WO2021135774A1

WO2021135774A1 - Tumor prediction method and device, cloud platform, and computer-readable storage medium

Info

Publication number: WO2021135774A1
Application number: PCT/CN2020/132372
Authority: WO
Inventors: 邓胡川; 赵安江; 高杰临; 丁瑞鹏; 谢庆国
Original assignee: 苏州瑞派宁科技有限公司
Priority date: 2020-01-02
Filing date: 2020-11-27
Publication date: 2021-07-08
Also published as: CN111210441A

Abstract

Disclosed in embodiments of the present application are a tumor prediction method and device, a cloud platform, and a computer-readable storage medium. The tumor prediction method is executed on the cloud platform, and comprises: calling an obtained target prediction model to segment an obtained target image of a target patient, so as to obtain a segmented image containing a tumor area; extracting a high-dimensional feature and a depth feature from the obtained segmented image; performing screening processing on the high-dimensional feature and the depth feature according to a preset condition; and calling the target prediction model to fuse the screened depth feature and high-dimensional feature to obtain a fusion feature, and predicting a tumor classification of the target patient according to the fusion feature. By means of the technical solution provided in the embodiments of the present application, the accuracy of a tumor classification prediction result can be improved, and a doctor can be efficiently assisted in diagnosis.

Description

Tumor prediction method, device, cloud platform and computer readable storage medium

Technical field

This application relates to the technical field of medical data processing, and in particular to a tumor prediction method, device, cloud platform, and computer-readable storage medium.

Background technique

Regardless of cell morphology or tissue structure, tumor tissues are different from the normal tissues from which they originated to varying degrees. This difference is called atypia. The size of the atypia can be expressed by the degree of differentiation and maturity of tumor tissue. The small atypia of the tumor tissue indicates that the degree of differentiation is high, and the degree of malignancy is low; on the contrary, it indicates that the degree of differentiation of tumor tissue is low, and the degree of malignancy is high.

Malignant tumors are divided into early stage, middle stage and late stage. Most of the early stage malignant tumors can be cured, and the intermediate stage malignant tumors can relieve pain and prolong life. Therefore, tumor classification prediction is particularly important. At present, the method of imagingomics is mainly used to predict tumor classification. This method is mainly used in Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (MRI). A large number of quantitative image features are extracted from the regions of interest in medical images such as Emission Tomography (PET), and machine learning methods are used to screen and analyze these quantitative image features, and select the most valuable features associated with clinical problems. Use selection Models are constructed with the features derived, and the constructed models are used for tumor diagnosis and clinical phenotype prediction.

In the process of realizing this application, the inventor found at least the following problems in the prior art:

(1) Existing imagingomics methods need to implement different processing steps on multiple systems or software, for example, segmentation of tumor regions on image processing software such as 3D slicer and Mazda, and then import the results into MATLAB, Python The model is trained on software such as SPSS and R, and finally the Area Under Curve (AUC) and survival curve are drawn using software such as SPSS and R. This method not only requires the configuration of the cumbersome environment required under multiple software, but also the multi-system processing makes the data collection and sorting work cumbersome, which can easily cause problems such as data loss, incomplete data information, and data sharing.

(2) Existing imagingomics methods generally use imaging data in a single data center for testing, which makes the established tumor classification prediction model low in repeatability, versatility, anti-interference and difficult to be widely used.

(3) The existing imaging omics method basically only manually outlines the segmentation of images layer by layer for a certain type of tumor (for example, kidney cancer or lung cancer), and extracts gray-scale intensity features and three-dimensional shapes from the outlined segmentation area. Features, texture features and wavelet features and other high-dimensional features, and then use these high-dimensional features for analysis and research. However, due to the large differences between different tumors, extracting the same features cannot fully express the deep information and hidden information of different tumor regions. Therefore, the existing imagingomics methods are not universal for different tumors.

