Artificial intelligence-based medical image processing method, apparatus, and electronic device

Abstract

The application provides an artificial intelligence-based medical image processing method and device, electronic equipment and computer readable storage medium; the method comprises: obtaining a medical image processing request comprising a target object identifier; according to the medical image processing request, extracting a region of interest matching a target event type from a medical image corresponding to the target object identifier; extracting a candidate image feature corresponding to the target event type from the region of interest; performing feature screening processing on the candidate image feature to obtain a target prediction image feature corresponding to the target event type; performing development information prediction processing based on the target prediction image feature of the target event type to obtain event development information corresponding to the target object identifier. Through the application, development information prediction can be accurately performed according to a medical image.

Description

TECHNICAL FIELD

[0001] The present application relates to artificial intelligence technology, and in particular to a medical image processing method and device based on artificial intelligence, an electronic device, and a computer readable storage medium. BACKGROUND

[0002] Artificial intelligence (AI) is a comprehensive technology of computer science, which studies the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, such as natural language processing technology and machine learning / deep learning. With the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0003] Medical image-based analysis is one of the important applications in the field of artificial intelligence. A medical image analysis system is a system that uses a computer to process, analyze and understand medical images to identify various different patterns of targets and objects.

[0004] In related technologies, when analyzing medical images, the intermediate results related to the medical images cannot be accurately predicted.

[0005] SUMMARY

[0006] The embodiments of the present application provide a medical image processing method and device based on artificial intelligence, an electronic device, and a computer readable storage medium, which can accurately predict development information according to medical images.

[0007] The technical solutions of the embodiments of the present application are as follows:

[0008] The embodiments of the present application provide a medical image processing method based on artificial intelligence, comprising:

[0009] Obtaining a medical image processing request including a target object identifier;

[0010] According to the medical image processing request, extracting a region of interest matching a target event type from a medical image corresponding to the target object identifier;

[0011] Extracting a candidate image feature corresponding to the target event type from the region of interest;

[0012] Performing feature screening processing on the candidate image feature to obtain a target prediction image feature corresponding to the target event type;

[0013] Performing development information prediction processing based on the target prediction image feature corresponding to the target event type to obtain event development information corresponding to the target object identifier.

[0014] The embodiment of the application provides a medical image processing device based on artificial intelligence, which comprises:

[0015] The region module is configured to acquire a medical image processing request comprising a target object identifier; and extract, according to the medical image processing request, a region of interest corresponding to a target event type from a medical image corresponding to the target object identifier.

[0016] The feature module is configured to extract candidate image features corresponding to the target event type from the region of interest.

[0017] The selection module is configured to perform feature screening processing on the candidate image features corresponding to the target event type to obtain target prediction image features corresponding to the target event type.

[0018] The prediction module is configured to perform development information prediction processing based on the target prediction image features corresponding to the target event type to obtain event development information corresponding to the target object identifier.

[0019] In the above scheme, the region module is further configured to: in response to a cropping operation on the medical image, collect an event region corresponding to the target event type from the medical image corresponding to the target object identifier, and determine the event region as the region of interest corresponding to the target event type.

[0020] In the above scheme, the region module is further configured to: call a first machine learning model to perform segmentation processing on the medical image corresponding to the target object identifier to obtain an event region corresponding to the target event type, and determine the event region as the region of interest corresponding to the target event type.

[0021] In the above scheme, the first machine learning model comprises N cascaded down-sampling networks, a convolution network and M cascaded up-sampling networks; N and M are integers greater than or equal to 2; the region module is further configured to: perform down-sampling processing on the medical image through the N cascaded down-sampling networks to obtain a down-sampling processing result comprising features of different parts in the medical image; perform convolution processing on the down-sampling processing result through the convolution network to obtain convolution features corresponding to the medical image; and perform up-sampling processing on the convolution features through the M cascaded up-sampling networks to obtain a region corresponding to an event manifestation comprising the target event type in the medical image as the corresponding event region; wherein the event manifestation comprises at least one of the following: a blood vessel manifestation, a bone manifestation and a space-occupying manifestation.

[0022] In the above scheme, the feature module is further configured to: acquire attribute data of a candidate image feature corresponding to the target event type, wherein the attribute data comprises at least one of the following: a type of the candidate image feature, a dimension of the candidate image feature; and extract at least one of the following features of an event manifestation corresponding to the target event type from a region of interest of the medical image: a histogram intensity candidate image feature, a shape candidate image feature, a texture candidate image feature, and a wavelet candidate image feature, which are consistent with the attribute data.

[0023] In the above scheme, the selection module is further configured to: acquire event history features corresponding to the target object identifier and the target event type; when an information amount of the event history features is greater than an information amount threshold, perform feature screening processing on the event history features and the candidate image features to obtain target prediction image features corresponding to the medical image; and when the information amount of the event history features is not greater than the information amount threshold, perform feature screening processing on the candidate image features to obtain the target prediction image features corresponding to the medical image.

[0024] In the above scheme, the selection module is further configured to: determine an importance degree of each of the candidate image features for the development information prediction processing; and determine target prediction image features corresponding to the target event type from the candidate image features based on the importance degree of each of the candidate image features for the development information prediction processing.

[0025] In the above scheme, the selection module is further configured to: determine a correlation degree between each of the candidate image features and event development information corresponding to the target event type; wherein the correlation degree comprises at least one of the following: a Pearson coefficient between the candidate image feature and the event development information, mutual information between the candidate image feature and the event development information, and an analysis of variance ratio between the candidate image feature and the event development information; and determine an importance degree positively correlated with the correlation degree.

[0026] In the above scheme, the time prediction processing is implemented by a second machine learning model comprising a plurality of decision trees; and the selection module is further configured to: determine a first error rate of each of the candidate image features by each of the decision trees in the second machine learning model, so as to determine an importance degree of each of the candidate image features among all the candidate image features based on the first error rate.

[0027] In the above scheme, the selection module is further configured to: sort each of the candidate image features in descending order based on the corresponding importance degree, and select a plurality of candidate image features with high ranking as candidate target prediction image features; when the number of the candidate target prediction image features is not greater than a target prediction image feature threshold, determine the candidate target prediction image features as target prediction image features corresponding to the target event type; when the number of the candidate target prediction image features is greater than the target prediction image feature threshold, update the second machine learning model based on the candidate image features, so as to continue to determine the importance degree of each of the candidate target prediction image features by using the updated second machine learning model, and delete the candidate target prediction image features with low ranking.

