Disease classification prediction method, device, electronic device and storage medium

Abstract

The embodiment of the present invention provides a disease grading prediction method, device, electronic device and storage medium, the method comprising: obtaining target image features, target gene features and target clinical features of a target user; the target user comprises a first user in a preset training set and a second user in a preset test set; determining the target features of the target user according to the target image features, target gene features and target clinical features; training a preset grading prediction model based on a target feature vector and a preset label corresponding to the first user to obtain a target grading prediction model; inputting the target features of the second user into the target grading prediction model to obtain a target grading corresponding to the second user. The target grading prediction model is obtained through training, and automatic grading prediction is performed based on the target grading prediction model, thereby reducing the interference of human factors and improving the accuracy of disease grading prediction.

Description

Technical Field

[0001] The present invention relates to the field of medical image analysis, and in particular to a disease grading prediction method, device, electronic equipment and storage medium. Background Art

[0002] Brain tumors are abnormal cells that divide and grow unnaturally in brain tissue. They have a high incidence rate and a mortality rate of more than 3%, which seriously endangers human health. Among them, glioma is one of the common brain tumors originating from the skull. According to the classification of the World Health Organization, gliomas can be divided into four levels: I-IV. Among them, I-II is a low-grade glioma, and III-IV is a high-grade glioma. High-grade gliomas (HGG), such as glioblastoma multiforme (GBM), have an average survival time of 23 months, a two-year survival rate of 47.4%, and a four-year survival rate of only 18.5%. Low-grade gliomas (LGG), such as oligodendrogliomas and astrocytomas, have a ten-year survival rate of 57%. Accurate grading of brain gliomas is a prerequisite for saving patients' lives, and has positive clinical significance for treatment decisions, monitoring and management of chemoradiotherapy, and prognosis evaluation.

[0003] For brain tumors, magnetic resonance imaging (MRI), as a typical non-invasive imaging technology, can produce high-quality brain images without damage and skull artifacts, and can provide more comprehensive information for the analysis of brain tumors. It is the main technical means to analyze and treat brain tumors.

[0004] In the prior art, the grading method for diseases such as brain tumors is usually performed by radiologists through qualitative analysis based on their own experience and MRI images. This method based on manual experience is greatly affected by human factors, and incorrect grading often occurs, and the accuracy of disease grading is low. Summary of the invention

[0005] The embodiments of the present invention provide a disease grading prediction method, device, electronic device and storage medium to solve the problem of low accuracy of existing manual disease grading.

[0006] In order to solve the above problems, the embodiment of the present invention is implemented as follows:

[0007] In a first aspect, an embodiment of the present invention discloses a disease classification prediction method, comprising:

[0008] Obtaining target image features, target gene features, and target clinical features of a target user; the target user includes a first user in a preset training set and a second user in a preset test set;

[0009] Determining target features of the target user according to the target image features, the target gene features, and the target clinical features;

[0010] Based on the target feature and the preset label corresponding to the first user, a preset grading prediction model is trained to obtain a target grading prediction model;

[0011] The target features of the second user are input into the target grading prediction model to obtain the target grading corresponding to the second user.

[0012] Optionally, the acquiring of target image features of the target user includes:

[0013] Acquire a target image of the target user;

[0014] Segmenting the target image according to a preset image segmentation algorithm, and determining a target area in the target image;

[0015] According to a preset image processing algorithm, the image features of the target area are extracted to obtain the target image features of the target user.

[0016] Optionally, the step of obtaining target gene characteristics of a target user includes:

[0017] Acquiring target genomics data of the target user;

[0018] Based on the type of the target disease, a candidate gene is selected from the target genomics data; the candidate gene is related to the target disease and / or the mutation rate of the candidate gene in the target user is higher than a first preset threshold;

[0019] The gene features of the candidate genes are extracted to obtain the target gene features of the target user.

[0020] Optionally, the step of obtaining target clinical characteristics of a target user includes:

[0021] Acquiring preset clinical data of the target user;

[0022] According to the type of the target disease, screening the preset clinical data related to the target disease from the preset clinical data to obtain target clinical information;

[0023] The clinical characteristics of the target clinical information are extracted to obtain the target clinical characteristics of the target user.

[0024] Optionally, determining the target feature of the target user according to the target image feature, the target gene feature, and the target clinical feature includes:

[0025] The target imaging feature, the target gene feature and the target clinical feature are combined to obtain the target full feature;

[0026] According to a preset feature selection algorithm, the target features of the target user are selected from the target full features; the target features are the features ranked in the top N positions in importance among the target full features; N is the target number of target features; and N is an integer greater than 0.

[0027] Optionally, the merging of the target image feature, the target gene feature and the target clinical feature includes:

[0028] Vectorizing the target image feature, the target gene feature, and the target clinical feature respectively to obtain a target image feature vector, a target gene feature vector, and a target clinical feature vector;

[0029] The target image feature vector, the target gene feature vector and the target clinical feature vector are combined to obtain a first target full feature vector corresponding to the target full feature.

[0030] Optionally, selecting the target features of the target user from the target full features according to a preset feature selection algorithm includes:

[0031] Normalizing the first target full quantity feature vector to obtain a second target full quantity feature vector corresponding to the first target full quantity feature;

[0032] Determining the importance of each feature in the second target full feature vector according to the preset feature selection algorithm;

[0033] Determine the number of targets N corresponding to the target feature according to the importance and the preset feature selection algorithm;

[0034] In the second target full feature vector, the features ranked in the top N in terms of importance are selected as the target features.

[0035] Optionally, determining the number of targets N corresponding to the target feature includes:

[0036] According to the ranking of the importance, different numbers of features are selected in turn;

[0037] Input each number of features into the preset feature selection algorithm for verification, and determine the average accuracy corresponding to each number of features;

[0038] The number of features with the highest average accuracy is used as the target number N corresponding to the target features.

