strategies. Traditional machine learning methods have been widely used in tumor radiotherapy response prediction, but there are still challenges, especially in model selection and hyperparameter tuning. These steps usually require a lot of manual operation and deep professional knowledge, and in addition, traditional machine learning models usually lack sufficient interpretability, which makes it more difficult to understand the logic behind the model prediction. Therefore, it is crucial to develop machine learning algorithm models that can provide higher accuracy and better interpretability for predicting and analyzing the changes in tumor and sub-regional SUV. SUMMARY

[0003] The embodiments of the present application provide a two-stage clustering diagnosis and automatic machine learning prediction method and model, which can accurately identify and predict high-risk and low-risk sub-regions of tumors, thereby improving the accuracy of personalized radiotherapy plans and improving the clinical prognosis of patients.

[0004] To solve the above technical problems, in a first aspect, the embodiments of the present application provide a two-stage clustering diagnosis and automatic machine learning prediction method, which comprises the following steps: first, based on the K-means clustering algorithm, the 3D spatial coordinates and SUV in the tumor PET image are characterized to realize the first stage clustering; then, based on the New_mid K-means clustering algorithm, the voxel level clustering analysis is performed on the tumor region in the tumor PET image to realize the second stage clustering, and the tumor sub-regions with different metabolic characteristics are outlined; next, based on the tumor sub-regions outlined by the two-stage clustering, the features of the whole tumor and different sub-regions are extracted; then, based on the extracted features, an automatic machine learning prediction algorithm is constructed, and based on the automatic machine learning prediction algorithm, the labels in the classification task and the numerical values in the regression task are predicted to obtain the prediction results; finally, based on the Kernel SHAP method, the feature importance of the prediction results is analyzed for interpretability.

[0005] In some example embodiments, based on the K-means clustering algorithm, the 3D spatial coordinates and SUV in the tumor PET image are characterized to realize the first stage clustering, comprising: using a step method to perform multi-group weighting on the 3D spatial coordinates and SUV features in the tumor PET image to obtain multiple weight combination clustering results; based on the K-means clustering algorithm, the clustering evaluation index of each weight combination clustering result is calculated; based on the clustering evaluation index, the 3D spatial coordinates and SUV in the tumor PET image are finally characterized.

[0006] In some example embodiments, using a step method to perform multi-group weighting on the 3D spatial coordinates and SUV features in the tumor PET image to obtain multiple weight combination clustering results, comprising: using a step method to take the weight of the SUV feature in the tumor PET image from 0.1 to 1.0 with a step of 0.1; according to the corresponding relationship of the 3D spatial coordinates and SUV feature weights in the tumor PET image, the 3D spatial coordinates feature in the tumor PET image is weighted in multiple groups to obtain multiple weight combination clustering results; the corresponding relationship of the 3D spatial coordinates and SUV feature weights in the tumor PET image is:

[0007] w x = w y = w z = (1-w SUV ) / 3

[0008] wherein w x represents the x-coordinate feature weight in the tumor PET image; w y represents the y-coordinate feature weight in the tumor PET image; w z represents the z-coordinate feature weight in the tumor PET image; wSUV SUV feature weight in tumor PET image.

[0009] In some example embodiments, a cluster evaluation index of each weight combination clustering result is calculated based on the K-means clustering algorithm, including: substituting each weight combination into the K-means clustering algorithm to calculate a comprehensive score as the cluster evaluation index of the weight combination clustering result; and the calculation formula of the comprehensive score is:

[0010]

[0011] wherein CS represents the comprehensive score; SC pre tumor cluster profile coefficient in the early stage of radiotherapy; SC mid tumor cluster profile coefficient in the middle stage of radiotherapy; DS high Dice similarity of high-risk sub-regions in the early stage of radiotherapy and the middle stage of radiotherapy; DS low Dice similarity of low-risk sub-regions in the early stage of radiotherapy and the middle stage of radiotherapy.

[0012] In some example embodiments, the 3D spatial coordinates and SUV in the tumor PET image are finally feature-weighted based on the cluster evaluation index, including: for one sample, based on the cluster evaluation index of each weight combination clustering result, the SUV feature weight in the tumor PET image corresponding to the optimal cluster evaluation index is selected; for all samples, based on the SUV feature weight in the tumor PET image corresponding to the optimal cluster evaluation index selected for each sample, the average value is obtained to obtain the final SUV feature weight in the tumor PET image for all samples; and according to the corresponding relationship between the 3D spatial coordinates and the SUV feature weight in the tumor PET image, the final 3D spatial coordinate feature weight in the tumor PET image for all samples is calculated.

[0013] In some example embodiments, voxel-level cluster analysis is performed on the tumor region in the tumor PET image based on the New_mid K-means clustering algorithm to realize the second-stage clustering, and tumor sub-regions with different metabolic characteristics are delineated, including: based on the K-means clustering algorithm, the clustering analysis of the tumor region in the middle stage of radiotherapy is performed, the early-stage SUV feature and the middle-stage SUV feature are fused and input for voxel-level cluster analysis, and the New_mid K-means clustering algorithm is constructed; based on the New_mid K-means clustering algorithm, the optimal weight of the middle-stage SUV feature is selected by step-by-step debugging; wherein the middle-stage SUV feature weight is greater than or equal to the early-stage SUV feature weight when selected, so as to avoid neglecting the data aggregation degree of itself in the clustering analysis process of the tumor region in the middle stage of radiotherapy to obtain better coincidence degree stability.

[0014] In some example embodiments, based on the tumor sub-regions outlined by the two-stage clustering, feature extraction is performed on the whole tumor and different sub-regions, including: the features extracted from the whole tumor and different sub-regions include traditional features and radiomics features; the traditional features include the volume of the whole tumor and different sub-regions, the average SUV value of the whole tumor and different sub-regions, and the radiotherapy dose of the whole tumor and different sub-regions; the radiomics features include 6 intensity histogram features, 19 texture features from a gray level co-occurrence matrix, 5 texture features from a neighborhood gray level difference matrix, and 11 texture features from a gray level zone size matrix, a total of 41 radiomics features.