[0028] In the above scheme, the selection module is further configured to: for each of the decision trees, perform the following processing: obtaining medical image out-of-bag data corresponding to the decision tree, determining a first error rate of development information prediction processing by using the medical image out-of-bag data, performing noise interference processing on each of the candidate image features of the medical image out-of-bag data to obtain medical noise image out-of-bag data corresponding to each of the candidate image features after noise interference, and determining a second error rate of development information prediction processing by using each of the medical noise image out-of-bag data; for each of the candidate image features, perform the following processing: determining noise influence of any one of the decision trees, wherein the noise influence is a difference between the first error rate and the second error rate corresponding to the decision tree, and the second error rate corresponds to the decision tree and the candidate image feature; and performing average processing on the differences of a plurality of decision trees to obtain a specificity degree of each of the candidate image features.

[0029] In the above scheme, the time prediction processing is implemented by using a second machine learning model including a plurality of decision trees; and the prediction module is further configured to: map the target prediction image features to event development information by using each of the decision trees; and perform average processing on the event development information obtained by using each of the decision trees, and take the average processing result as the event development information.

[0030] In the foregoing scheme, the development information prediction processing is implemented through a second machine learning model including multiple decision trees; the apparatus further includes a training module configured to: before obtaining event development information corresponding to a target object identifier through development information prediction processing based on target prediction image features of the target event type, acquire a medical image sample having a pre-labeled event development information and a pre-labeled target event type; extract a region of interest sample corresponding to the pre-labeled target event type from the medical image sample; acquire candidate image feature samples corresponding to the region of interest sample; perform feature selection processing on the candidate image feature samples of the region of interest to obtain target prediction image feature samples corresponding to the medical image sample; perform forward propagation of the target prediction image feature samples in the second machine learning model, and perform backward propagation of the pre-labeled event development information in the second machine learning model to update parameters of the second machine learning model.

[0031] An electronic device is provided in an embodiment of the application, and the electronic device includes:

[0032] A memory configured to store executable instructions.

[0033] A processor configured to execute the executable instructions stored in the memory to implement the medical image processing method based on artificial intelligence provided in the embodiments of the application.

[0034] A computer-readable storage medium is provided in an embodiment of the application, and the computer-readable storage medium stores executable instructions, and the executable instructions are configured to be executed by a processor to implement the medical image processing method based on artificial intelligence provided in the embodiments of the application.

[0035] The embodiments of the application have the following beneficial effects:

[0036] By extracting effective features in the statistical features of the region of interest of the medical image as the basis for development information prediction, on the one hand, the calculation resource consumption and feature engineering consumption of the prediction process are effectively saved, and on the other hand, the effective features can accurately reflect the development of the disease type, thereby ensuring the accuracy of the predicted development information. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 FIG. 1 is a schematic diagram of an application scenario of a medical image processing system provided in an embodiment of the application;

[0038] Figure 2 FIG. 2 is a structural schematic diagram of an electronic device for medical image processing provided in an embodiment of the application;

[0039] Figures 3-5 FIG. 3 is a flowchart of a medical image processing method provided in an embodiment of the application;

[0040] Figure 6 is a medical image schematic diagram provided by an embodiment of the application;

[0041] Figure 7 is a medical image schematic diagram provided by an embodiment of the application;

[0042] Figure 8 is a flow schematic diagram of a medical image processing method provided by an embodiment of the application. DETAILED DESCRIPTION

[0043] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings, and the described embodiments should not be regarded as limiting the present application. All other embodiments obtained by those of ordinary skill in the art without making creative labor fall within the scope of protection of the present application.

[0044] In the following description, "some embodiments" are described, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0045] In the following description, the terms "first\second" are only to distinguish similar objects, and do not represent a specific order of the objects. It can be understood that "first\second" can be interchanged in a specific order or sequence as allowed, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0047] The related data collection and processing in the embodiments of the present application should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of authorization of laws and regulations and the personal information subject.

[0048] Before the embodiments of the present application are further described in detail, the terms and terms involved in the embodiments of the present application are explained, and the terms and terms involved in the embodiments of the present application are applicable to the following explanations.

[0049] 1) Deep Learning (DL): a new research direction in the field of Machine Learning (ML), which is introduced into ML to make it closer to the original goal - Artificial Intelligence (AI). Deep learning is the learning of the internal rules and representation hierarchy of sample data, and the information obtained in this learning process is very helpful for the interpretation of data such as text, images and sound. The ultimate goal of deep learning is to enable machines to have the ability to analyze and learn like humans, and to be able to recognize data such as text, images and sound.

[0050] 2) Magnetic Resonance Imaging (MRI): using the principle of nuclear magnetic resonance, based on the different decay of released energy in different structural environments inside the material, through the detection of electromagnetic waves emitted by the application of gradient magnetic field, the position and type of atomic nucleus constituting the object can be determined, and accordingly the structural image of the object inside can be drawn.

[0051] 3) Computed Tomography (CT): a technique that irradiates the head with X-rays from multiple directions along a selected tomographic layer, measures the amount of X-rays transmitted, digitizes and calculates the absorption coefficient of each unit volume of the layer by computer, and then reconstructs the image.

[0052] 4) Imageomics: high-throughput extraction of a large number of high-dimensional quantitative image features from medical images such as MRI and CT, and analysis.

[0053] 5) Neural Network: a nonlinear, self-adaptive information processing system composed of a large number of processing units interconnected.

[0054] 6) Transverse plane: a section that divides the patient into two parts in the vertical direction.

[0055] 7) Lasso (Least absolute shrinkage and selection operator): this method is a compression estimate. It obtains a more refined model by constructing a penalty function, so that it compresses some regression coefficients, i.e. forces the sum of the absolute values of the coefficients to be less than a fixed value; at the same time, some regression coefficients are set to zero. Therefore, it retains the advantage of subset shrinkage and is a biased estimate for handling data with complex collinearity.