(4) In the existing imaging omics technology, the precise and rapid segmentation of tumors is a great challenge. In terms of accuracy, the doctor’s manual segmentation results are still the gold standard, which largely relies on the doctor’s expertise and experience, and the reproducibility rate is low. The traditional manual segmentation method is mainly performed by professional imaging doctors, but they cannot process the patient's image data in large quantities, and cannot avoid the limitation of time-consuming and labor-intensive. Even if the semi-automatic segmentation method is adopted, it is necessary for the doctor to mark the target area and the background area on multiple images of each patient. Although the operation frequency of the doctor is reduced, it is still time-consuming and labor-intensive.

Summary of the invention

The purpose of the embodiments of the present application is to provide a tumor prediction method, device, cloud platform, and computer-readable storage medium to solve at least one problem in the prior art.

In order to solve the above technical problems, embodiments of the present application provide a tumor prediction method. The tumor prediction method can be executed on a cloud platform and can include:

Calling the acquired target prediction model to segment the acquired target image of the target patient to obtain a segmented image containing the tumor area;

Extracting high-dimensional features and depth features from the obtained segmented image;

Screening the high-dimensional features and the depth features according to preset conditions;

The target prediction model is called to fuse the selected depth features and the high-dimensional features to obtain fusion features, and the tumor classification of the target patient is predicted according to the fusion features.

Optionally, the target prediction model is obtained in the following manner:

Obtain the target prediction model from an external device or locally.

Optionally, obtaining the target prediction model locally includes:

Training and verifying a pre-built machine learning model by using the acquired sample image data, the sample image data including training data and verification data, and matching the target image;

The machine learning model that has achieved the optimal training effect and passed the verification is determined as the target prediction model.

Optionally, before training the machine learning model, the tumor prediction method includes:

Select the pre-stored sample image data from a local database; or

The sample image data is obtained by processing the received patient data.

Optionally, obtaining the sample image data by processing the received patient data includes:

Analyze the format of the received patient data;

The sample image data is selected from the parsed patient data according to a preset standard.

Optionally, the preset criteria include whether the patient data is complete, whether it has been clinically verified, and whether it meets clinical indicators.

Optionally, filtering the depth features and the high-dimensional features extracted from the target image according to preset conditions includes:

Use sparse representation algorithm, lasso algorithm, Fisher discriminant, feature selection algorithm based on maximum correlation-minimum redundancy, or feature selection algorithm based on conditional mutual information to filter the high-dimensional features and the depth features to filter out those that satisfy The high-dimensional features and depth features of the preset conditions.

Optionally, the target image includes a CT image, an MRI image, a PET image, a US image, a SPECT image, and/or a PET/CT image.

Optionally, the target prediction model includes an AlexNet model or a VGGNet model.

The embodiment of the present application also provides a tumor prediction device, which may be set on a cloud platform and may include:

A segmentation unit configured to call the acquired target prediction model to segment the acquired target image of the target patient to obtain a segmented image containing the tumor region;

An extraction unit configured to extract high-dimensional features and depth features from the obtained segmented image;

A screening unit configured to screen the high-dimensional features and the depth features according to preset conditions;

A fusion unit configured to call the target prediction model to fuse the selected depth feature and the high-dimensional feature to obtain a fusion feature;

A prediction unit configured to predict the tumor classification of the target patient according to the fusion feature.

Optionally, the tumor prediction device further includes:

The acquiring unit is configured to acquire the target prediction model by using the acquired sample image data to train and verify the pre-built machine learning model, and to optimize the training effect and pass the verification. A machine learning model is determined as the target prediction model, wherein the sample image data includes training data and verification data, and matches the target image.

The embodiment of the present application also provides a cloud platform, which includes the above-mentioned tumor prediction device.

Optionally, the cloud platform also includes:

A data management device configured to manage user permissions and received user data, the user data including patient data and user account information.

Optionally, the cloud platform further includes one or more of the following devices:

A resource monitoring device, which is configured to monitor resource usage and network performance parameters according to received monitoring instructions;

A visualization processing device configured to display the received user data, the processing result output by the tumor prediction device, and the constructed nomogram and/or survival curve diagram;

A data storage device configured to store various data output by the data management device and the tumor prediction device;

A control device configured to operate the tumor prediction device, the data management device, the resource monitoring device, the visualization processing device, and the data storage device.