[0015] In some example embodiments, based on the extracted features, an automatic machine learning prediction algorithm is constructed, including: based on the extracted features, an automatic machine learning core framework is built; the automatic machine learning core framework includes a meta-learning module, a Bayesian optimization module, and an ensemble learning module; wherein the meta-learning module uses past data sets and model performance metadata to guide model selection and hyperparameter initialization for new tasks; the Bayesian optimization module predicts performance under different configurations by constructing a probability model, thereby guiding the search and tuning process of hyperparameters; the ensemble learning module assigns weights to each model based on the performance indicators of the models, and then integrates them into a powerful ensemble model, which can effectively reduce overfitting and improve the generalization ability of the model on unknown data.

[0016] In some example embodiments, based on the Kernel SHAP method, the feature importance of the prediction result is analyzed for interpretability, including: based on the SHAP method, a Kernel SHAP method is constructed by fusing kernel regression and weight sampling technology; based on the Kernel SHAP method, the estimation problem of Shapley value is converted into a weighted least squares regression problem, and efficient linear model solving technology is used to estimate the contribution value of each feature; based on the contribution value of each feature, the feature importance of the prediction result is analyzed for interpretability.

[0017] In a second aspect, the embodiments of the present application also provide a two-stage clustering diagnosis and automatic machine learning prediction model, comprising a two-stage clustering module and a prediction analysis module connected in sequence; the two-stage clustering module comprises a feature weighting module and a New_mid K-means clustering module; the feature weighting module is configured to perform feature weighting on 3D spatial coordinates and SUV value features in a tumor PET image according to a K-means clustering algorithm to realize first-stage clustering; the New_mid K-means clustering module is configured to perform voxel-level clustering analysis on a tumor region in the tumor PET image according to a New_mid K-means clustering algorithm to realize second-stage clustering and delineate tumor sub-regions with different metabolic characteristics; the prediction analysis module comprises a feature extraction module, an automatic machine learning module and a Kernel SHAP interpretability analysis module; the feature extraction module is configured to perform feature extraction on the whole tumor and different sub-regions according to the tumor sub-regions delineated by the two-stage clustering; the automatic machine learning module is configured to construct an automatic machine learning prediction algorithm according to the extracted features, and predict labels in a classification task and numerical values in a regression task based on the automatic machine learning prediction algorithm to obtain a prediction result; and the Kernel SHAP interpretability analysis module is configured to perform interpretability analysis on feature importance of the prediction result according to a Kernel SHAP method.

[0018] The technical solutions provided by the embodiments of the present application have at least the following advantages:

[0019] The embodiments of the present application provide a two-stage clustering diagnosis and automatic machine learning prediction method and model, which comprises the following steps: first, performing feature weighting on 3D spatial coordinates and SUV in a tumor PET image according to a K-means clustering algorithm to realize first-stage clustering; then, performing voxel-level clustering analysis on a tumor region in the tumor PET image according to a New_mid K-means clustering algorithm to realize second-stage clustering and delineate tumor sub-regions with different metabolic characteristics; next, performing feature extraction on the whole tumor and different sub-regions according to the tumor sub-regions delineated by the two-stage clustering; then, constructing an automatic machine learning prediction algorithm based on the extracted features, and predicting labels in a classification task and numerical values in a regression task based on the automatic machine learning prediction algorithm to obtain a prediction result; finally, performing interpretability analysis on feature importance of the prediction result based on a Kernel SHAP method.

[0020] The application provides a two-stage clustering diagnosis and automatic machine learning prediction method and model. Firstly, the application uses a two-stage clustering algorithm to efficiently cluster tumor PET images, thereby distinguishing high-risk and low-risk sub-regions with different metabolic characteristics in the tumor, and providing data support for formulating more detailed radiotherapy strategies in clinical practice. Secondly, the application extracts traditional features and image features, combines the data obtained from voxel clustering and the clinical information of patients, and uses an automatic machine learning prediction algorithm to construct a precise radiotherapy response prediction model. The model can automatically select the optimal feature combination and adopt the most suitable machine learning algorithm for integration, thereby realizing precise prediction of the radiotherapy response of patients. Finally, the application verifies the excellent performance of the model through the application results of patients with locally advanced non-small cell lung cancer, and shows the significant advantages of the model in accurate segmentation of tumor sub-regions and prediction analysis of radiotherapy response. The model provided by the application can accurately identify high-risk and low-risk sub-regions of tumors and perform prediction analysis, thereby improving the accuracy of individualized radiotherapy plans and improving the clinical prognosis of patients. BRIEF DESCRIPTION OF DRAWINGS

[0021] One or more embodiments are illustrated by way of example in the accompanying drawings that are not intended to be limiting of the application, unless otherwise specifically indicated, the drawing figures are not necessarily to scale.

[0022] Figure 1 A flowchart of a two-stage clustering diagnosis and automatic machine learning prediction method provided by an embodiment of the application.

[0023] Figure 2 A core framework diagram of automatic machine learning provided by an embodiment of the application.

[0024] Figure 3 A structure diagram of a two-stage clustering diagnosis and automatic machine learning prediction model provided by an embodiment of the application.

[0025] Figures 4A-4B A visualization diagram of a two-stage clustering diagnosis and automatic machine learning prediction model provided by an embodiment of the application for clustering sub-regions in the early radiotherapy period and clustering sub-regions in the middle radiotherapy period of patient 001.