[0056] In an actual medical scenario, the target object (patient) may lose consciousness or be unaccompanied during the onset and subsequent development of the disease, and thus cannot provide accurate event development information such as the onset time. In the related art, no method for directly predicting event development information of a disease is found. Through imageomics, it can be determined whether the event development information of the target object is within a specific time window. For example, when the event development information is the onset time, the onset time is predicted to be within 4.5 hours, that is, only whether the onset time is within a certain range can be determined, and any time stamp in 4.5 hours may be the event development information, but an accurate event development information cannot be given. Moreover, when the time range is determined through influenceomics, it is only effective for a specific disease type and does not have universality.

[0057] The embodiment of the present application provides a medical image processing method and device based on artificial intelligence, electronic equipment and computer readable storage medium, which can accurately predict the event development information of multiple diseases according to the medical image.

[0058] The medical image processing method provided in the embodiment of the present application can be realized by a terminal / server alone; or can be realized by a terminal and a server in cooperation, for example, the terminal alone bears the medical image processing method described below, or the terminal sends the medical image of the target object to the server, and the server executes the medical image processing method according to the received medical image of the target object.

[0059] The electronic equipment for medical image processing provided in the embodiment of the present application can be various types of terminals or servers, wherein the server can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services; the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application.

[0060] Taking the server as an example, for example, it can be a server cluster deployed in the cloud, which opens an artificial intelligence cloud service (AiaaS, AI as a Service) to users. The platform will split several common AI services and provide independent or packaged services in the cloud. This service mode is similar to an AI theme mall, and all users can access one or more artificial intelligence services provided by the AIaaS platform through the application programming interface.

[0061] For example, one of the artificial intelligence cloud services can be a medical image processing service, that is, the server in the cloud encapsulates the medical image processing program provided in the embodiments of the present application. The user calls the medical image processing service in the cloud service through the terminal (running a client, such as a map positioning client, a medical client, etc.) to make the server deployed in the cloud call the encapsulated medical image processing program to obtain a medical image processing request including a target object identifier; extract a region of interest matching the target event type from the medical image corresponding to the target object identifier according to the medical image processing request; extract candidate image features of the target event type from the region of interest; perform feature screening processing on the candidate image features to obtain target prediction image features of the target event type; perform development information prediction processing based on the target prediction image features of the target event type to obtain event development information corresponding to the target object identifier, for example, event development information of a heart disease, etc. The medical image processing method provided in the embodiments of the present application is not for a living or animal body as an object, and is not directly aimed at obtaining a disease diagnosis result or a health condition. The disease diagnosis result or the health condition cannot be directly obtained according to the event development information of the disease type, and is only used to assist doctors and researchers in disease diagnosis, re-examination, and research on treatment methods.

[0062] Referring to Figure 1 , Figure 1 FIG. 1 is a schematic diagram of an application scenario of a medical image processing system provided in the embodiments of the present application. The terminal 400 is connected to the server 200 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0063] The terminal 400 (running a client, such as a medical client, etc.) can be used to obtain a medical image. For example, the terminal 400 collects a medical image through an image collection device of the terminal 400, or sends a collected medical image to the terminal 400 through other image collection equipment, and the terminal 400 receives the medical image.

[0064] In some embodiments, a medical image processing plug-in can be implanted in the client running in the terminal to implement a medical image processing method locally in the client. For example, after the terminal 400 obtains an event development information prediction request (including a medical image) based on a target event type of the medical image, the terminal 400 calls an image processing plug-in to implement an image processing method to predict event development information of the target event type according to target prediction image features of a region of interest corresponding to the target event type in the medical image.

[0065] In some embodiments, after the terminal 400 acquires the event development information prediction request (including the medical image) based on the target event type of the medical image, the image processing interface of the server 200 (which can be provided in the form of a cloud service, i.e., an image processing service) is called, and the server 200 predicts the event development information of the corresponding target event type according to the target prediction image features of the region of interest corresponding to the target event type in the medical image. For example, after the medical image is input to the medical application by the patient, the doctor, or the researcher, the medical application automatically acquires the event development information prediction request (including the medical image) based on the target event type of the medical image, and then calls the image processing interface of the server 200 to predict the event development information of the corresponding target event type for the medical image. The medical image processing method provided in the embodiments of the present application is not for a living or animal body as the object, and is not directly aimed at obtaining a disease diagnosis result or a health condition. The disease diagnosis result or the health condition cannot be directly obtained according to the event development information of the target event type, i.e., the event development information of the target event type is not directly used for disease diagnosis, but only serves as intermediate data to assist the patient in predicting the disease, and to assist the doctor, the researcher in diagnosing the disease, reexamining the disease, and researching the treatment method.

[0066] The structure of the electronic device for medical image processing provided in the embodiments of the present application will be described below. Referring to Figure 2 , Figure 2 FIG. 1 is a structural schematic diagram of an electronic device for medical image processing provided in the embodiments of the present application, which takes the server 200 as an example. Figure 2 The server 200 for medical image processing shown in FIG. 1 includes at least one processor 210, a memory 250, at least one network interface 220, and a user interface 230. The various components in the server 200 are coupled together through a bus system 240. It can be understood that the bus system 240 is used to realize the connection and communication between the components. The bus system 240 includes not only a data bus, but also a power bus, a control bus, and a status signal bus. However, for the purpose of clear illustration, all the buses are marked as the bus system 240 in Figure 2 .

[0067] The processor 210 can be an integrated circuit chip with a signal processing capability, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0068] The memory 250 includes volatile memory or nonvolatile memory, and can include both volatile and nonvolatile memory. Among them, the nonvolatile memory can be a read only memory (ROM), and the volatile memory can be a random access memory (RAM). The memory 250 described in the embodiments of the present application is intended to include any suitable type of memory. The memory 250 can optionally include one or more storage devices that are physically located away from the processor 210.

[0069] In some embodiments, the memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, which are exemplarily illustrated below.

[0070] The operating system 251 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks.

[0071] The network communication module 252 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including Bluetooth, wireless compatibility authentication (WiFi), and universal serial bus (USB), etc.

[0072] In some embodiments, the medical image processing apparatus provided by the embodiments of the present application can be implemented in a software manner, for example, it can be a medical image processing plug-in in the terminal described above, or a medical image processing service in the server described above. Of course, it is not limited thereto, and the medical image processing apparatus provided by the embodiments of the present application can be provided in various software embodiments, including various forms such as application programs, software, software modules, scripts or codes.