[0039] In a second aspect, an embodiment of the present invention discloses a disease classification prediction device, comprising:

[0040] An acquisition module, used to acquire target image features, target gene features, and target clinical features of a target user; the target user includes a first user in a preset training set and a second user in a preset test set;

[0041] A determination module, used to determine the target characteristics of the target user according to the target image characteristics, the target gene characteristics and the target clinical characteristics;

[0042] A training module, used to train a preset grading prediction model based on the target feature and a preset label corresponding to the first user to obtain a target grading prediction model;

[0043] An input module is used to input the target features of the second user into the target grading prediction model to obtain the target grading corresponding to the second user.

[0044] Optionally, the acquisition module is specifically used to:

[0045] Acquire a target image of the target user;

[0046] Segmenting the target image according to a preset image segmentation algorithm, and determining a target area in the target image;

[0047] According to a preset image processing algorithm, the image features of the target area are extracted to obtain the target image features of the target user.

[0048] Optionally, the acquisition module is further specifically used for:

[0049] Acquiring target genomics data of the target user;

[0050] Based on the type of the target disease, a candidate gene is selected from the target genomics data; the candidate gene is related to the target disease and / or the mutation rate of the candidate gene in the target user is higher than a first preset threshold;

[0051] The gene features of the candidate genes are extracted to obtain the target gene features of the target user.

[0052] Optionally, the acquisition module is further specifically used for:

[0053] Acquiring preset clinical data of the target user;

[0054] According to the type of the target disease, screening the preset clinical data related to the target disease from the preset clinical data to obtain target clinical information;

[0055] The clinical characteristics of the target clinical information are extracted to obtain the target clinical characteristics of the target user.

[0056] Optionally, the determining module is specifically used to:

[0057] The target imaging feature, the target gene feature and the target clinical feature are combined to obtain the target full feature;

[0058] According to a preset feature selection algorithm, the target features of the target user are selected from the target full features; the target features are the features ranked in the top N positions in importance among the target full features; N is the target number of target features; and N is an integer greater than 0.

[0059] Optionally, the determination module is specifically used to: vectorize the target image feature, the target gene feature, and the target clinical feature, respectively, to obtain a target image feature vector, a target gene feature vector, and a target clinical feature vector;

[0060] The target image feature vector, the target gene feature vector and the target clinical feature vector are combined to obtain a first target full feature vector corresponding to the target full feature.

[0061] Optionally, the determining module is further specifically configured to:

[0062] Normalizing the first target full quantity feature vector to obtain a second target full quantity feature vector corresponding to the first target full quantity feature;

[0063] Determining the importance of each feature in the second target full feature vector according to the preset feature selection algorithm;

[0064] Determine the number of targets N corresponding to the target feature according to the importance and the preset feature selection algorithm;

[0065] In the second target full feature vector, the features ranked in the top N in terms of importance are selected as the target features.

[0066] Optionally, the determining module is further specifically configured to:

[0067] According to the ranking of the importance, different numbers of features are selected in turn;

[0068] Input each number of features into the preset feature selection algorithm for verification, and determine the average accuracy corresponding to each number of features;

[0069] The number of features with the highest average accuracy is used as the target number N corresponding to the target features.

[0070] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of disease grade prediction described in the first aspect.

[0071] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of disease grade prediction described in the first aspect are implemented.

[0072] In an embodiment of the present invention, the target image features, target gene features, and target clinical features of the target user are obtained; the target user includes a first user in a preset training set and a second user in a preset test set; the target features of the target user are determined according to the target image features, target gene features, and target clinical features; based on the target feature vector and the preset label corresponding to the first user, the preset grading prediction model is trained to obtain the target grading prediction model; the target features of the second user are input into the target grading prediction model to obtain the target grade corresponding to the second user. In this way, in an embodiment of the present invention, the target grading prediction model is obtained by training the target grading prediction model by introducing the target image features, target gene features, and target clinical features of the target user, and automatic grading prediction is performed based on the target grading prediction model, thereby reducing the interference of human factors and improving the accuracy of disease grading prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A flowchart showing the steps of a disease classification prediction method according to an embodiment of the present invention is shown;

[0074] Figure 2 A schematic diagram of a process of disease classification prediction according to an embodiment of the present invention is shown;

[0075] Figure 3 A structural block diagram of a disease grading prediction device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] Reference Figure 1, showing a flow chart of the steps of a disease grading prediction method of the present invention. The execution subject of the embodiment of the present invention may refer to a computing device, specifically a medical device, an intelligent device, and a server, etc., which is not specifically limited in the embodiment of the present invention. The disease grading prediction method may specifically include:

[0078] Step 101: Obtain target image features, target gene features, and target clinical features of a target user; the target user includes a first user in a preset training set and a second user in a preset test set.

[0079] In the embodiment of the present invention, the target user may refer to a target patient suffering from a disease. Before obtaining the final target classification prediction model, it is necessary to complete the corresponding training and testing process based on the relevant data of the target user. The first user in the preset training set can be used for model training, and the second user in the preset test set can be used for model accuracy testing.

[0080] The target image feature may refer to the relevant features of the target user's MRI image. Exemplarily, the target influencing feature may include first-order statistical features, shape features, and texture features, among which the first-order statistical features may specifically include mean value, variance, etc., the shape features may specifically include volume, surface area, etc., and the texture features may specifically include grayscale co-occurrence matrix, grayscale region size matrix, etc., which may be determined based on the specific type of actual disease and the specific graded prediction requirements, and the embodiments of the present invention are not limited to this.

[0081] The target gene feature may refer to the gene information of the target user, and may specifically include the gene expression level, gene mutation status, etc. The target clinical feature may refer to the clinical data of the target user, and may specifically include the age, gender, height, weight, blood pressure, etc. of the target user. The embodiments of the present invention do not limit the specific types of the target gene feature and the target clinical feature.