[0026] Figures 5A-5C A Kernel SHAP explainability analysis diagram of a two-stage clustering diagnosis and automatic machine learning prediction model provided by an embodiment of the application for performing 3 traditional feature classification tasks and regression tasks on the entire tumor, the high-risk tumor sub-region, and the low-risk tumor sub-region.

[0027] Figures 6A-6CA two-stage clustering diagnosis and automatic machine learning prediction model provided by an embodiment of the present application provides Kernel SHAP explainability analysis diagrams for 41 radiomic feature classification tasks and regression tasks of the entire tumor, high-risk tumor sub-regions, and low-risk tumor sub-regions. DETAILED DESCRIPTION

[0028] As can be known from the background, tumor heterogeneity is one of the characteristics of malignant tumors, which can cause differences in tumor growth rate, invasion and metastasis, drug sensitivity, prognosis, and other aspects. The SUV value of the tumor before radiotherapy has a significant correlation with the radiotherapy effect and survival rate of the patient, and the change in the SUV value during and after radiotherapy can also provide important information about the tumor response to radiotherapy. Traditional machine learning methods have been widely used in tumor radiotherapy response prediction, but usually require a large amount of manual operation and deep professional knowledge in model selection and hyperparameter adjustment, and in addition, traditional machine learning models usually lack sufficient explainability, which makes it more difficult to understand the logic behind the model prediction. It is crucial to develop machine learning algorithm models that can provide higher accuracy and better explainability to predict and analyze the changes in the SUV of the tumor and its sub-regions.

[0029] Traditional machine learning methods usually require a large amount of manual operation and deep professional knowledge in model selection and hyperparameter adjustment, and in addition, traditional machine learning models usually lack sufficient explainability, which makes it more difficult to understand the logic behind the model prediction. In view of the deficiencies in the prior art, the embodiments of the present application provide a two-stage clustering diagnosis and automatic machine learning prediction method and model, which comprises the following steps: first, based on the K-means clustering algorithm, the 3D spatial coordinates and SUV in the tumor PET image are feature-weighted to realize the first-stage clustering; then, based on the New_mid K-means clustering algorithm, the tumor region in the tumor PET image is analyzed at the voxel level to realize the second-stage clustering, and the tumor sub-regions with different metabolic characteristics are outlined; next, based on the tumor sub-regions outlined by the two-stage clustering, features are extracted from the entire tumor and different sub-regions; then, based on the extracted features, an automatic machine learning prediction algorithm is constructed, and based on the automatic machine learning prediction algorithm, the labels in the classification task and the numerical values in the regression task are predicted to obtain the prediction results; finally, based on the Kernel SHAP method, the feature importance of the prediction results is analyzed for explainability. The purpose of the present application is to provide a two-stage clustering diagnosis and automatic machine learning prediction method and model, which can accurately identify and predict the high-risk and low-risk sub-regions of the tumor, thereby improving the accuracy of individualized treatment plans and improving the clinical prognosis of patients.

[0030] The embodiments of the present application will be described in detail below with reference to the drawings. However, those skilled in the art can understand that in the embodiments of the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, the technical solutions claimed by the present application can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0031] Referring to Figure 1 The embodiments of the present application provide a two-stage clustering diagnosis and automatic machine learning prediction method and model, which comprises the following steps:

[0032] Step S101, based on the K-means clustering algorithm, the 3D spatial coordinates and SUV in the tumor PET image are feature-weighted to realize the first-stage clustering.

[0033] Step S102, based on the New_mid K-means clustering algorithm, the tumor region in the tumor PET image is analyzed at the voxel level to realize the second-stage clustering, and the tumor sub-region with different metabolic characteristics is outlined.

[0034] Step S103, based on the tumor sub-region outlined by the two-stage clustering, the features of the whole tumor and different sub-regions are extracted.

[0035] Step S104, based on the extracted features, an automatic machine learning prediction algorithm is constructed, and based on the automatic machine learning prediction algorithm, the labels in the classification task and the numerical values in the regression task are predicted to obtain the prediction results.

[0036] Step S105, based on the Kernel SHAP method, the feature importance of the prediction results is analyzed for interpretability.

[0037] The two-stage tumor sub-region clustering diagnosis and automatic machine learning prediction model provided by the present application is based on the K-means clustering algorithm and the New_mid K-means clustering algorithm, and uses the two-stage clustering algorithm to efficiently cluster the PET image, thereby distinguishing the high-risk and low-risk sub-regions with different metabolic characteristics inside the tumor, which can provide data support for the clinical development of more detailed treatment strategies. In addition, the two-stage tumor sub-region clustering diagnosis and automatic machine learning prediction model is based on the automatic machine learning prediction algorithm and the Kernel SHAP method, and by combining the data obtained from voxel clustering and the clinical information of the patient after traditional feature and image feature extraction, a precise radiotherapy response prediction model is constructed using the automatic machine learning prediction algorithm, which can automatically select the optimal feature combination and adopt the most suitable machine learning algorithm for integration, thereby realizing the precise prediction of the radiotherapy response of the patient.

[0038] In some embodiments, in step S101, the 3D spatial coordinates and SUV in the tumor PET image are characterized to realize the first-stage clustering based on the K-means clustering algorithm, including:

[0039] Step S1011, the 3D spatial coordinates and SUV characteristics in the tumor PET image are characterized in multiple groups by using a step-by-step method, and multiple weight combination clustering results are obtained.

[0040] Step S1012, based on the K-means clustering algorithm, the clustering evaluation index of each weight combination clustering result is calculated.

[0041] Step S1013, based on the clustering evaluation index, the 3D spatial coordinates and SUV in the tumor PET image are finally characterized.

[0042] In some embodiments, in step S1011, the 3D spatial coordinates and SUV characteristics in the tumor PET image are characterized in multiple groups by using a step-by-step method, and multiple weight combination clustering results are obtained, including:

[0043] Step S1011a, the weight of the SUV characteristics in the tumor PET image is stepped from 0.1 to 1.0 by using a step-by-step method.