[0073] Figure 2 The medical image processing apparatus 255 stored in the memory 250 is shown, which can be software in the form of programs and plug-ins, such as image processing plug-ins, and includes a series of modules, including a region module 2551, a feature module 2552, a selection module 2553, a prediction module 2554, and a training module 2555.

[0074] As before, the medical image processing method provided by the embodiments of the present application can be implemented by various types of electronic devices. Referring to Figure 3 , Figure 3 is a flowchart of the medical image processing method provided by the embodiments of the present application, which is described in conjunction with the steps shown in Figure 3 .

[0075] In step 101, a medical image processing request including a target object identifier is acquired to extract a region of interest matching a target event type from a medical image corresponding to the target object identifier according to the medical image processing request.

[0076] As an example, the target object is a patient, such as a human or an animal, the target object identifier is identity information of the target object, the medical image can be a medical image, such as a medical image obtained by computed tomography, a medical image obtained by magnetic resonance imaging, etc., the target event type can be an event type in a medical scenario, such as a disease type such as hemorrhagic stroke or pancreatic cancer, and the region of interest is a region most closely associated with the target event type.

[0077] In some embodiments, the extraction of the region of interest matching the target event type from the medical image corresponding to the target object identifier in step 101 can be achieved by the following technical solution: in response to a cropping operation on the medical image, an event region corresponding to the target event type is collected from the medical image corresponding to the target object identifier, and the region of interest corresponding to the target event type is determined.

[0078] As an example, the cropping operation on the medical image can be an operation initiated by medical personnel or researchers, and the location data of the region of interest in the medical image is carried in the cropping operation. According to the location data, an event region (for example, a lesion region) corresponding to the target event type is collected from the medical image. For example, for hemorrhagic stroke, the bleeding point is the event region. The event region is determined as the region of interest corresponding to the target event type, and the region of interest is displayed in a display manner different from other regions to distinguish the region of interest from other regions in the medical image except the region of interest. See Figure 6 , Figure 6 is a medical image diagram provided by an embodiment of the present application, Figure 6 shows a medical image 601 with a target event type of hemorrhagic stroke. See Figure 7 , Figure 7 is a medical image diagram provided by an embodiment of the present application, Figure 7 in 701 is a region of interest of a medical image with a target event type of hemorrhagic stroke, and the region of interest is presented in a shadow manner to facilitate the later differentiation of the region of interest by medical personnel or researchers.

[0079] In some embodiments, the extraction of the region of interest matching the target event type from the medical image corresponding to the target object identifier in step 101 can be achieved by the following technical solution: a first machine learning model is called to perform segmentation processing on the medical image to obtain an event region corresponding to the target event type, and the region of interest corresponding to the target event type is determined.

[0080] As an example, the first machine learning model can be various convolutional neural network models, such as a U-shaped convolutional neural network, when the first machine learning model is a U-shaped convolutional neural network, the first machine learning model includes N cascaded down-sampling networks, a convolutional network, and M cascaded up-sampling networks; N and M are integers greater than or equal to 2.

[0081] In some embodiments, the above-mentioned calling the first machine learning model to segment the corresponding medical image of the target object to obtain an event region corresponding to the target event type can be implemented by the following technical solutions: performing down-sampling processing on the medical image through N cascaded down-sampling networks to obtain a down-sampling processing result including features of different parts in the medical image; performing convolution processing on the down-sampling processing result through the convolutional network to obtain convolutional features corresponding to the medical image; performing up-sampling processing on the convolutional features through M cascaded up-sampling networks to obtain a region in the medical image corresponding to an event manifestation (e.g., a lesion manifestation) including the target event type as the corresponding event region; wherein the event manifestation includes at least one of the following: a blood vessel manifestation, a bone manifestation, and an occupying manifestation. The U-shaped network structure is formed by the multiple down-sampling networks and the multiple up-sampling networks, and the connection operation is used between the networks in the same level instead of directly supervising the high-level semantic features, which ensures that the feature map recovered finally fuses more low-level candidate image features, and also makes the features of different scales be fused, which is beneficial to accurately predicting the event development information subsequently.

[0082] As an example, through the nth down-sampling network in the N cascaded down-sampling networks, the input of the nth down-sampling network is down-sampled and processed, and the nth down-sampling processing result output by the nth down-sampling network is transmitted to the (n+1)th down-sampling network to continue the down-sampling processing to obtain the (n+1)th down-sampling processing result; wherein n is an integer variable starting from 1 and increasing, and the value range of n is n When n is 1, the input of the nth down-sampling network is the medical image, when n is 2 n , the input of the nth down-sampling network is the (n-1)th down-sampling processing result output by the (n-1)th down-sampling network.

[0083] As an example, through the mth upsampling network in the M cascaded upsampling networks, the input of the mth upsampling network is upsampled to obtain an upsampling result corresponding to the mth upsampling network, and the upsampling result corresponding to the mth upsampling network is transmitted to the (m+1)th upsampling network to continue the upsampling process to obtain an upsampling result corresponding to the (m+1)th upsampling network; wherein m is an integer variable starting from 1 and increasing, and the value of m ranges from 1 to M. m When m is 1, the input of the mth sampling network is the convolution feature, and when m is 2 , the input of the mth upsampling network is the (m-1)th upsampling processing result output by the (m-1)th upsampling network.

[0084] As an example, before the mth upsampling network in the M cascaded upsampling networks is used to upsample the input of the mth upsampling network to obtain an upsampling result corresponding to the mth upsampling network, when m is , the output of the Mth downsampling network is spliced with the output of the convolution network, and the spliced result is used as the input of the mth upsampling network to replace the output of the convolution network as the input of the mth upsampling network; when m is , the output of the (M+1-m)th downsampling network is spliced with the output of the (m-1)th upsampling network, and the spliced result is used as the input of the mth upsampling network to replace the upsampling processing result corresponding to the (m-1)th upsampling network as the input of the mth upsampling network.

[0085] As an example, before the first machine learning model is called to segment the medical image to obtain the region of interest corresponding to the target event type, the medical image sample is forward propagated in the first machine learning model to obtain a predicted region of interest corresponding to the target event type; the predicted region of interest and the pre-labeled region of interest corresponding to the target event type are substituted into the image segmentation loss function to determine the parameters of the first machine learning model when the image segmentation loss function reaches the minimum value.