An embodiment of the present application also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, and the computer program can implement the above-mentioned tumor prediction method when the computer program is executed.

It can be seen from the technical solutions provided by the above embodiments of the application that the embodiments of the application predict the tumor classification of the target patient by calling the target prediction model on the cloud platform instead of executing on multiple systems or software, which simplifies the realization of tumor classification prediction The operating environment can improve the accuracy of tumor classification prediction. In addition, the embodiments of the present application not only extract high-dimensional features in the segmented images, but also extract deep features in the segmented images. This takes into account the differences in the features that need to be extracted from images of different tumors or different imaging devices, and fully interprets the tumors. The heterogeneity of this method can fully express the deep information and hidden information of different tumor regions. Therefore, this method has universal applicability. In addition, the tumor prediction method provided by the embodiments of the present application can realize automatic segmentation of images, thereby improving the speed and accuracy of image segmentation, and can also save labor and time costs.

Description of the drawings

In order to more clearly describe the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments described in this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative labor.

Fig. 1 is an application environment diagram of a tumor prediction method in an embodiment of the present application;

Fig. 2 is a schematic flowchart of a tumor prediction method provided in an embodiment of the present application;

Fig. 3 is a schematic structural diagram of a tumor prediction device provided in an embodiment of the present application;

Fig. 4 is a schematic structural diagram of a cloud platform provided in an embodiment of the present application.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only used to explain some of the embodiments of the present application, not all of them. The embodiments are not intended to limit the scope of the application or the claims. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work should fall within the protection scope of this application.

It should be noted that when an element is referred to as being "disposed on" another element, it can be directly disposed on the other element or there may be a centered element. When an element is referred to as being "connected/coupled" to another element, it can be directly connected/coupled to the other element or a central element may be present at the same time. The term "connection/connection" as used herein may include electrical and/or mechanical physical connection/connection. The term "comprising/comprising" as used herein refers to the presence or addition of features, steps or elements, but does not exclude the presence or addition of one or more other features, steps or elements. The term "and/or" as used herein includes any and all combinations of one or more of the associated listed items.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field of this application. The terminology used herein is only for the purpose of describing specific embodiments, and is not intended to limit the application.

In addition, in the description of this application, the terms "first", "second", "third", etc. are only used for the purpose of description and to distinguish similar objects. There is no sequence between the two, nor can they be understood as Indicates or implies relative importance. In addition, in the description of the present application, unless otherwise specified, "plurality" means two or more.

Fig. 1 is an application environment diagram of a tumor prediction method in an embodiment. Referring to Figure 1, this method can be applied to a cloud platform. The cloud platform includes a terminal 100 and a server 200 connected through a network. This method can be executed in the terminal 100 or the server 200. For example, the terminal 100 can directly acquire patient data including image data of a target patient from a medical device, and execute the above method on the terminal side; or, the terminal 100 can also acquire the target patient After the patient data is sent to the server 200, the server 200 obtains the patient data of the target patient and executes the above-mentioned method. The terminal 100 may specifically be a desktop terminal (for example, a desktop computer) or a mobile terminal (for example, a notebook computer or a tablet computer) or the like. The server 200 may be implemented as an independent server or a server cluster composed of multiple servers.

Figure 2 is a tumor prediction method provided in an embodiment of the application. The method can be executed on a cloud platform and can include the following steps:

S1: Obtain the target prediction model.

The target prediction model may be any neural network model used to predict tumor classification, for example, an AlexNet model or a VGGNet model. Among them, the AlexNet model mainly includes an 8-layer structure such as 5-layer convolutional layer and 3-layer fully connected layer; VGGNet model can include 16-layer structure such as 8-layer convolutional layer, 5-layer pooling layer and 3-layer fully connected layer, or It is a 19-layer structure, but it is not limited to this.

After receiving the instruction indicating that tumor prediction is about to be performed, the target prediction model can be obtained from an external device or locally. The external device here can refer to a device external to the cloud platform, correspondingly, the local device can refer to the cloud platform; or, the external device can also refer to the device on the cloud platform other than the tumor prediction device, and correspondingly, the local device can refer to the tumor Forecasting device.