[0044] Step S1011b, according to the corresponding relationship of the 3D spatial coordinates and SUV characteristics weight in the tumor PET image, the 3D spatial coordinates characteristics in the tumor PET image are characterized in multiple groups, and multiple weight combination clustering results are obtained.

[0045] The corresponding relationship of the 3D spatial coordinates and SUV characteristics weight in the tumor PET image is:

[0046] w x = w y = w z = (1-w SUV ) / 3

[0047] Wherein, w x represents the x-coordinate characteristic weight in the tumor PET image; w y represents the y-coordinate characteristic weight in the tumor PET image; w z represents the z-coordinate characteristic weight in the tumor PET image; w SUV represents the SUV characteristic weight in the tumor PET image.

[0048] In some embodiments, in step S1012, a cluster evaluation index of each weight combination clustering result is calculated based on the K-means clustering algorithm, including: substituting each weight combination into the K-means clustering algorithm, and calculating a comprehensive score as the cluster evaluation index of the weight combination clustering result; the calculation formula of the comprehensive score is:

[0049]

[0050] wherein CS represents the comprehensive score; SC pre represents the tumor cluster profile coefficient in the early radiotherapy stage; SC mid represents the tumor cluster profile coefficient in the middle radiotherapy stage; DS high represents the Dice similarity of the high-risk sub-region in the early radiotherapy stage and the middle radiotherapy stage; DS low represents the Dice similarity of the low-risk sub-region in the early radiotherapy stage and the middle radiotherapy stage.

[0051] In some embodiments, in step S1013, the 3D spatial coordinates and SUV in the tumor PET image are finally weighted based on the cluster evaluation index, including:

[0052] Step S1013a, for one sample, based on the cluster evaluation index of each weight combination clustering result, the SUV feature weight in the tumor PET image corresponding to the optimal cluster evaluation index is selected.

[0053] Step S1013b, for all samples, based on the SUV feature weight in the tumor PET image corresponding to the optimal cluster evaluation index selected for each sample, the average value is obtained to obtain the final SUV feature weight in the tumor PET image for all samples, and the final 3D spatial coordinate feature weight in the tumor PET image for all samples is calculated according to the corresponding relationship between the 3D spatial coordinates and the SUV feature weight in the tumor PET image in step S1011b.

[0054] In some embodiments, in step S102, the tumor region in the tumor PET image is clustered and analyzed at the voxel level based on the New_mid K-means clustering algorithm to realize the second stage clustering, and the tumor sub-region with different metabolic characteristics is delineated, including:

[0055] Step S1021, based on the K-means clustering algorithm, the radiotherapy middle period tumor region is clustered and analyzed, the radiotherapy early period SUV feature and the radiotherapy middle period SUV feature are fused and input for voxel-level clustering analysis, and the New_mid K-means clustering algorithm is constructed.

[0056] Step S1022: Based on the New_mid K-means clustering algorithm, the optimal weight of the SUV feature in the middle stage of radiotherapy is selected by debugging through the stepwise method.

[0057] It is worth noting that the weight of the SUV feature in the middle stage of radiotherapy should be greater than or equal to the weight of the SUV feature in the early stage of radiotherapy, so as to avoid ignoring the data aggregation degree in the process of cluster analysis of tumor regions in the middle stage of radiotherapy in order to obtain better coincidence stability.

[0058] Specifically, the process of selecting the optimal weight of the SUV feature in the middle stage of radiotherapy by debugging the step method is similar to the process of assigning multiple groups of weights to the 3D spatial coordinates and SUV features in the tumor PET image by the step method. However, in the process of selecting the optimal weight of the SUV feature in the middle stage of radiotherapy by debugging the step method, it is required that the weight of the SUV feature in the middle stage of radiotherapy should be greater than or equal to the weight of the SUV feature in the early stage of radiotherapy. At the same time, the weight of the 3D spatial coordinate feature in the early stage of radiotherapy should be calculated based on the correspondence between the 3D spatial coordinates and the SUV feature weight in the tumor PET image in step S1011b, while the weight of the 3D spatial coordinate feature in the middle stage of radiotherapy should be calculated based on the correspondence between the new 3D spatial coordinates and the SUV feature weight. The correspondence between the new 3D spatial coordinates and the SUV feature weight is:

[0059] w xmid =w ymid =w zmid =[1-(w SUVpre +w SUVmid )] / 3

[0060] w SUVpre +w SUVmid =w SUV

[0061] w SUVmid ≥w SUVpre

[0062] Among them, w xmid represents the x-coordinate feature weight; w ymid represents the y-coordinate feature weight; w zmid represents the z-coordinate feature weight; w SUVpre represents the weight of SUV features in the early stage of radiotherapy; w SUVmid represents the characteristic weight of SUV in the middle stage of radiotherapy; w SUV Represents the total weight of SUV features.

[0063] In some embodiments, in step S103, based on the tumor subregions delineated by the two-stage clustering, feature extraction is performed on the entire tumor and different subregions, including:

[0064] Step S1031, feature extraction is performed on the entire tumor, and the extracted features include traditional features and radiomic features.

[0065] Step S1032, feature extraction is performed on the high-risk tumor sub-region, and the extracted features include traditional features and radiomic features.

[0066] Step S1033, feature extraction is performed on the low-risk tumor sub-region, and the extracted features include traditional features and radiomic features.

[0067] Specifically, feature extraction is performed on the entire tumor, the high-risk tumor sub-region, and the low-risk tumor sub-region respectively, and the extracted features include 3 traditional features and 41 radiomic features. The 3 traditional features include the volume corresponding to the entire tumor or the tumor sub-region, the average SUV value corresponding to the entire tumor or the tumor sub-region, and the radiotherapy dose corresponding to the entire tumor or the tumor sub-region. The 41 radiomic features include 6 intensity histogram features, 19 texture features from a gray tone space-dependent matrix, 5 texture features from a neighborhood gray tone difference matrix, and 11 texture features from a gray region size matrix.