[0082] The traditional manual film reading method can only be based on manual experience and some histological classification features in medical images. The genes introduced in the embodiments of the present invention are more critical predictors of changes in some molecular markers related to treatment and prognosis. For example, some molecular genetic markers (including IDH1 / 2 mutations, TP53 mutations, MGMT promoter methylation, 1p / 19q co-deletion, etc.) play an important role in the occurrence of tumors. In the embodiments of the present invention, the target image features are fused with the target gene features, and the target clinical features of the target user are combined. It can combine the advantages of both imaging and genomics, integrate biomolecular level information into imaging methods, and can also be combined with the actual clinical data of the target user, so as to further improve the accuracy of disease grading prediction.

[0083] Step 102: Determine the target characteristics of the target user according to the target image characteristics, the target gene characteristics and the target clinical characteristics.

[0084] In the embodiment of the present invention, the target feature may refer to each feature of the target user used for model training. The target user may have many target impact features, target gene features, and target clinical features. In this step, the features with high importance for disease grading prediction can be screened out through feature fusion and extraction to obtain the target feature. The model is subsequently trained based on the target feature, which can improve the accuracy of model prediction.

[0085] Step 103: Based on the target features and the preset label corresponding to the first user, a preset grading prediction model is trained to obtain a target grading prediction model.

[0086] In an embodiment of the present invention, a preset label may refer to a hierarchical label (label) corresponding to the first user in a preset training set. The preset label may be predetermined and labeled. A preset hierarchical prediction model may refer to a preset classification model, specifically a support vector machine (SVM) classification model, etc. Of course, other classifiers may also be used. The embodiment of the present invention does not limit the specific type of the preset hierarchical prediction model. A target hierarchical prediction model may refer to a trained hierarchical prediction model that can be used for prediction.

[0087] In this step, after obtaining the target features of the target user, the preset hierarchical prediction model can be trained based on the data of the first user in the preset training set, and the target features of multiple first users and the preset labels of the first users are input into the preset hierarchical prediction model. The training process of the preset hierarchical prediction model is executed to obtain the final target hierarchical prediction model.

[0088] Step 104: input the target features of the second user into the target classification prediction model to obtain the target classification corresponding to the second user.

[0089] In an embodiment of the present invention, the target classification may refer to the predicted result of the disease grade of the second user. After the preset classification prediction model is trained and the target classification prediction model is obtained, the target features of the second user in the preset test set may be input into the target classification prediction model to obtain the target classification of the second user. When it is necessary to predict the disease grade of other patients in the future, the target imaging features, target gene features and target clinical features of other patients may be collected, and the target features of other patients may be obtained based on these features. The target features of other patients may be input into the target classification prediction model to obtain the corresponding disease grades of other patients, which is convenient, fast and has a high accuracy rate.

[0090] In an embodiment of the present invention, the target image features, target gene features, and target clinical features of the target user are obtained; the target user includes a first user in a preset training set and a second user in a preset test set; the target features of the target user are determined according to the target image features, target gene features, and target clinical features; based on the target feature vector and the preset label corresponding to the first user, the preset grading prediction model is trained to obtain the target grading prediction model; the target feature vector of the second user is input into the target grading prediction model to obtain the target grade corresponding to the second user. In this way, in an embodiment of the present invention, the target grading prediction model is obtained by training the target grading prediction model by introducing the target image features, target gene features, and target clinical features of the target user, and automatic grading prediction is performed based on the target grading prediction model, thereby reducing the interference of human factors and improving the accuracy of disease grading prediction.

[0091] Optionally, in the embodiment of the present invention, obtaining the target image features of the target user in step 101 can be specifically implemented by the following steps S21 to S23:

[0092] Step S21: Acquire the target image of the target user.

[0093] In an embodiment of the present invention, the target image may refer to a multimodal MRI magnetic resonance image of the target user (including four sequences: T1, FLAIR, T1c, and T2), wherein T1 and T2 are physical quantities used to measure electromagnetic waves, T1 can be used to display anatomical images, and T2 can be used to display lesion images. FLAIR water-pressurized images are based on the liquid attenuation inversion sequence of magnetic resonance imaging, also known as water suppression imaging technology. The T1c sequence is to create a contrast agent (pigment) in the blood before doing MR. The bright areas have abundant blood supply, and the enhanced display indicates abundant blood flow, and the tumor site is exactly where the blood flow is very fast. The T1c sequence can further display the situation inside the tumor and distinguish between tumors and non-tumor lesions. Acquiring multimodal target images can ensure the comprehensiveness and integrity of the image data and improve the accuracy of subsequent grading predictions.

[0094] Exemplarily, when performing graded prediction for the target disease of glioma, an embodiment of the present invention can obtain and use the TCGA-LGG low-grade glioma dataset and the TCGA-GBM high-grade glioma dataset, which contains a total of 65 low-grade brain tumor patients and 102 high-grade brain tumor patients preoperative multimodal MRI magnetic resonance images, to perform feature extraction and subsequent model training and testing processes.

[0095] Step S22: segment the target image according to a preset image segmentation algorithm, and determine a target area in the target image.

[0096] In the embodiment of the present invention, a preset image segmentation algorithm can be used for image segmentation, which can specifically refer to a semantic feature pyramid network (SFPN) deep learning model, etc. The embodiment of the present invention does not limit the specific algorithm for image segmentation. The target area can refer to the disease area in the target image. For a brain tumor, the target area can refer to the tumor area.