Obtaining the target prediction model locally can include:

(1) Select pre-stored sample image data from a local database or obtain sample image data by processing the received patient data.

Sample image data can include CT (Computed Tomography), MRI (Magnetic Resonance Imaging), PET (Positron Emission Tomography, Positron Emission Tomography), US (ultra sound, ultrasound), SPECT (Single-Photon Emission Computed Tomography) and other image data, and PET/CT (Positron Emission Tomography/Computed Tomography, referred to as PET/CT) and other multimodal image data At least one of. The sample image data can be divided into training data and verification data. Among them, the training data can be used to train the machine learning model, and the verification data can be used to verify the training result of the machine learning model. In the sample image data, the ratio of these two types of data is generally 7:3 or 8:2.

The patient data may include image data of different types of patients and patient case information, such as gender, age, height, weight, and so on.

In one embodiment, a large amount of corresponding image data can be selected from the local database as sample image data according to the received instruction, and most of the image data can be randomly used as training data and the remaining image data as the training data according to the preset ratio. verify the data.

In another embodiment, the received patient data can be formatted according to the received instruction. For example, different types of image data can be parsed into DICOM format; then, the parsed patient data can be retrieved according to a preset standard. Select the corresponding image data from the data as the sample image data. The preset standard may include whether the patient data is complete, whether it has been clinically verified, whether it meets clinical indicators, and so on. For example, it can be judged whether the patient’s clinical data and its case information are not complete. If it is incomplete, the patient’s image data is not selected as the sample image data; it can also be judged whether the patient’s image data has been confirmed by clinical means, such as biopsy testing If it is diagnosed as malignant tumor, the image data of the patient can be selected as sample image data; it can also be judged whether the image data of the patient is judged by the doctor as whether the lesion is too small or abnormal in shape. If it does not meet the requirements, it is not selected. The image data of the patient; the image data of the patient can also be selected according to the actual research needs of the doctor and combined with clinical indicators.

(2) Use the acquired sample image data to train and verify the pre-built machine learning model, and determine the machine learning model that has achieved the best training effect and passed the verification as the target prediction model.

After obtaining the sample image data, the obtained training data can be used to train the machine learning model. Specifically, the segmentation model in the machine learning model can be trained according to the received user instructions, the tumor area and the background area in the sample image data can be separated, and then high-dimensional features and depth features can be extracted from the tumor area, and then the high-dimensional features can be extracted from the tumor area. Dimensional features and depth features are screened, and then the filtered high-dimensional features and deep features are fused to obtain the fused features. Finally, machine learning algorithms such as support vector machine, lasso (LASSO) logistic regression or random forest are used to fuse The feature and the classification label of the tumor (for example, the label of benign tumor is 1, and the label of malignant tumor is 0) are processed, and the features that are highly related to the benign and malignant tumors are selected from the fusion features. At this time, the training effect can be considered as the best Optimal, and determine the various network parameters in the machine learning model.

After it is determined that the training effect of the machine learning model is optimal, the verification data can be used to cross-validate the trained machine learning model with a 50% or 10% discount to calculate the corresponding accuracy, precision, and recall rates. When the obtained accuracy rate, precision rate, and recall rate reach the corresponding preset thresholds, the trained machine learning model can be determined as the target prediction model.

S2: Call the acquired target prediction model to segment the acquired target image of the target patient to obtain a segmented image containing the tumor region.

The target image may refer to an image obtained by scanning a target patient with a medical imaging device, which contains the area where the tumor site of the target patient is located (ie, the tumor area). The target image may include a CT image, an MRI image, a PET image, a US image, a SPECT image, and/or a PET/CT image. The sample image data is matched with the target image, including matching of type and/or content, etc., for example, they are all CT images of the patient, or all lung images.

After acquiring the target prediction model and receiving the target image of the target patient, the acquired target prediction model can be called to segment the acquired target image of the target patient on the tumor region and the background region to extract the tumor region in the target image. Thus, a segmented image containing the tumor area is obtained.