[0068] In some embodiments, in step S104, an automatic machine learning prediction algorithm is constructed based on the extracted features to predict labels in a classification task and numerical values in a regression task, including:

[0069] Step S1041, based on the extracted features, an automatic machine learning core framework is built; the automatic machine learning core framework includes a meta-learning module, a Bayesian optimization module, and an ensemble learning module.

[0070] Step S1042, based on the automatic machine learning prediction algorithm, the labels in the classification task are predicted.

[0071] Step S1043, based on the automatic machine learning prediction algorithm, the numerical values in the regression task are predicted.

[0072] The meta-learning module of the automatic machine learning core framework in step S1041 uses past data sets and metadata of model performance to guide model selection and initialization of hyperparameters for new tasks; the Bayesian optimization module predicts the performance under different configurations by constructing a probability model, thereby guiding the search and optimization process of hyperparameters; the ensemble learning module assigns weights to each model based on the performance indicators of the models, and then integrates into a powerful ensemble model, which can effectively reduce overfitting and improve the generalization ability of the model on unknown data.

[0073] Specifically, the value to be predicted in the regression task is the change rate of the SUV mean of the whole tumor or tumor sub-region. At the same time, samples satisfying the corresponding relationship are classified into the "responsive" class and are additionally labeled with "1"; samples not satisfying the corresponding relationship are classified into the "non-responsive" class and are additionally labeled with "0", and the label to be predicted in the classification task is "1" or "0"; the corresponding relationship is:

[0074]

[0075] wherein, ΔSUVmean represents the change rate of the SUV mean of the whole tumor or tumor sub-region; SUVpre represents the average SUV value of the whole tumor or tumor sub-region in the pre-radiotherapy period; and SUVmid represents the average SUV value of the whole tumor or tumor sub-region in the mid-radiotherapy period.

[0076] Referring to Figure 2 The embodiments of the present application provide a core framework diagram of automatic machine learning, including a meta-learning module, a Bayesian optimization module and an ensemble learning module. In step S1041, the core framework of automatic machine learning is built, including:

[0077] In step S1041a, the meta-learning module is built, and the past data set and the metadata of model performance are used to guide the model selection and the initialization of hyperparameters of a new task.

[0078] In step S1041b, the Bayesian optimization module is built, and a probability model is constructed to predict the performance under different configurations, thereby guiding the search and optimization process of hyperparameters.

[0079] In step S1041c, the ensemble learning module is built, and weights are assigned to each model based on the performance indicators of the models, and then integrated into a powerful ensemble model, which can effectively reduce overfitting and improve the generalization ability of the model on unknown data.

[0080] In some embodiments, in step S105, the feature importance of the prediction result is analyzed for interpretability based on the Kernel SHAP method, including:

[0081] In step S1051, the Kernel SHAP method is constructed based on the SHAP method by fusing kernel regression and weight sampling technology.

[0082] In step S1052, the Kernel SHAP method is used to convert the estimation problem of Shapley value into a weighted least squares regression problem, and efficient linear model solving technology is used to estimate the contribution value of each feature.

[0083] In step S1053, based on the contribution value of each feature in step S1052, the feature importance of the prediction result in step S104 is analyzed for interpretability.

[0084] Referring to Figure 3 The embodiment of the present application also provides a two-stage clustering diagnosis and automatic machine learning prediction model, comprising a two-stage clustering module 101 and a prediction analysis module 102 connected in sequence; wherein the two-stage clustering module comprises a feature weighting module 1011 and a New_mid K-means clustering module 1012; the feature weighting module 1011 is used for feature weighting of 3D space coordinates and SUV value features in tumor PET images according to a K-means clustering algorithm to realize first-stage clustering; the New_mid K-means clustering module 1012 is used for voxel-level clustering analysis of tumor regions in tumor PET images according to a New_mid K-means clustering algorithm to realize second-stage clustering and delineate tumor sub-regions with different metabolic characteristics; the prediction analysis module 102 comprises a feature extraction module 1021, an automatic machine learning module 1022 and a Kernel SHAP interpretability analysis module 1023; the feature extraction module 1021 is used for feature extraction of the whole tumor and different sub-regions according to the tumor sub-regions delineated by the two-stage clustering; the automatic machine learning module 1022 is used for constructing an automatic machine learning prediction algorithm according to the extracted features, and predicting labels in a classification task and numerical values in a regression task based on the automatic machine learning prediction algorithm to obtain a prediction result; and the Kernel SHAP interpretability analysis module 1023 is used for interpretability analysis of feature importance of the prediction result according to a Kernel SHAP method.

[0085] The two-stage clustering diagnosis and automatic machine learning prediction model provided by the present application is described in detail below with respect to a specific embodiment.

[0086] Firstly, the two-stage clustering diagnosis and automatic machine learning prediction model proposed by the embodiment of the present application is based on a Windows 11 system and implemented using Python 3.9.

[0087] In addition, it should be noted that the embodiment of the present application is aimed at analyzing and evaluating the radiotherapy response of locally advanced non-small cell lung cancer.

[0088] Firstly, the SUV feature weight in the tumor PET image is taken from 0.1 to 1.0 with a step of 0.1 by the feature weighting module, and the 3D spatial coordinate feature is weighted in multiple groups according to the corresponding relationship between the 3D spatial coordinates and the SUV feature weight in the tumor PET image, then each weight combination is substituted into the K-means clustering algorithm to obtain the optimal SUV feature weight corresponding to the optimal comprehensive score of each patient, and the final optimal SUV feature weight for all patients is obtained by averaging, and the final 3D spatial coordinate feature weight for all patients is calculated according to the corresponding relationship between the 3D spatial coordinates and the SUV feature weight in the tumor PET image. The final optimal SUV feature weight is 0.8, and the corresponding x coordinate feature weight, y coordinate feature weight and z coordinate feature weight are all 1 / 15.