[0097] In this step, after acquiring the target image, the target image can be preprocessed first, which can specifically include size scaling and standardization (Z-score standardization). Through preprocessing, the target image can be scaled to a uniform size and processing standard to ensure the accuracy of subsequent feature extraction and to minimize the impact of irrelevant factors such as image size on the accuracy of graded prediction. Afterwards, the SFPN deep learning model can be used to automatically segment the target image at the pixel level, and then the area to which each pixel belongs is classified and judged to determine the target area in the target image. For example, after pixel-level segmentation of the target image of a brain tumor patient, the classification result of each pixel can be determined, and 1 is output for the tumor area and 0 is output for the non-tumor area.

[0098] Step S23: extracting the image features of the target area according to a preset image processing algorithm to obtain the target image features of the target user.

[0099] In the embodiment of the present invention, the preset image processing algorithm can be used to extract image features. After the target area of ​​the disease is determined according to the original target image and the preset image segmentation algorithm, the image processing algorithm can be used to extract the image features of the target area image under four modalities (T1, FLAIR, T1c, T2) to form the target image features.

[0100] Exemplarily, for a target user with a brain tumor, in an embodiment of the present invention, an MRI image of the target user is first obtained, the MRI image is segmented and the target region, i.e., the tumor region image, is determined, and then the influencing features of the tumor region image under four modalities are extracted, including first-order statistical features, shape features, and texture features, totaling 386, to form an image feature vector R, wherein the first-order statistical features include 18 features, such as mean value, variance, kurtosis, energy, entropy, etc., the shape features include 14 features, such as volume, surface area, sphericity, maximum three-dimensional diameter, etc., and the texture features include 75 features, such as grayscale co-occurrence matrix, grayscale region size matrix, grayscale run matrix, grayscale correlation matrix, and adjacent grayscale difference matrix. Table 1 shows the types and numbers of image features extracted by different modalities in an embodiment of the present invention. It should be noted that since the shape feature does not depend on the modality and is only related to the segmentation result of the target region, only one set is extracted.

[0101]

[0102] Table 1

[0103] In the embodiment of the present invention, a target image of a target user is obtained; the target image is segmented according to a preset image segmentation algorithm, and a target area in the target image is determined; the image features of the target area are extracted according to a preset image processing algorithm to obtain the target image features of the target user. In this way, in the embodiment of the present invention, the target area can be automatically segmented and determined, and the image features of the target area can be subsequently extracted to obtain the target image features, thereby improving the efficiency of image processing, and the extracted image features are interpretable, further ensuring the accuracy of subsequent model predictions.

[0104] Optionally, in the embodiment of the present invention, obtaining the target gene characteristics of the target user in step 101 can be specifically implemented by the following steps S31 to S33:

[0105] Step S31: Acquire target genomics data of the target user.

[0106] In the embodiment of the present invention, the target genomics data may refer to the gene data of the target user.

[0107] Step S32: based on the type of the target disease, select a candidate gene from the target genomics data; the candidate gene is related to the target disease and / or the mutation rate of the candidate gene in the target user is higher than a first preset threshold.

[0108] In an embodiment of the present invention, the target disease may refer to a disease targeted by hierarchical prediction, such as glioma. The candidate gene may refer to a gene in the target genomics data that is highly correlated with the target disease and has a high mutation rate. The first preset threshold may refer to a pre-set mutation rate threshold. When the mutation rate is higher than the first preset threshold, it can be determined that the mutation rate of the gene is high.

[0109] Specifically, in this step, the first n genes with the strongest correlation with the target disease classification can be selected as the first candidate gene according to the type of the target disease; then the gene mutation state of the target user is obtained, and the gene mutation state includes two types: wild type (wildtype, marked as 0) and mutant (mutant, marked as 1). According to the different mutation rates of different genes in the target user (the mutation rate of a certain gene = the number of patients with this gene mutation in the data set / the total number of patients in the data set), the first q genes with the highest mutation rate are selected as the second candidate gene. Of course, the gene with a mutation rate higher than the first preset threshold can also be directly selected as the second candidate gene, and the embodiment of the present invention does not limit this. After obtaining the first candidate gene and the second candidate gene, the intersection or union of the two can be taken to determine the final candidate gene, so that the gene with the strongest correlation with the target disease or the highest mutation rate can be accurately determined, ensuring the accuracy and comprehensiveness of the subsequent gene feature extraction.

[0110] Step S33: extract the gene features of the candidate genes to obtain the target gene features of the target user.

[0111] In the embodiment of the present invention, after the candidate genes are determined, the gene characteristics of each target user for each candidate gene can be obtained respectively, specifically including gene expression level, gene mutation status, etc., to obtain the target gene characteristics of the target user.

[0112] In the embodiment of the present invention, the target genomics data of the target user is obtained; based on the type of the target disease, a candidate gene is selected from the target genomics data; the candidate gene is related to the target disease and / or the mutation rate of the candidate gene in the target user is higher than a first preset threshold; the gene features of the candidate gene are extracted to obtain the target gene features of the target user. In this way, in the embodiment of the present invention, by determining the candidate gene that is most related to the target disease or has the highest mutation rate, and extracting the gene features of the candidate gene of the target user, the genomic features can be integrated with the imaging features, thereby improving the accuracy of disease grading prediction.

[0113] Optionally, in the embodiment of the present invention, obtaining the target clinical characteristics of the target user in step 101 can be specifically implemented by the following steps S41 to S43:

[0114] Step S41: Acquire preset clinical data of the target user.

[0115] In the embodiment of the present invention, the preset clinical data may refer to various clinical data of the target user.

[0116] Step S42: According to the type of the target disease, the preset clinical data related to the target disease is screened from the preset clinical data to obtain target clinical information.

[0117] In the embodiment of the present invention, the target clinical information may refer to clinical data related to the target disease in the preset clinical data. Through data screening, clinical data irrelevant to the target disease can be filtered out to ensure the validity of the clinical data.

[0118] Step S43: extract clinical characteristics of the target clinical information to obtain target clinical characteristics of the target user.