[0086] In step 102, candidate image features corresponding to the target event type are extracted from the region of interest.

[0087] As an example, the candidate image features are image features of the region of interest in the medical image that have relevance to the target event type.

[0088] In some embodiments, referring to Figure 4 , Figure 4 is a medical image processing method based on artificial intelligence provided by the embodiments of the present application, and the candidate image features corresponding to the target event type are extracted from the region of interest in step 102, which can be performed through steps 1021-1022.

[0089] In step 1021, attribute data of the candidate image feature corresponding to the target event type is obtained.

[0090] As an example, the attribute data includes at least one of the following: type of the candidate image feature, dimension of the candidate image feature, the attribute data defines the type of the candidate image feature and the dimension of the candidate image feature of each type.

[0091] In step 1022, at least one of the following features of the event performance corresponding to the target event type is extracted from the region of interest of the medical image: histogram intensity candidate image feature, shape candidate image feature, texture candidate image feature, wavelet candidate image feature.

[0092] As an example, when statistical calculation is performed on the intensity information of the medical image in the region of interest, due to the presence of the event performance (e.g., lesion performance) corresponding to the target event type in the region of interest, a candidate image feature with statistical significance can be obtained, and the obtained candidate image feature has a strong correlation with the lesion performance corresponding to the target event type. The candidate image feature includes at least one of the following: histogram intensity candidate image feature, shape candidate image feature, texture candidate image feature, wavelet candidate image feature, etc. The histogram intensity candidate image feature represents the structural features of the edge (gradient), so it can describe the local shape information. The histogram intensity candidate image feature is obtained based on the quantization of the position and direction space, which can suppress the effects of translation and rotation to a certain extent. The histogram intensity candidate image feature is obtained based on the processing method of block division, so that the relationship between the local pixels of the region of interest of the medical image can be well represented. The shape candidate image feature is described by length, width, circularity, and moment invariants. The texture candidate image feature is a visual feature that reflects the homogeneity phenomenon in the medical image. It embodies the surface structure organization arrangement properties with slow change or periodic change of the object surface. Texture has three major signs: certain local sequence repetition, non-random arrangement, and uniformity within the texture region. Texture is different from image features such as gray scale and color, which are represented by the gray scale distribution of pixels and their surrounding spatial neighborhood.

[0093] In step 103, the candidate image feature is subjected to feature screening processing to obtain a target prediction image feature corresponding to the target event type.

[0094] As an example, the target prediction image feature is screened from the candidate image feature and can be an effective feature for development information prediction processing.

[0095] In some embodiments, referring to Figure 5 , Figure 5The medical image processing method based on artificial intelligence is provided by the embodiment of the present application, and the candidate image features of the target event type are subjected to feature screening processing in step 103 to obtain the target prediction image features corresponding to the target event type, which can be realized through steps 1031-1032.

[0096] In step 1031, the event history features corresponding to the target object identifier and the target event type are obtained.

[0097] As an example, when the medical record information corresponding to the target object identifier and the target event type (for example, the disease type) is available, the event history features (for example, the medical record features) corresponding to the medical record information can be combined with the candidate image features to perform feature screening processing based on the combination of the event history features and the candidate image features, and the event history features include at least one of the following: admission diastolic blood pressure, admission systolic blood pressure, whether drinking or smoking, relevant medical history, and the like.

[0098] In step 1032, when the information amount of the event history features is greater than the information amount threshold, the event history features and the candidate image features are subjected to feature screening processing to obtain the target prediction image features corresponding to the medical image.

[0099] As an example, when the information amount of the event history features is greater than the information amount threshold, the number of event history features is greater than the number threshold and / or the type of event history features is greater than the type threshold, then the event history features and the candidate image features are both used as the basis for feature screening processing, and the subsequent event development information prediction can be based on the event history features and the candidate image features, so that the event development information is predicted based on more abundant features, and the prediction accuracy is improved.

[0100] In step 1033, when the information amount of the event history features is not greater than the information amount threshold, the candidate image features are subjected to feature screening processing to obtain the target prediction image features corresponding to the medical image.

[0101] As an example, when the information amount of the event history features is not greater than the information amount threshold, the number of event history features is not greater than the number threshold and the type of event history features is not greater than the type threshold, then only the candidate image features are used as the basis for feature screening processing, and the subsequent event development information prediction is only based on the candidate image features, so as to save the cost of feature engineering and reduce the consumption of computing resources.

[0102] In some embodiments, the feature screening processing on the candidate image features in step 103 to obtain the target prediction image features corresponding to the target event type can be implemented through the following technical solutions: determining an importance degree of each candidate image feature corresponding to the development information prediction processing; and determining the target prediction image features corresponding to the target event type from the candidate image features based on the importance degree of each candidate image feature corresponding to the development information prediction processing. Through the target prediction image features, the event development information is predicted, on the one hand, the calculation resource consumption and the feature engineering consumption of the prediction process are effectively saved, and on the other hand, the target prediction image features can accurately reflect the development of the target event type, thereby ensuring the accuracy of the predicted development time.

[0103] As an example, the event development information can be medical history information or medical future information, wherein the medical history information includes the onset time, the onset state and the like of a disease, and the medical future information includes the time of the disease reaching a future stage, for example, the medical image is taken in the early stage of cancer, and the medical future information includes the time of the middle stage of cancer, the state of the middle stage of cancer and the like.

[0104] As an example, the two steps of determining the importance degree of each candidate image feature corresponding to the development information prediction processing, and determining the target prediction image features corresponding to the target event type from the candidate image features based on the importance degree of each candidate image feature corresponding to the development information prediction processing are also applicable to the combination of the candidate image features and the event history features, that is, determining the importance degree of each candidate image feature and the event history feature corresponding to the development information prediction processing, and determining the target prediction image features corresponding to the target event type from the candidate image features and the event history features based on the importance degree of each candidate image feature and each event history feature corresponding to the development information prediction processing.

[0105] In some embodiments, the determination of the importance degree of each feature corresponding to the development information prediction processing can be implemented through the following technical solutions: determining a correlation degree between each candidate image feature and the event development information corresponding to the target event type; wherein the correlation degree includes at least one of the following: a Pearson coefficient between the candidate image feature and the event development information; mutual information between the candidate image feature and the event development information; analysis of variance ratio between the candidate image feature and the event development information; and determining the importance degree positively correlated with the correlation degree.