[0089] Then, the optimal weight of the mid-course SUV feature in radiotherapy is selected by the New_mid K-means clustering module using a similar step-by-step debugging method, while ensuring that the mid-course SUV feature weight should be greater than or equal to the pre-radiotherapy SUV feature weight. The optimal mid-course SUV feature weight for clustering the mid-course tumor PET image in radiotherapy is 0.4, the optimal pre-radiotherapy SUV feature weight is 0.4, and the corresponding x coordinate feature weight, y coordinate feature weight and z coordinate feature weight are all 1 / 15. According to the feature weighting result, the New_mid K-means clustering algorithm is used to cluster the pre-radiotherapy and mid-course tumor PET images of each patient to obtain the cluster 2 class results, wherein the cluster 2 class results of the pre-radiotherapy and mid-course tumor PET images of patient 001 are shown in FIG. 2. Figures 4A-4B . Figure 4A FIG. 1 shows a visualization schematic diagram of the two-stage clustering diagnosis and automatic machine learning prediction model provided by the present application for the pre-radiotherapy cluster sub-region of patient 001; Figure 4B FIG. 2 shows a visualization schematic diagram of the two-stage clustering diagnosis and automatic machine learning prediction model provided by the present application for the mid-course cluster sub-region of patient 001, wherein the light color part is the high-risk sub-region and the dark color part is the low-risk sub-region.

[0090] The comparative analysis of the contour coefficient, CH-index, Dice similarity and Euclidean distance of the New_mid K-means algorithm, K-means algorithm, spectral clustering algorithm and GMM algorithm in cluster 2 class is shown in Table 1 and Table 2. It can be seen that the New_mid K-means clustering has a greater improvement in sub-region delineation stability compared with K-means, and the clustering effect is better than that of spectral clustering algorithm and GMM algorithm.

[0091] Table 1. The profile coefficient, CH-index of New_mid K-means algorithm, K-means algorithm, spectral clustering algorithm and GMM algorithm when clustering 2 classes in an embodiment

[0092]

[0093] Table 2. The Dice similarity, Euclidean distance of New_mid K-means algorithm, K-means algorithm, spectral clustering algorithm and GMM algorithm when clustering 2 classes in an embodiment

[0094]

[0095] Next, through the feature extraction module, based on the tumor sub-region outlined by the two-stage clustering module, feature extraction is performed on the whole tumor, high-risk tumor sub-region and low-risk tumor sub-region respectively. The extracted features include 3 traditional features and 41 radiomics features. The 3 traditional features include the volume corresponding to the whole tumor or tumor sub-region, the average SUV value corresponding to the whole tumor or tumor sub-region, and the radiotherapy dose corresponding to the whole tumor or tumor sub-region. The 41 radiomics features include 6 intensity histogram features, 19 texture features from the gray level co-occurrence matrix, 5 texture features from the neighborhood gray level difference matrix, and 11 texture features from the gray level zone size matrix.

[0096] Then, through the automatic machine learning module, the LOOCV (Leave-One-Out Cross-Validation) method is adopted, and in the classification task, the automatic machine learning (Auto Machine Learning, AutoML), logistics regression (Logistics Regression, LR), support vector machine (Support Vector Machine, SVM), random forest (Random Forest, RF), and gradient boosting decision tree (Gradient Boosting Decision Tree, GBDT) methods are used for prediction and comparative analysis. The AUC values obtained by 3 traditional feature classification and 41 radiomics feature classification are shown in Table 3 and Table 4. In the regression task, the automatic machine learning, logistics regression, SVM, RF, and GBDT methods are used for prediction and comparative analysis. The RMSE values obtained by 3 traditional feature regression and 41 radiomics feature regression are shown in Table 5 and Table 6. It can be seen that the AUC values obtained by the automatic machine learning algorithm in the classification task are generally higher than those of other classic machine learning algorithms, indicating that its classification prediction in the whole tumor, high-risk tumor sub-region, and low-risk tumor sub-region is more accurate. At the same time, the RMSE values obtained by the automatic machine learning algorithm in the regression task are generally lower than those of other classic machine learning algorithms, indicating that its regression prediction in the whole tumor, high-risk tumor sub-region, and low-risk tumor sub-region is more accurate. It can be seen that the automatic machine learning algorithm has excellent prediction performance.

[0097] Table 3 AUC values obtained by automatic machine learning algorithm and classic machine learning algorithm 3 traditional feature classification in automatic machine learning module in an embodiment

[0098]

[0099] Table 4 AUC values obtained by automatic machine learning algorithm and classic machine learning algorithm 41 radiomics feature classification in automatic machine learning module in an embodiment

[0100]

[0101]

[0102] Table 5 RMSE values obtained by automatic machine learning algorithm and classic machine learning algorithm 3 traditional feature regression in automatic machine learning module in an embodiment

[0103]

[0104] Table 6 RMSE values obtained by automatic machine learning algorithm and classic machine learning algorithm 41 radiomics feature regression in automatic machine learning module in an embodiment

[0105]

[0106] Finally, the Kernel SHAP explainability analysis module is used to analyze the feature importance of the prediction results based on the contribution value of each feature. The Kernel SHAP explainability analysis is performed on the prediction results of the 3 traditional feature classification tasks and regression tasks of the entire tumor, the high-risk tumor sub-region, and the low-risk tumor sub-region, as shown in Figures 5A-5C . Among them, Figure 5A FIG. 7 shows the Kernel SHAP explainability analysis diagram of the two-stage clustering diagnosis and automatic machine learning prediction model provided by the present application for the 3 traditional feature classification tasks (top) and regression tasks (bottom) of the entire tumor; Figure 5B FIG. 8 shows the Kernel SHAP explainability analysis diagram of the model for the 3 traditional feature classification tasks (top) and regression tasks (bottom) of the high-risk tumor sub-region; Figure 5C FIG. 9 shows the Kernel SHAP explainability analysis diagram of the model for the 3 traditional feature classification tasks (top) and regression tasks (bottom) of the low-risk tumor sub-region.