[0119] In the embodiment of the present invention, after the target clinical information related to the target disease is determined, data processing operations such as discretization and labeling can be performed on the target clinical information to extract the target clinical characteristics of the target user.

[0120] In the embodiment of the present invention, the preset clinical data of the target user is obtained; according to the type of the target disease, the preset clinical data related to the target disease is screened from the preset clinical data to obtain the target clinical information; the clinical features of the target clinical information are extracted to obtain the target clinical features of the target user. In this way, in the embodiment of the present invention, the target clinical data of the target user is determined for the subsequent model training process through data screening and data extraction processing, which can improve the accuracy of disease classification.

[0121] Optionally, in the embodiment of the present invention, step 102 may be implemented by the following steps 1021 to 1022:

[0122] Step 1021: merge the target image features, the target gene features, and the target clinical features to obtain the target full features.

[0123] In the embodiment of the present invention, the target full feature may refer to a feature set composed of target image features, target gene features, and target clinical features. The target full feature includes all the features of the target image features, target gene features, and target clinical features determined in the above steps, and can serve as the data basis for subsequent feature screening.

[0124] Step 1022: select the target features of the target user from the target full set of features according to a preset feature selection algorithm; the target features are the features ranked in the top N positions in importance among the target full set of features; N is the target number of target features; and N is an integer greater than 0.

[0125] In the embodiment of the present invention, the preset feature selection algorithm may be a pre-set feature screening algorithm, specifically, a cross-validation-recursive feature elimination method, etc. The embodiment of the present invention does not limit the specific type of the preset feature selection algorithm. The degree of importance may refer to the importance of a certain feature in the full set of target features for hierarchical prediction. The target number N may refer to the optimal number of target features. It should be noted that the accuracy of hierarchical prediction for different numbers of target features is different. In this step, the target number N may be first determined based on the importance and the preset feature selection algorithm, and then the target number N of target features may be selected to form the target feature set ultimately used for hierarchical prediction.

[0126] In the embodiment of the present invention, the target image features, target gene features and target clinical features are merged to obtain the target full features; the target features of the target user are selected from the target full features according to the preset feature selection algorithm; the target features are the features ranked in the top N positions in terms of importance among the target full features. In this way, the target full features are obtained by feature merging, and then the target features with higher importance are determined based on the feature selection algorithm, so that the optimal target feature subset can be selected, and the classifier can be trained using the optimal feature subset, which can further improve the accuracy of the target classification prediction model.

[0127] Optionally, in the embodiment of the present invention, step 1021 may be implemented by the following steps S51 to S52:

[0128] Step S51, vectorizing the target image feature, the target gene feature and the target clinical feature respectively to obtain a target image feature vector, a target gene feature vector and a target clinical feature vector.

[0129] In the embodiment of the present invention, after obtaining the target image features, target gene features and target clinical features, the feature set can be vectorized to obtain the target image feature vector R, the target gene feature vector G and the target clinical feature vector C.

[0130] Step S52: Merge the target image feature vector, the target gene feature vector and the target clinical feature vector to obtain a first target full feature vector corresponding to the target full feature.

[0131] In an embodiment of the present invention, the first target full feature vector may refer to the vector corresponding to the original target full feature including all imaging genomics features. Specifically, the target imaging feature vector R, the target gene feature vector G, and the target clinical feature vector C may be combined to obtain the first target full feature vector F, that is, F=R+G+C.

[0132] In an embodiment of the present invention, the target image feature, the target gene feature, and the target clinical feature are vectorized respectively to obtain a target image feature vector, a target gene feature vector, and a target clinical feature vector; the target image feature vector, the target gene feature vector, and the target clinical feature vector are merged to obtain a first target full feature vector corresponding to the target full feature. In this way, by vectorizing and merging the features to obtain the first target full feature vector, feature processing based on vectors can improve the operation speed, and merging the vectors can ensure the comprehensiveness of the features during subsequent feature screening.

[0133] Optionally, in the embodiment of the present invention, step 1022 may specifically include the following steps S61 to S64:

[0134] Step S61: normalize the first target full quantity feature vector to obtain a second target full quantity feature vector corresponding to the first target full quantity feature.

[0135] In an embodiment of the present invention, the second target full feature vector may refer to the first target full feature vector after normalization. In this step, after the first target full feature is obtained by feature merging, the first target full feature vector F obtained after merging may be normalized, and all quantitative feature data in F may be normalized to the range of [-1,1] to obtain F', thereby obtaining the normalized second target full feature vector F'. The specific normalization formula may be:

[0136]

[0137] Step S62: Determine the importance of each feature in the second target full feature vector according to the preset feature selection algorithm.

[0138] In an embodiment of the present invention, after normalization, the importance of each feature in the second target full feature vector can be calculated using the second target full feature vector and a preset feature selection algorithm. Exemplarily, taking the preset feature selection algorithm as a support vector machine-based recursive feature elimination method (SVM-RFE) as an example, when performing feature selection based on the second target full feature vector F', the specific method is to randomly divide the data set into 5 subsets, 4 of which are used as training sets and the other as a validation set, so that 5 groups of training sets and validation sets can be obtained, and the 5 groups of training sets and validation sets are modeled using SVM classifiers respectively, and the importance of each feature is calculated, and the recursive feature elimination method (RFE) is used to obtain the importance ranking of all features.

[0139] Step S63: Determine the number of targets N corresponding to the target feature according to the importance and the preset feature selection algorithm.

[0140] In an embodiment of the present invention, after determining the importance of each feature in the second target full feature vector, verification can be performed again based on the importance and a preset feature selection algorithm to determine the optimal number N corresponding to the target features, thereby further improving the accuracy in subsequent model training processes.

[0141] Step S64: In the second target full feature vector, select the features ranked in the top N in terms of importance as the target features.