S3: Extract high-dimensional features and depth features from the obtained segmented image.

High-dimensional features may refer to data with a higher dimensionality, which may include at least one of histogram features, three-dimensional shape features, texture features, and filtering features. Among them, the histogram feature can indicate the gray level of the image, which can include features such as maximum, minimum, median, mean, span (maximum-minimum), variance, standard deviation, mean absolute deviation, and/or root mean square. ; Three-dimensional shape features can include volume, surface area, compactness features, sphericity, spherical asymmetry, and/or surface area to volume ratio, etc.; texture features can indicate information about the relative positions of various gray levels of the image, which can include Gray Level Dependence Matrix (GLDM), Gray Level Co-occurence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Area Size Matrix (Gray Level Size Zone Matrix, GLSZM for short) and/or Neighborhood Gray Tone Difference Matrix (NGTDM for short) and other types of features. These features can be constructed based on the spatial distribution of tumor voxel intensity ; Filtering features can be obtained in the following ways: Decompose image texture information through wavelet transform to obtain high-frequency and low-frequency sampled images, and decompose the image into multiple components, that is, they can be respectively along the X, Y, and Z directions High- and low-pass filtering is performed to obtain sub-bands in different directions. For each sub-band, the histogram feature, three-dimensional shape feature, and texture feature are calculated separately to obtain the filtered feature after filtering. The filtering feature can be used to eliminate the noise mixed in the image and improve the clarity of the image.

The depth feature may refer to the hidden information of the image that cannot be seen by the naked eye and cannot be represented by ordinary features. It is a feature that needs to be extracted by calling a deep neural network and can be used to predict tumor classification.

After the segmented image of the target patient is obtained, high-dimensional features and depth features can be extracted from the segmented image.

Regarding the method for extracting high-dimensional features from the image, reference can be made to the description in the prior art, which will not be repeated here.

Regarding the depth feature, the target prediction model can be called to extract from the target image.

In one embodiment, when the target prediction model is the AlexNet model, first, the first convolutional layer in the AlexNet model can be called to perform convolution processing such as local response normalization and pooling on the input target image, and output the result. The extracted feature map; then, the second convolutional layer can be called to perform local response normalization, maximum pooling and other convolution processing on the feature map output by the first convolutional layer and output the corresponding feature map; then, you can sequentially Call the third convolutional layer and the fourth convolutional layer to perform convolution processing and output the corresponding feature maps; call the fifth convolutional layer to directly perform maximum pooling processing on the feature maps output by the fourth convolutional layer; finally call three The fully connected layer performs classification processing to extract depth features and output the extracted depth features.

In one embodiment, when the target prediction model is the VGGNet model, the convolutional layer in the VGGNet model can be called to perform convolution processing on the input target image, then the pooling layer is called for maximum pooling processing, and finally the full connection is called The layer performs classification processing to extract depth features and output the extracted depth features.

It should be noted that the number, size, step length and other settings of the convolution kernels of the multiple network layers of the same type may be the same or different, and there is no limitation here. Moreover, each network layer processes the feature map extracted by the previous network layer connected to it and outputs the extracted feature map to the next network layer connected to it.

S4: Perform screening processing on the extracted high-dimensional features and depth features according to preset conditions.

The preset conditions can be set according to empirical data or actual needs, which can indicate that the selected features have the greatest impact on the prediction target. The greatest impact on the forecast target can be reflected in the relevance, but also in other measurement indicators.

After extracting high-dimensional features and depth features from the target image, you can use sparse representation algorithm, lasso algorithm, Fisher discriminant, feature selection algorithm based on maximum correlation-minimum redundancy, or feature selection algorithm based on conditional mutual information, etc. The algorithm screens high-dimensional features and depth features to filter out the features that meet the preset conditions from these features, that is, to filter out the features that have the greatest impact on the prediction target.