[0107] The Kernel SHAP explainability analysis is performed on the prediction results of the 41 radiomics feature classification tasks and regression tasks of the entire tumor, the high-risk tumor sub-region, and the low-risk tumor sub-region, as shown in Figures 6A-6C . Among them, Figure 6A FIG. 10 shows the Kernel SHAP explainability analysis diagram of the two-stage clustering diagnosis and automatic machine learning prediction model provided by the present application for the 41 radiomics feature classification tasks (left) and regression tasks (right) of the entire tumor; Figure 6B FIG. 11 shows the Kernel SHAP explainability analysis diagram of the model for the 41 radiomics feature classification tasks (left) and regression tasks (right) of the high-risk tumor sub-region; Figure 6C FIG. 12 shows the Kernel SHAP explainability analysis diagram of the model for the 41 radiomics feature classification tasks (left) and regression tasks (right) of the low-risk tumor sub-region.

[0108] Based on the above technical solutions, the embodiment of the present application provides a two-stage clustering diagnosis and automatic machine learning prediction method and model, which comprises the following steps: first, based on the K-means clustering algorithm, the 3D space coordinates and SUV in the tumor PET image are characterized to realize the first stage clustering; then, based on the New_mid K-means clustering algorithm, the tumor region in the tumor PET image is clustered at the voxel level to realize the second stage clustering, and the tumor sub-region with different metabolic characteristics is outlined; next, based on the tumor sub-region outlined by the two-stage clustering, the features of the whole tumor and different sub-regions are extracted; then, based on the extracted features, an automatic machine learning prediction algorithm is constructed, and the labels in the classification task and the numerical values in the regression task are predicted based on the automatic machine learning prediction algorithm to obtain the prediction results; finally, based on the Kernel SHAP method, the feature importance of the prediction results is analyzed for interpretability.

[0109] The present application provides a two-stage clustering diagnosis and automatic machine learning prediction method and model, first, the present application uses a two-stage clustering algorithm to efficiently cluster PET images, thereby distinguishing high-risk and low-risk sub-regions with different metabolic characteristics within the tumor, which can provide data support for developing more detailed treatment strategies in clinical practice. Secondly, the present application combines the data obtained from voxel clustering and the clinical information of patients after traditional feature and imageomics feature extraction, and uses an automatic machine learning prediction algorithm to construct a precise radiotherapy response prediction model, which can automatically select the optimal feature combination and adopt the most suitable machine learning algorithm for integration, thereby realizing precise prediction of patient radiotherapy response. Finally, the application of the present application to patients with locally advanced non-small cell lung cancer verifies the excellent performance of the model, which shows its significant advantages in accurate segmentation of tumor sub-regions and prediction analysis of radiotherapy response. The model provided by the present application can accurately identify high-risk and low-risk sub-regions of tumors and perform prediction analysis, thereby improving the accuracy of individualized treatment plans and improving the clinical prognosis of patients.

[0110] Those skilled in the art can understand that the above embodiments are specific examples of the present application, and in actual application, various changes can be made in form and detail without departing from the spirit and scope of the present application. Any person skilled in the art can make changes and modifications without departing from the spirit and scope of the present application, therefore the protection scope of the present application should be limited by the scope defined in the claims.

Claims

1. A two-stage clustering diagnosis and automated machine learning prediction method, characterized in that, The method comprises the following steps: Based on the K-means clustering algorithm, the 3D spatial coordinates and SUV in the tumor PET image are characterized to realize the first stage clustering; Based on the New_mid K-means clustering algorithm, the voxel-level clustering analysis of the tumor region in the tumor PET image is performed to realize the second stage clustering, and the tumor sub-regions with different metabolic characteristics are outlined; Based on the tumor sub-regions outlined by the two-stage clustering, the features of the whole tumor and different sub-regions are extracted; Based on the extracted features, an automatic machine learning prediction algorithm is constructed, and based on the automatic machine learning prediction algorithm, the labels in the classification task and the numerical values in the regression task are predicted to obtain the prediction results; Based on the Kernel SHAP method, the feature importance of the prediction results is analyzed for interpretability; Based on the K-means clustering algorithm, the 3D spatial coordinates and SUV in the tumor PET image are characterized to realize the first stage clustering, comprising: Using the step-by-step method, the 3D spatial coordinates and SUV features in the tumor PET image are characterized in multiple groups to obtain multiple weight combination clustering results; Based on the K-means clustering algorithm, the clustering evaluation index of each weight combination clustering result is calculated; Based on the clustering evaluation index, the 3D spatial coordinates and SUV in the tumor PET image are finally characterized; Using the step-by-step method, the 3D spatial coordinates and SUV features in the tumor PET image are characterized in multiple groups to obtain multiple weight combination clustering results, comprising: Using the step-by-step method, the weight of the SUV feature in the tumor PET image is taken from 0.1 to 1.0 with a step of 0.1; According to the corresponding relationship between the 3D spatial coordinates and SUV feature weights in the tumor PET image, the 3D spatial coordinates feature in the tumor PET image is characterized in multiple groups to obtain multiple weight combination clustering results; The corresponding relationship between the 3D spatial coordinates and SUV feature weights in the tumor PET image is: wherein, represents an x-coordinate feature weight in the tumor PET image; represents a y-coordinate feature weight in the tumor PET image; represents a z-coordinate feature weight in the tumor PET image; represents an SUV feature weight in the tumor PET image; Based on the New_mid K-means clustering algorithm, the voxel-level clustering analysis of the tumor region in the tumor PET image is performed to realize the second stage clustering, and the tumor sub-regions with different metabolic characteristics are outlined, comprising: Based on the K-means clustering algorithm, the pre-radiotherapy SUV feature and the mid-radiotherapy SUV feature are fused and input for voxel-level clustering analysis to construct the New_mid K-means clustering algorithm for mid-radiotherapy tumor region clustering analysis; Based on the New_mid K-means clustering algorithm, the optimal weight of the mid-radiotherapy SUV feature is selected by the step-by-step method; wherein, the mid-radiotherapy SUV feature weight is greater than or equal to the pre-radiotherapy SUV feature weight to avoid ignoring the data aggregation degree of itself in the mid-radiotherapy tumor region clustering analysis process to obtain better coincidence stability.