[0142] In the embodiment of the present invention, after the number of targets N is determined, the features ranked as the top N in importance can be selected from the second target full feature vector as target features based on the order of importance to form a target feature set.

[0143] In the embodiment of the present invention, the first target full feature vector is normalized to obtain the second target full feature vector corresponding to the first target full feature; the importance of each feature in the second target full feature vector is determined according to a preset feature selection algorithm; the target number N corresponding to the target feature is determined according to the importance and the preset feature selection algorithm; in the second target full feature vector, the features ranked in the top N in importance are selected as target features. In this way, by determining the importance of each feature and determining the optimal number of target features to select the target features, the rationality of feature selection can be improved, the selection of the optimal target feature set can be achieved, and the accuracy of classification of the target hierarchical prediction model obtained by subsequent training can be improved.

[0144] Optionally, in the embodiment of the present invention, step S63 may be implemented by the following steps S631 to S633:

[0145] Step S631: Select different numbers of features in sequence according to the order of importance.

[0146] In an embodiment of the present invention, after determining the importance of each feature in the second target full feature vector, the optimal number of target features can be further determined. In this step, different numbers of features can be first selected, for example, the top 5 features, the top 10 features, the top 20 features, etc. in order of importance can be selected from the second target full feature vector, and then feature sets composed of different numbers of features can be verified to determine the optimal number.

[0147] Step S632: Input each number of features into the preset feature selection algorithm for verification to determine the average accuracy corresponding to each number of features.

[0148] In the embodiment of the present invention, the average accuracy may refer to the accuracy of the model trained based on the number of features when it is finally used for the preset classification. Exemplarily, in this step, a 5-fold cross validation may be used to determine the optimal number of features to be selected, i.e., the target number N. Different numbers of features are selected in order of importance, and the selected feature sets are cross-validated to obtain the average accuracy corresponding to the feature sets with different numbers of features.

[0149] Step S633: Taking the number of features with the highest average accuracy as the target number N corresponding to the target features.

[0150] In an embodiment of the present invention, after determining the average accuracy corresponding to feature sets with different numbers of features, the number of features in the feature set with the highest average accuracy can be used as the optimal number of features to be selected. After that, an RFE (recursive feature elimination) operation can be performed on the entire data set to obtain all target features and form a target feature vector set. Exemplarily, taking the aforementioned data set of 65 patients with low-grade brain tumors and 102 patients with high-grade brain tumors as an example, the SVM-RFE method is used to determine that the optimal number of features is 19. Table 2 shows the types of target features determined after feature selection in an embodiment of the present invention.

[0151]

[0152] Table 2

[0153] In the embodiment of the present invention, different numbers of features are selected in sequence according to the order of importance; each number of features is input into a preset feature selection algorithm for verification, and the average accuracy corresponding to each number of features is determined; the number of features with the highest average accuracy is used as the target number N corresponding to the target feature. In this way, the optimal number of features is determined through the importance and the preset feature selection algorithm, and the optimal target feature subset can be selected. The classifier is trained using the optimal feature subset, which can further improve the accuracy of the model classification prediction.

[0154] For example, Figure 2 The following is a flow chart of a disease classification prediction method according to an embodiment of the present invention. The specific flow of the disease classification prediction method according to an embodiment of the present invention is described in combination with an example in the case where the target disease is glioma and the target user is a glioma patient:

[0155] Step 201, obtaining a multimodal MRI image of a patient with a brain glioma; automatically segmenting the multimodal MRI image according to a preset image segmentation algorithm to determine the tumor region (target region); and extracting target image features of the tumor region image through a preset image processing algorithm.

[0156] Step 202: Obtain the genetic data (target genomics data) of patients with brain glioma, determine candidate genes through correlation analysis and mutation rate statistics; then extract the genetic features of the candidate genes of each patient to obtain the target gene features.

[0157] Step 203: Acquire clinical data (preset clinical data) of patients with brain glioma, determine target clinical information through data screening, and then obtain target clinical features through feature extraction.

[0158] Steps 201 to 203 are described in detail in the foregoing content, and will not be repeated here in the embodiment of the present invention.

[0159] Step 204: merge the target image features, target clinical features, and target gene features to obtain the target full-quantity features, and then extract the target features to obtain the target features.

[0160] In an embodiment of the present invention, the TCGA-LGG low-grade glioma data set and the TCGA-GBM high-grade glioma data set are first obtained, which include 65 low-grade brain tumor patients and 102 high-grade brain tumor patients preoperative multimodal MRI magnetic resonance images (including T1, FLAIR, T1c, T2 four sequences) and preset clinical data and target genomics data corresponding to each patient. For MRI image data, the semantic feature pyramid network (SFPN) is used to automatically segment the tumor area, and the image processing algorithm is used to extract the first-order statistical features, shape features and texture features of the tumor area to form target image features; for gene data, candidate genes are determined by correlation analysis and mutation rate statistics, and the gene expression level and gene mutation state of the candidate genes are extracted as target gene features; for clinical data, target clinical information such as the patient's age and gender are obtained by data screening, and the target clinical information is discretized and labeled to obtain the target clinical features. The three features are merged to obtain the target full-quantity features, and the cross-validation-recursive feature elimination method is used to select the target full-quantity features to obtain the target feature vector set corresponding to the target feature.

[0161] Step 205: Input the target features and preset labels into a preset classification prediction model (SVM classifier) ​​and train to obtain a target classification prediction model.

[0162] In this step, the target feature vector set corresponding to the target feature is randomly divided into 5 subsets, and the support vector machine (SVM) classification model is used as the preset classification model for classification and the 5-fold cross validation method is used for model training to obtain the target classification model. After that, the target feature vectors of the patients (second users) in the preset test set can be sequentially input into the target classification model, and the model outputs the target classification, that is, the tumor classification prediction result. The cross validation method used in the embodiment of the present invention can specifically refer to grouping the original data, one part as a training set, and the other part as a test set, first using the training set to train the classifier, and then using the test set to test the trained model, so as to use this as a performance indicator for evaluating the classifier.