The main idea based on the sparse representation algorithm is that natural signals can be sparsely represented by a dictionary. Generally speaking, the model of the algorithm can be expressed as follows:

Wherein, y is the sparse representation classification label set (e.g., benign, malignant; metastasis, no _{metastasis); D = [d 1,} d 2, K, d i, K, d k] is a high-dimensional feature and depth sparse feature set representation constituting, d _i denotes a characteristic high-dimensional feature and depth features; [alpha] is a coefficient sparse representation, in the form of a matrix, and

It is its estimated value and contains some non-zero elements; μ is a regularization parameter greater than 0, and is used to balance the balance between fidelity and sparsity.

By solving the above formula, α is obtained, and the features whose coefficients are not 0 are selected as high-dimensional features and depth features.

Fisher-based discriminant method is a qualitative classification discrimination method, which is mainly based on the idea of projection. It obtains the feature vector corresponding to the largest feature value in high-dimensional features and depth features, and projects the image data to a feature vector composed of these feature vectors. In this high-dimensional space, the high-dimensional features and depth features with the smallest distance of the same category and the largest distance of different categories are selected in the high-dimensional space.

For related descriptions of other algorithms, reference can be made to the prior art, which will not be repeated here.

By screening high-dimensional features and depth features, redundant features and low-relevance features in these two features can be effectively screened out, and thus the accuracy of the prediction results can be improved.

S5: Call the target prediction model to fuse the selected high-dimensional features and deep features to obtain fusion features, and predict the tumor classification of the target patient based on the fusion features.

After screening the high-dimensional features and depth features that meet the preset conditions, the target prediction model can be called to fuse the screened high-dimensional features and depth features to obtain the fused features.

Regarding how to use the machine learning model to perform fusion processing on image data, reference may be made to related descriptions in the prior art.

After the fusion feature is obtained, the obtained fusion feature can be matched with the preset tumor classification label, so as to predict the tumor classification of the target patient according to the matching result. For example, when the fusion feature matches a benign tumor signature, the patient's tumor can be predicted to be benign; when the fusion feature matches the malignant tumor signature, the patient's tumor can be predicted to be malignant.

It should be noted that the matching relationship between the fusion feature and the tumor classification label may be determined when the machine learning model is trained.

It can be seen from the above description that the embodiment of the present application predicts the tumor classification of the target patient by calling the target prediction model on the cloud platform instead of executing on multiple systems or software, which simplifies the operating environment for achieving tumor classification prediction, and It can prevent data loss and incomplete data information, and can also realize data sharing and improve data processing efficiency. Moreover, the sample image data used comes from multiple imaging equipment or medical institutions, which can make the established target prediction model highly repeatable, versatile and anti-interference, and can be widely used. In addition, the embodiments of the present application not only extract high-dimensional features in the segmented images, but also extract deep features in the segmented images. This takes into account the differences in the features that need to be extracted from images of different tumors or different imaging devices, and fully interprets the tumors. The heterogeneity of this method can fully express the deep information and hidden information of different tumor regions. Therefore, this method has universal applicability. In addition, the tumor prediction method provided by the embodiments of the present application can realize automatic segmentation of images, thereby improving the speed and accuracy of image segmentation, and can also save labor and time costs.

As shown in FIG. 3, an embodiment of the present application also provides a tumor prediction apparatus 300, which may be set on a cloud platform and may include:

An obtaining unit 310, which may be configured to obtain a target prediction model for tumor prediction;

The segmentation unit 320 may be configured to call the acquired target prediction model to segment the acquired target image of the target patient to obtain a segmented image containing the tumor region;

An extraction unit 330, which may be configured to extract high-dimensional features and depth features from the obtained segmented images;

The screening unit 340, which can be configured to screen high-dimensional features and depth features according to preset conditions;

The fusion unit 350 may be configured to call the target prediction model to fuse the selected depth features and high-dimensional features to obtain fusion features;

The prediction unit 360 may be configured to predict the tumor classification of the target patient according to the fusion feature.

In an embodiment, the acquiring unit 310 may also be specifically configured to use the acquired sample image data to train a pre-built machine learning model, and determine the machine learning model that has reached the optimal training effect and passed the verification as the target prediction model.

For the detailed description of the above-mentioned units, reference may be made to the relevant description in the above method embodiments, which will not be repeated here.