2. The two-stage clustering diagnosis and automated machine learning prediction method of claim 1, wherein, Based on the K-means clustering algorithm, the clustering evaluation index of each weight combination clustering result is calculated, comprising: The comprehensive score is calculated by substituting each weight combination into the K-means clustering algorithm, and the comprehensive score is taken as a clustering evaluation index of the weight combination clustering result; The calculation formula of the comprehensive score is: where CS represents a composite score; represents a tumor cluster profile coefficient in the early stage of radiotherapy; represents a tumor cluster profile coefficient in the middle stage of radiotherapy; represents a Dice similarity of high-risk sub-regions in the early stage of radiotherapy and the middle stage of radiotherapy; represents a Dice similarity of low-risk sub-regions in the early stage of radiotherapy and the middle stage of radiotherapy.

3. The two-stage clustering diagnosis and automated machine learning prediction method of claim 1, wherein, Based on the clustering evaluation index, the 3D spatial coordinates and SUV in the tumor PET image are finally weighted, including: For a sample, based on the clustering evaluation index of each weight combination clustering result, the SUV feature weight in the tumor PET image corresponding to the optimal clustering evaluation index is selected; For all samples, based on the SUV feature weight in the tumor PET image corresponding to the optimal clustering evaluation index selected for each sample, the average value of the SUV feature weight in the tumor PET image for all samples is obtained; according to the corresponding relationship between the 3D spatial coordinates and the SUV feature weight in the tumor PET image, the final 3D spatial coordinate feature weight in the tumor PET image for all samples is calculated.

4. The two-stage clustering diagnosis and automated machine learning prediction method of claim 1, wherein, Based on the tumor sub-regions outlined by the two-stage clustering, features are extracted from the whole tumor and different sub-regions, including: The features extracted from the whole tumor and different sub-regions include traditional features and radiomics features; The traditional features include the volume of the whole tumor and different sub-regions, the average SUV value of the whole tumor and different sub-regions, and the radiotherapy dose of the whole tumor and different sub-regions; The radiomics features include 6 intensity histogram features, 19 texture features from a gray tone space dependence matrix, 5 texture features from a neighborhood gray tone difference matrix, and 11 texture features from a gray region size matrix, a total of 41 radiomics features.

5. The two-stage clustering diagnosis and automated machine learning prediction method of claim 1, wherein, Based on the extracted features, an automatic machine learning prediction algorithm is constructed, including: Based on the extracted features, an automatic machine learning core framework is built; The automatic machine learning core framework includes a meta-learning module, a Bayesian optimization module, and an ensemble learning module; wherein, The meta-learning module uses past data sets and metadata of model performance to guide model selection and initialization of hyperparameters for new tasks; The Bayesian optimization module predicts the performance under different configurations by constructing a probability model, and then guides the search and optimization process of hyperparameters; The ensemble learning module assigns weights to each model based on the performance indicators of the models, and then integrates them into a powerful ensemble model, which can effectively reduce overfitting and improve the generalization ability of the model on unknown data.

6. The two-stage clustering diagnosis and automated machine learning prediction method of claim 1, wherein, Based on the Kernel SHAP method, the feature importance of the prediction result is analyzed for interpretability, including: Based on the SHAP method, the Kernel SHAP method is constructed by integrating kernel regression and weight sampling technology; Based on the Kernel SHAP method, the estimation problem of Shapley value is converted into a weighted least squares regression problem, and efficient linear model solving technology is used to estimate the contribution value of each feature; Based on the contribution value of each feature, the feature importance of the prediction result is analyzed for interpretability.

7. A two-stage clustering diagnosis and automatic machine learning prediction system for implementing the two-stage clustering diagnosis and automatic machine learning prediction method according to any one of claims 1 to 6, characterized in that, It includes a two-stage clustering module and a prediction analysis module connected in sequence; The two-stage clustering module comprises a feature weighting module and a New_mid K-means clustering module; The feature weighting module is configured to perform feature weighting on 3D spatial coordinates and SUV value features in the tumor PET image according to a K-means clustering algorithm to realize first-stage clustering; The New_mid K-means clustering module is configured to perform voxel-level clustering analysis on the tumor region in the tumor PET image according to a New_mid K-means clustering algorithm to realize second-stage clustering, and to delineate tumor sub-regions with different metabolic characteristics; The prediction analysis module comprises a feature extraction module, an automatic machine learning module, and a Kernel SHAP interpretability analysis module; The feature extraction module is configured to perform feature extraction on the entire tumor and different sub-regions according to the tumor sub-regions delineated by the two-stage clustering; The automatic machine learning module is configured to construct an automatic machine learning prediction algorithm according to the extracted features, and to predict labels in a classification task and numerical values in a regression task based on the automatic machine learning prediction algorithm to obtain a prediction result; The Kernel SHAP interpretability analysis module is configured to perform interpretability analysis on feature importance of the prediction result according to a Kernel SHAP method.

Images