[0163] In an embodiment of the present invention, for the prediction results, the classification accuracy can be used as the performance evaluation index of the model, and the calculation formula is accuracy = (TP + TN) / (P + N), where TP represents the number of samples correctly classified as positive examples, TN represents the number of samples correctly classified as negative examples, and P + N represents the sum of all positive examples and negative examples, that is, the number of all samples.

[0164] The five-fold cross-validation evaluation results of the 167 patients in the above example are shown in Table 3. In order to verify the performance of the disease grading prediction method of the embodiment of the present invention, the imaging genomics method (grading prediction using only imaging features) is compared with the embodiment of the present invention. It can be seen that the disease grading prediction method of the embodiment of the present invention has higher grading accuracy on the four subsets (1, 2, 4, 5). The average accuracy of the 167 patients is 94%, which is 2.4% higher than that of the imaging genomics method alone.

[0165] Table 3 Five-fold cross validation evaluation results

[0166]

[0167]

[0168] Table 3

[0169] It can be seen from the above results that the disease grading method proposed in the embodiment of the present invention can not only comprehensively quantify the tumor phenotypic characteristics and reduce the interference of human factors through automated quantitative analysis, but also introduces the patient's genomic characteristics and clinical characteristics for prediction compared with the imaging genomics method that only uses image features, thereby improving the accuracy of grading prediction and providing a basis for formulating the best diagnosis and treatment plan for patients.

[0170] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0171] Reference Figure 3 , shows a structural block diagram of a disease grading prediction device according to an embodiment of the present invention. Specifically, the device 30 may include the following modules:

[0172] An acquisition module 301 is used to acquire target image features, target gene features, and target clinical features of a target user; the target user includes a first user in a preset training set and a second user in a preset test set;

[0173] A determination module 302 is used to determine the target feature of the target user according to the target image feature, the target gene feature and the target clinical feature;

[0174] A training module 303 is used to train a preset grading prediction model based on the target feature and the preset label corresponding to the first user to obtain a target grading prediction model;

[0175] The input module 304 is used to input the target features of the second user into the target grading prediction model to obtain the target grading corresponding to the second user.

[0176] In summary, the disease grading prediction device provided in the embodiment of the present invention obtains the target image features, target gene features and target clinical features of the target user; the target user includes the first user in the preset training set and the second user in the preset test set; the target features of the target user are determined according to the target image features, target gene features and target clinical features; based on the target feature vector and the preset label corresponding to the first user, the preset grading prediction model is trained to obtain the target grading prediction model; the target features of the second user are input into the target grading prediction model to obtain the target grade corresponding to the second user. In this way, in the embodiment of the present invention, the target grading prediction model is obtained by training the target grading prediction model by introducing the target image features, target gene features and target clinical features of the target user, and automatic grading prediction is performed based on the target grading prediction model, thereby reducing the interference of human factors and improving the accuracy of disease grading prediction.

[0177] Optionally, the acquisition module 301 is specifically used to:

[0178] Acquire a target image of the target user;

[0179] Segmenting the target image according to a preset image segmentation algorithm, and determining a target area in the target image;

[0180] According to a preset image processing algorithm, the image features of the target area are extracted to obtain the target image features of the target user.

[0181] Optionally, the acquisition module 301 is further specifically configured to:

[0182] Acquiring target genomics data of the target user;

[0183] Based on the type of the target disease, a candidate gene is selected from the target genomics data; the candidate gene is related to the target disease and / or the mutation rate of the candidate gene in the target user is higher than a first preset threshold;

[0184] The gene features of the candidate genes are extracted to obtain the target gene features of the target user.

[0185] Optionally, the acquisition module 301 is further specifically configured to:

[0186] Acquiring preset clinical data of the target user;

[0187] According to the type of the target disease, screening the preset clinical data related to the target disease from the preset clinical data to obtain target clinical information;

[0188] The clinical characteristics of the target clinical information are extracted to obtain the target clinical characteristics of the target user.

[0189] Optionally, the determining module 302 is specifically configured to:

[0190] The target imaging feature, the target gene feature and the target clinical feature are combined to obtain the target full feature;

[0191] According to a preset feature selection algorithm, the target features of the target user are selected from the target full features; the target features are the features ranked in the top N positions in importance among the target full features; N is the target number of target features; and N is an integer greater than 0.

[0192] Optionally, selecting the target features of the target user from the target full features according to a preset feature selection algorithm includes:

[0193] Normalizing the first target full quantity feature vector to obtain a second target full quantity feature vector corresponding to the first target full quantity feature;

[0194] Determining the importance of each feature in the second target full feature vector according to the preset feature selection algorithm;

[0195] Determine the number of targets N corresponding to the target feature according to the importance and the preset feature selection algorithm;

[0196] In the second target full feature vector, the features ranked in the top N in terms of importance are selected as the target features.

[0197] Optionally, determining the number of targets N corresponding to the target feature includes:

[0198] According to the ranking of the importance, different numbers of features are selected in turn;

[0199] Input each number of features into the preset feature selection algorithm for verification, and determine the average accuracy corresponding to each number of features;

[0200] The number of features with the highest average accuracy is used as the target number N corresponding to the target features.

[0201] Optionally, an embodiment of the present invention further provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the various processes of the above-mentioned disease grading prediction method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0202] Optionally, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the various processes of the above-mentioned disease grading prediction method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0203] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0204] It is easy for a person skilled in the art to think that any combination of the above embodiments is feasible, so any combination of the above embodiments is an implementation scheme of the present invention. However, due to space limitations, this specification will not describe them in detail here.