[0106] As an example, the correlation degree index of each candidate image feature and the event development information is independently calculated to ensure that each candidate image feature has strong reference significance in the event development information prediction task. The index of independent feature selection includes at least one of the following: mutual information, Pearson coefficient, analysis variance ratio (ANOVA F-value), and the like. The mutual information is a measure of the mutual dependence between the candidate image feature and the event development information. The Pearson coefficient is a linear correlation coefficient. The Pearson correlation coefficient is a statistical quantity used to reflect the linear correlation degree of the candidate image feature and the event development information.

[0107] In some embodiments, the development information prediction processing is implemented by a second machine learning model including a plurality of decision trees. The determination of the importance of each candidate image feature for the development information prediction processing can be implemented by the following technical solution: determining a first error rate of each candidate image feature by each decision tree in the second machine learning model, and determining the importance of each candidate image feature among all candidate image features based on the first error rate.

[0108] As an example, the second machine learning model can calculate the importance of a single feature. The second machine learning model has many candidate image features, some of which have noise. Therefore, the importance of each feature needs to be calculated and the candidate image features are sorted, and then the candidate image features with high importance can be selected from all features. The second machine learning model can be a random forest model, or the target prediction image feature can be selected from a plurality of candidate image features by lasso algorithm.

[0109] In some embodiments, the determination of the importance of each candidate image feature among all candidate image features based on the first error rate can be implemented by the following technical solution: for each decision tree, obtaining medical image out-of-bag data corresponding to the decision tree, determining a first error rate of the development information prediction processing by the medical image out-of-bag data, performing noise interference processing on each candidate image feature of the medical image out-of-bag data to obtain medical noise image out-of-bag data of each candidate image feature after noise interference, and determining a second error rate of the development information prediction processing by each medical noise image out-of-bag data; for each candidate image feature, determining the noise influence of any one decision tree, wherein the noise influence is the difference between the first error rate and the second error rate of the corresponding decision tree, and the second error rate corresponds to the decision tree and the candidate image feature; and performing average processing on the differences of the plurality of decision trees to obtain the specificity degree of each candidate image feature.

[0110] As an example, the calculation method of the importance degree of a certain candidate image feature X in the second machine learning model is as follows: for each decision tree in the second machine learning model, the first error rate of the decision tree is calculated using the corresponding medical image out-of-bag data, noise interference is randomly added to the candidate image feature X of the medical image out-of-bag data, that is, the value of the medical image out-of-bag data at the candidate image feature X is randomly changed, the second error rate is calculated again, assuming that there are multiple decision trees in the second machine learning model, then the importance degree of the candidate image feature X is the average of the noise influence of each decision tree, the noise influence is the difference between the second error rate and the first error rate. The reason why this expression can be used as the importance degree of the candidate image feature X is that: if noise interference is randomly added to a certain candidate image feature, the error rate of the medical image out-of-bag data increases significantly, which means that the candidate image feature has a great influence on the prediction result of the event development information of the medical image out-of-bag data, that is, its importance degree is high.

[0111] As an example, the calculation method of the first error rate is as follows: the event development information is predicted by the decision tree A for the medical image out-of-bag data, the predicted event development information is compared with the actual event development information, if the comparison result is within the set range, it is determined that the prediction is accurate, the prediction process is repeated multiple times, and the number of accurate predictions is determined to obtain the first error rate. The calculation method of the second error rate is similar to that of the first error rate, the difference is only that the prediction basis is changed to the medical noise image out-of-bag data, that is, the medical image out-of-bag data with noise interference, and the value of the corresponding medical noise image out-of-bag data at the candidate image feature X is also subjected to noise interference. When building each tree of the second machine learning model, the training set is randomly and with replacement extracted, for each decision tree (assuming for the kth decision tree), about half of the training instances do not participate in the generation of the kth decision tree, this half of the data is called the out-of-bag data of the kth decision tree.

[0112] In some embodiments, the determining of the target prediction image feature corresponding to the target event type from the candidate image features based on the importance of each candidate image feature in predicting the development information can be achieved by the following technical solution: performing descending order sorting of each candidate image feature based on the corresponding importance, and taking the candidate image features with higher ranking as candidate target prediction image features; when the number of the candidate target prediction image features is not greater than the target prediction image feature number threshold, determining the candidate target prediction image features as the target prediction image features corresponding to the target event type; when the number of the candidate target prediction image features is greater than the target prediction image feature number threshold, updating the second machine learning model based on the candidate image features, so as to continue to determine the importance of each candidate target prediction image feature by using the updated second machine learning model, and delete the candidate target prediction image features with lower importance ranking.

[0113] As an example, the target of feature selection is to determine the candidate image features highly related to the event development information, and select a smaller number of candidate image features to sufficiently predict the event development information. First, the candidate image features in the second machine learning model are sorted in descending order of importance, and the candidate image features with corresponding proportion of unimportance are removed from the current candidate image features, so as to obtain a new feature set. The plurality of decision trees are updated by using the new feature set, and the importance of each candidate image feature in the feature set is calculated for sorting. The above steps are repeated until only a specified number of candidate image features are left.

[0114] In some embodiments, a plurality of groups of candidate image features can also be obtained by the above-mentioned embodiments, and each group of candidate image features has a specified number of candidate image features. For the plurality of groups of candidate image features, the corresponding second machine learning models are established, and the error rates of the corresponding groups of candidate image features are calculated. The group of features with the lowest error rate is selected as the finally selected target prediction image features.

[0115] In step 104, the development information prediction processing is performed based on the target prediction image feature of the target event type, and the event development information corresponding to the target object identifier is obtained.

[0116] As an example, the event development information corresponding to the target object identifier is the event development information of the target event type (for example, the disease type) of the patient. The event development information can be medical history information or medical future information. The medical history information includes the onset time of the disease, the development time of each stage of the disease development, the severity level of the disease, and the like. The medical future information includes the time when the disease reaches a future stage, for example, the medical image is taken in the early stage of cancer. The medical future information includes the time when the cancer reaches the middle stage, the state of the cancer in the middle stage, and the like.