By using the tumor prediction device provided in the embodiments of the present application, fully automatic segmentation of tumor regions can be realized, and the accuracy of tumor classification and prediction can be improved, and the diagnosis can be effectively assisted by doctors.

The embodiment of the present application also provides another cloud platform, which may include the tumor prediction device 300 in FIG. 3, and may also include a data management device 100, which may be configured to manage user rights and user data, and the user data includes Patient data and user account information. Specifically, the data management device 100 can manage user permissions based on account authorization information. For example, a user account has authorized N image centers or medical institutions to have upload permissions, but only M image centers or medical institutions are authorized to have operating data. The authorized N imaging centers or medical institutions can upload data to this account, but only M imaging centers or medical institutions have the right to manipulate data, where N and M are both positive integers greater than 1. And N is greater than M. The data management device 100 can also manage the account information registered by the user, and filter the patient data uploaded by the user to filter out the patient data that meets the preset requirements, and can also set the preset size and format to meet the preset requirements. The requested patient data is sent to the data storage device for storage.

In addition, the cloud platform may also include one or more of a resource monitoring device 200, a visualization processing device 400, a data storage device 500, and a control device 600.

The resource monitoring device 200 can be configured to monitor resource usage and network performance parameters, including CPU, memory, GPU, concurrency, bandwidth, packet loss rate, etc., according to received monitoring instructions, and perform according to resource usage The corresponding scheduling.

The visual processing device 400 may display corresponding data according to the received instructions (including user instructions or preset script instructions). For example, the visualization processing device 400 can display the received user data, can also display the processing results (including image segmentation results and/or tumor classification prediction results) output by the tumor prediction device 300, and can also combine the high-dimensional features that have passed the screening with the The combination of imaging omics labels and clinical indicators (for example, age, gender, gene mutation, etc.) obtained by linear combination of in-depth features and their feature coefficients, to construct a personalized and intuitive nomogram and/or survival curve Figures, etc., to effectively assist doctors in medical diagnosis.

Regarding the specific forms of the nomogram and the survival curve diagram, reference can be made to the prior art, which will not be repeated here.

The data storage device 500 may be used to store various data output by the data management device 100 and/or the tumor prediction device 300. The data storage device 500 can be provided with a MySQL database, which can be used to store high-dimensional features, depth features, system dynamics information, DICOM file storage paths, system usage records, and the like. The data storage device 500 supports cloud storage and real-time viewing of raw data and analysis results, as well as data sharing across regions and multiple centers.

The control device 600 can be used to control the operations of the data management device 100, the resource monitoring device 200, the tumor prediction device 300, the visualization processing device 400, and the data storage device 500.

By using the above cloud platform, it is possible to achieve efficient management of image data, accurate prediction of tumor classification, and real-time data sharing.

In one embodiment, the present application also provides a computer-readable storage medium in which a computer program is stored, and when the computer program is executed, the corresponding function described in the foregoing method embodiment can be realized. The computer program can also be run on the terminal or server as shown in Figure 1.

A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium. Here, when the program is executed, it may include the procedures of the above-mentioned method embodiments. Wherein, any references to memories, storage media, databases or other media used in the embodiments provided in this application may include non-volatile and/or volatile memories. Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not a limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The systems, equipment, devices, units, etc. described in the foregoing embodiments may be specifically implemented by semiconductor chips, computer chips, and/or entities, or implemented by products with certain functions. For the convenience of description, when describing the above device, the functions are divided into various units and described separately. Of course, when implementing this application, the functions of each unit can be implemented in the same or multiple chips.

Although the present application provides the method operation steps described in the above-mentioned embodiments or flowcharts, more or fewer operation steps may be included in the method based on conventional or no creative labor. In steps where there is no necessary causal relationship logically, the execution order of these steps is not limited to the execution order provided in the embodiments of the present application.

The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In addition, the technical features of the above embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, All should be considered as the scope of this specification.

The above-mentioned embodiments are described to facilitate those skilled in the art to understand and use this application. It is obvious that those skilled in the art can easily make various modifications to these embodiments, and apply the general principles described here to other embodiments without creative work. Therefore, this application is not limited to the above-mentioned embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of this application without departing from the scope of this application should fall within the protection scope of this application.