[0205] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0206] Similarly, it should be understood that in order to streamline the present invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the intention that the claimed invention requires more features than those expressly recited in each claim. Rather, as reflected in the claims, inventive aspects lie in less than all of the features of the individual embodiments previously disclosed. Therefore, the claims that follow the detailed description are hereby expressly incorporated into the detailed description, with each claim itself serving as a separate embodiment of the present invention.

[0207] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0208] In addition, those skilled in the art will appreciate that, although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, in the claims, any one of the claimed embodiments may be used in any combination.

[0209] It should be noted that the above embodiments illustrate the present invention rather than limit it, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising a number of different elements and by means of a suitably programmed computer. In a unit claim enumerating a number of devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc., does not indicate any order. These words may be interpreted as names.

Claims

1. A disease classification prediction method, characterized in that: include: Obtaining target imaging features, target gene features, and target clinical features of a target user; The target users include a first user in a preset training set and a second user in a preset test set; Determining target features of the target user according to the target image features, the target gene features, and the target clinical features; Based on the target feature and the preset label corresponding to the first user, a preset grading prediction model is trained to obtain a target grading prediction model; Inputting the target feature of the second user into the target classification prediction model to obtain the target classification corresponding to the second user; The determining the target feature of the target user according to the target image feature, the target gene feature and the target clinical feature includes: The target imaging feature, the target gene feature and the target clinical feature are combined to obtain the target full feature; According to a preset feature selection algorithm, the target feature of the target user is selected from the target full set of features; the target feature is the feature ranked in the top N positions in terms of importance among the target full set of features; N is the target number of target features; N is an integer greater than 0; The combining of the target image feature, the target gene feature and the target clinical feature includes: Vectorizing the target image feature, the target gene feature, and the target clinical feature respectively to obtain a target image feature vector, a target gene feature vector, and a target clinical feature vector; The target image feature vector, the target gene feature vector and the target clinical feature vector are combined to obtain a first target full feature vector corresponding to the target full feature; The step of selecting the target features of the target user from the target full set of features according to a preset feature selection algorithm includes: Normalizing the first target full quantity feature vector to obtain a second target full quantity feature vector corresponding to the first target full quantity feature; Determining the importance of each feature in the second target full feature vector according to the preset feature selection algorithm; Determine the number of targets N corresponding to the target feature according to the importance and the preset feature selection algorithm; In the second target full feature vector, the features ranked in the top N in terms of importance are selected as the target features; The determining the number of targets N corresponding to the target feature includes: According to the ranking of the importance, different numbers of features are selected in turn; Input each number of features into the preset feature selection algorithm for verification, and determine the average accuracy corresponding to each number of features; The number of features with the highest average accuracy is used as the target number N corresponding to the target features.

2. The method according to claim 1, characterized in that The step of obtaining target image features of a target user includes: Acquire a target image of the target user; Segmenting the target image according to a preset image segmentation algorithm, and determining a target area in the target image; According to a preset image processing algorithm, the image features of the target area are extracted to obtain the target image features of the target user.

3. The method according to claim 1, characterized in that The step of obtaining target gene characteristics of a target user includes: Acquiring target genomics data of the target user; Based on the type of the target disease, a candidate gene is selected from the target genomics data; the candidate gene is related to the target disease and / or the mutation rate of the candidate gene in the target user is higher than a first preset threshold; The gene features of the candidate genes are extracted to obtain the target gene features of the target user.

4. The method according to claim 1, characterized in that: The step of obtaining target clinical characteristics of a target user includes: Acquiring preset clinical data of the target user; According to the type of the target disease, screening the preset clinical data related to the target disease from the preset clinical data to obtain target clinical information; The clinical characteristics of the target clinical information are extracted to obtain the target clinical characteristics of the target user.

5. A disease classification prediction device, characterized in that: include: An acquisition module, used to acquire target image features, target gene features, and target clinical features of a target user; The target users include a first user in a preset training set and a second user in a preset test set; A determination module, used to determine the target characteristics of the target user according to the target image characteristics, the target gene characteristics and the target clinical characteristics; A training module, used to train a preset grading prediction model based on the target feature and a preset label corresponding to the first user to obtain a target grading prediction model; An input module, used for inputting the target feature of the second user into the target classification prediction model to obtain the target classification corresponding to the second user; The determination module is specifically used to merge the target image feature, the target gene feature and the target clinical feature to obtain a target full feature; select the target feature of the target user from the target full feature according to a preset feature selection algorithm; the target feature is the feature ranked in the top N positions in importance among the target full features; N is the target number of target features; N is an integer greater than 0; The merging of the target image feature, the target gene feature and the target clinical feature comprises: vectorizing the target image feature, the target gene feature and the target clinical feature respectively to obtain a target image feature vector, a target gene feature vector and a target clinical feature vector; merging the target image feature vector, the target gene feature vector and the target clinical feature vector to obtain a first target full feature vector corresponding to the target full feature; The step of selecting the target feature of the target user from the target full features according to the preset feature selection algorithm includes: normalizing the first target full feature vector to obtain a second target full feature vector corresponding to the first target full feature; determining the importance of each feature in the second target full feature vector according to the preset feature selection algorithm; determining the number of targets N corresponding to the target feature according to the importance and the preset feature selection algorithm; and selecting the features ranked in the top N positions in importance in the second target full feature vector as the target feature; The determining of the target number N corresponding to the target feature includes: selecting different numbers of features in sequence according to the order of importance; inputting each number of features into the preset feature selection algorithm for verification, and determining the average accuracy corresponding to each number of features; and taking the number of features with the highest average accuracy as the target number N corresponding to the target feature.

6. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the disease grading prediction method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the disease grading prediction method according to any one of claims 1 to 4 is implemented.

Images