[0117] In some embodiments, the development information prediction processing is implemented by a second machine learning model comprising a plurality of decision trees; and the development information prediction processing based on the target prediction image feature of the target event type in step 104 to obtain the event development information corresponding to the target object identifier can be implemented by the following technical solution: mapping the target prediction image feature to the event development information of the corresponding target event type by each decision tree; and performing average processing on the event development information obtained by each decision tree, and taking the average processing result as the event development information of the corresponding target event type. By predicting the event development information through a plurality of decision trees respectively, the prediction accuracy and stability can be effectively improved, the probability of prediction error can be reduced, and thus the event development information obtained as an intermediate result has reference value.

[0118] For example, when the second machine learning model comprises a plurality of decision trees, the target prediction image feature is mapped to the event development information of the corresponding target event type by each decision tree, and the event development information comprises the onset time, the node time of reaching a certain disease development node of the corresponding target event type, and the like. When the event development information is the onset time, the onset time obtained by each decision tree is averaged, and the average processing result is taken as the onset time of the corresponding target event type. For example, there are three decision trees, and the predicted event development information is 3.5 hours, 3 hours and 4 hours, and the onset time predicted by the second machine learning model is 3.5 hours. When the event development information is the onset severity level, the onset severity level obtained by each decision tree is averaged, and the average processing result is taken as the onset severity level of the corresponding target event type. For example, there are three decision trees, and the predicted event development information is level 1, level 2 and level 3, and the onset severity level predicted by the second machine learning model is level 2.

[0119] For example, the process of obtaining the event development information corresponding to the target object identifier based on the target prediction image feature of the target event type in step 104 can also be implemented by a Lasso model, a gradient descent tree model, an extreme gradient boosting tree model and a category gradient boosting tree model.

[0120] In some embodiments, the development information prediction processing is implemented by a second machine learning model including a plurality of decision trees; before the development information prediction processing based on the target prediction image features of the target event type in step 104 is performed to obtain the event development information corresponding to the target object identification, a medical image sample with pre-labeled event development information and a pre-labeled target event type is obtained; a region of interest sample corresponding to the pre-labeled target event type is extracted from the medical image sample; a candidate image feature sample corresponding to the region of interest sample is obtained; the candidate image feature sample of the region of interest is subjected to feature screening processing to obtain a target prediction image feature sample corresponding to the medical image sample; the target prediction image feature sample is forward propagated in the second machine learning model, and the forward propagation result is backward propagated in the second machine learning model with the pre-labeled event development information to update the parameters of the second machine learning model.

[0121] As an example, the pre-labeled event development information can be pre-labeled medical history information or pre-labeled medical future information, wherein the pre-labeled medical history information includes pre-labeled onset time, pre-labeled onset state, etc. of a disease, and the pre-labeled medical future information includes time of the disease reaching a future stage, for example, the medical image is taken in the early stage of cancer, and the medical future information includes pre-labeled time of the cancer in the middle stage, pre-labeled state of the cancer in the middle stage, etc.

[0122] As an example, before the event development information prediction is performed, the second machine learning model needs to be constructed by using the labeled data, the target object with the pre-labeled event development information is used as the training data, the target prediction image features are constructed, the obtained target prediction image features are used as the input, the pre-labeled event development information is used as the label, the second machine learning model is trained, and after the training is completed, the second machine learning model can be directly used to predict the event development information of the target object with unknown event development information.

[0123] In the following, an exemplary application of the embodiments of the present application in an actual application scenario will be described.

[0124] The embodiments of the present application provide a medical image processing method based on artificial intelligence, which is widely applicable to the dependence of disease event development information in various medical scenarios, for example, prediction of onset time, especially when the relevant medical image has a great reference significance for disease development trend, the event development information prediction based on the target prediction image features of the region of interest of the medical image can effectively improve the prediction accuracy of the event development information.

[0125] In some embodiments, the prediction of the event development information includes predicting the onset time, and in the case where the onset time of the target object (patient) is unknown, the real onset time of the target object is predicted based on the medical scan image by an algorithm. The following will take the target event type of hemorrhagic stroke as an example to specifically describe the steps of the medical image processing method based on artificial intelligence provided by the embodiments of the present application.

[0126] In some embodiments, referring to Figure 8 , Figure 8 is a flowchart of the medical image processing method provided by the embodiments of the present application. In step 201, a medical image is received, and in step 202, a region of interest is extracted. The region of interest in the medical scan image needs to be outlined and segmented, and the region of interest under different target event types is different. Taking hemorrhagic stroke as an example, the blood clot region is highly related to the onset of the disease of the target object, so this region is the region that medical staff and researchers pay more attention to, and is marked as the region of interest. At the same time, there are two methods for labeling the region of interest, for example, segmentation can be performed based on the operation of professionals, or segmentation can be performed through a neural network model (for example, a U-shaped neural network). The former has higher accuracy and better labeling effect, and the latter has faster processing speed and saves manpower. Referring to Figure 6 and Figure 7 , Figure 6 shows a brain CT image under the target event type of hemorrhagic stroke. Figure 7 shows an example of labeling the hemorrhagic region of the brain CT under the target event type of hemorrhagic stroke.

[0127] In some embodiments, in step 203, candidate image feature extraction is performed. The intensity information of the medical image in the region of interest obtained in the previous step is statistically calculated by a mathematical calculation method to obtain statistically significant candidate image features (representing image omics information). The candidate image features include histogram intensity candidate image features, shape candidate image features, texture candidate image features, wavelet candidate image features, etc. In the target event type of hemorrhagic stroke, a total of 851-dimensional candidate image features in the CT image can be extracted based on the pyradiomics tool package in the programming language for subsequent use. The image omics extraction method and the number of candidate image feature categories used in this step can be flexibly adjusted.

[0128] In some embodiments, in step 204, event history feature combination (optional) is performed. When other medical record information of the target object is available, important information in the medical record can be combined with the candidate image features as auxiliary candidate image features, and input into the next step of candidate image feature selection. The event history features include at least one of the following: admission diastolic blood pressure, admission systolic blood pressure, whether drinking or smoking, relevant medical history, etc. When the event history features are lacking, only the candidate image features can be used for onset time prediction.