Claims

A tumor prediction method, characterized in that the tumor prediction method is executed on a cloud platform, and includes:

Calling the acquired target prediction model to segment the acquired target image of the target patient to obtain a segmented image containing the tumor area;

Extracting high-dimensional features and depth features from the obtained segmented image;

Screening the high-dimensional features and the depth features according to preset conditions;

The target prediction model is called to fuse the selected depth features and the high-dimensional features to obtain fusion features, and the tumor classification of the target patient is predicted according to the fusion features.
The tumor prediction method according to claim 1, wherein the target prediction model is obtained in the following manner:

Obtain the target prediction model from an external device or locally.
The tumor prediction method according to claim 2, wherein obtaining the target prediction model locally comprises:

Training and verifying a pre-built machine learning model by using the acquired sample image data, the sample image data including training data and verification data, and matching the target image;

The machine learning model that has achieved the optimal training effect and passed the verification is determined as the target prediction model.
The tumor prediction method according to claim 3, characterized in that, before training the machine learning model, the tumor prediction method comprises:

Select the pre-stored sample image data from a local database; or

The sample image data is obtained by processing the received patient data.
The tumor prediction method according to claim 4, wherein obtaining the sample image data by processing the received patient data comprises:

Analyze the format of the received patient data;

The sample image data is selected from the parsed patient data according to a preset standard.
The tumor prediction method according to claim 5, wherein the preset criteria include whether the patient data is complete, whether it has been clinically verified, and whether it meets clinical indicators.
The tumor prediction method according to claim 1, wherein the screening of the depth features and the high-dimensional features extracted from the target image according to preset conditions comprises:

Use sparse representation algorithm, lasso algorithm, Fisher discriminant, feature selection algorithm based on maximum correlation-minimum redundancy, or feature selection algorithm based on conditional mutual information to filter the high-dimensional features and the depth features to filter out those that satisfy The high-dimensional features and depth features of the preset conditions.
The tumor prediction method according to claim 1, wherein the target image includes a CT image, an MRI image, a PET image, a US image, a SPECT image, and/or a PET/CT image.
The tumor prediction method according to claim 1, wherein the target prediction model comprises an AlexNet model or a VGGNet model.
A tumor prediction device, characterized in that the tumor prediction device is set on a cloud platform and includes:

A segmentation unit configured to call the acquired target prediction model to segment the acquired target image of the target patient to obtain a segmented image containing the tumor region;

An extraction unit configured to extract high-dimensional features and depth features from the obtained segmented image;

A screening unit configured to screen the high-dimensional features and the depth features according to preset conditions;

A fusion unit configured to call the target prediction model to fuse the selected depth feature and the high-dimensional feature to obtain a fusion feature;

A prediction unit configured to predict the tumor classification of the target patient according to the fusion feature.
The tumor prediction device according to claim 10, wherein the tumor prediction device further comprises:

The acquiring unit is configured to acquire the target prediction model by using the acquired sample image data to train and verify the pre-built machine learning model, and to optimize the training effect and pass the verification. A machine learning model is determined as the target prediction model, wherein the sample image data includes training data and verification data, and matches the target image.
A cloud platform, wherein the cloud platform comprises the tumor prediction device according to any one of claims 10-11.
The cloud platform according to claim 12, wherein the cloud platform further comprises:

A data management device configured to manage user permissions and received user data, the user data including patient data and user account information.
The cloud platform according to claim 13, wherein the cloud platform further comprises one or more of the following devices:

A resource monitoring device, which is configured to monitor resource usage and network performance parameters according to received monitoring instructions;

A visualization processing device configured to display the received user data, the processing result output by the tumor prediction device, and the constructed nomogram and/or survival curve diagram;

A data storage device configured to store various data output by the data management device and the tumor prediction device;

A control device configured to operate the tumor prediction device, the data management device, the resource monitoring device, the visualization processing device, and the data storage device.
A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, which can realize the tumor prediction method according to any one of claims 1 to 9 when the computer program is executed.