[0129] In some embodiments, in step 205, feature selection is performed. The candidate image features are selected by at least one of the following methods: independent feature selection and model feature selection. A large number of candidate image features obtained in the previous step are screened to retain important candidate image features and exclude useless candidate image features, thereby improving the robustness and accuracy of the final onset time prediction algorithm. In independent feature selection, the correlation degree index of each type of candidate image feature and the onset time is calculated independently, thereby ensuring that each candidate image feature has strong reference significance in the onset time prediction task. The index of independent feature selection includes at least one of the following: mutual information, Pearson coefficient, ANOVA F-value, etc. In model feature selection, a related machine learning model is used to measure the redundancy between each candidate image feature, and important candidate image features with strong specificity are retained. The candidate image feature selection model includes: lasso model, random forest model, etc.

[0130] In some embodiments, in step 206, machine learning model prediction is performed to obtain the predicted onset time. The target prediction image features selected by the candidate image feature selection are input into the machine learning method to predict the real onset time of the target object corresponding to the target event type. Before performing the onset time prediction, a second machine learning model needs to be constructed using the labeled data. The target object with pre-labeled onset time is used as training data to construct the target prediction image features. The obtained target prediction image features are input, and the pre-labeled onset time is used as the label to train the second machine learning model. After the training is completed, the second machine learning model can be directly used to predict the onset time of the target object with unknown onset time.

[0131] The embodiment of the application provides a medical image processing method based on artificial intelligence, effectively utilizes candidate image features and machine learning to predict disease onset time, provides a novel onset time prediction method, predicts event development information for patients who cannot provide accurate event development information, is used for assisting doctors and researchers to diagnose diseases, recheck and research treatment methods, has strong flexibility, according to different requirements, various methods can be selected in each process module; strong expansibility, can be applied to various medical images of target event types having strong reference significance for disease development.

[0132] The following continues to illustrate an example structure of the implementation of the medical image processing device 255 based on artificial intelligence provided by the embodiment of the application as a software module, in some embodiments, as shown in Figure 2 The software module stored in the medical image processing device 255 based on artificial intelligence in the memory 250 can include: a region module 2551, configured to acquire a medical image processing request including a target object identifier; extract a region of interest matched with a target event type from a medical image corresponding to the target object identifier according to the medical image processing request; a feature module 2552, configured to extract candidate image features of the target event type from the region of interest; a selection module 2553, configured to perform feature screening processing on the candidate image features of the target event type to obtain target prediction image features of the target event type; and a prediction module 2554, configured to perform development information prediction processing based on the target prediction image features of the target event type to obtain event development information corresponding to the target object identifier.

[0133] In the above scheme, the region module 2551 is further configured to: in response to a cropping operation on the medical image, collect an event region corresponding to the target event type from the medical image corresponding to the target object identifier, and determine the event region as the region of interest corresponding to the target event type.

[0134] In the above scheme, the region module 2551 is further configured to: call a first machine learning model to perform segmentation processing on the medical image corresponding to the target object identifier, obtain an event region corresponding to the target event type, and determine the event region as the region of interest corresponding to the target event type.

[0135] In the above scheme, the first machine learning model comprises N cascaded down-sampling networks, a convolution network, and M cascaded up-sampling networks; N and M are integers greater than or equal to 2; the region module 2551 is further configured to: perform down-sampling processing on the medical image by using the N cascaded down-sampling networks to obtain a down-sampling processing result comprising features of different parts in the medical image; perform convolution processing on the down-sampling processing result by using the convolution network to obtain convolution features corresponding to the medical image; and perform up-sampling processing on the convolution features by using the M cascaded up-sampling networks to obtain a region in the medical image corresponding to an event manifestation comprising the target event type, as a corresponding event region; wherein the event manifestation comprises at least one of the following: a blood vessel manifestation, a bone manifestation, and a space-occupying manifestation.

[0136] In the above scheme, the feature module 2552 is further configured to: obtain attribute data of a candidate image feature corresponding to the target event type; wherein the attribute data comprises at least one of the following: a type of the candidate image feature, and a dimension of the candidate image feature; and extract at least one of the following features from the region of interest in the medical image: a histogram intensity candidate image feature, a shape candidate image feature, a texture candidate image feature, and a wavelet candidate image feature, which are consistent with the attribute data and correspond to the event manifestation of the target event type.

[0137] In the above scheme, the selection module 2553 is further configured to: obtain event history features corresponding to the target object identifier and the target event type; when an information amount of the event history features is greater than an information amount threshold, perform feature screening processing on the event history features and the candidate image features to obtain target prediction image features corresponding to the medical image; and when the information amount of the event history features is not greater than the information amount threshold, perform feature screening processing on the candidate image features to obtain the target prediction image features corresponding to the medical image.

[0138] In the above scheme, the selection module 2553 is further configured to: determine an importance degree of each candidate image feature for development information prediction processing; and determine target prediction image features corresponding to the target event type from the candidate image features based on the importance degree of each candidate image feature for the development information prediction processing.

[0139] In the above scheme, the selection module 2553 is further configured to: determine a correlation degree between each candidate image feature and event development information corresponding to the target event type; wherein the correlation degree comprises at least one of the following: a Pearson coefficient between the candidate image feature and the event development information; mutual information between the candidate image feature and the event development information; and an analysis of variance ratio between the candidate image feature and the event development information; and determine an importance degree positively correlated with the correlation degree.

[0140] In the above scheme, the time prediction processing is implemented through a second machine learning model including a plurality of decision trees; the selection module 2553 is further configured to determine, through each decision tree in the second machine learning model, a first error rate of each candidate image feature, so as to determine, based on the first error rate, an importance degree of each candidate image feature among all candidate image features.

[0141] In the above scheme, the selection module 2553 is further configured to: sort each candidate image feature in descending order based on the corresponding importance degree, and take a plurality of candidate image features with high sorting order as candidate target prediction image features; when the number of the candidate target prediction image features is not greater than a target prediction image feature number threshold, determine the candidate target prediction image features as the target prediction image features of the corresponding target event type; when the number of the candidate target prediction image features is greater than the target prediction image feature number threshold, update the second machine learning model based on the candidate image features, so as to continue to determine the importance degree of each candidate target prediction image feature through the updated second machine learning model, and delete the candidate target prediction image features with low importance degree.

[0142] In the above scheme, the selection module 2553 is further configured to: for each decision tree, perform the following processing: obtain medical image out-of-bag data corresponding to the decision tree, determine a first error rate of development information prediction processing performed through the medical image out-of-bag data, perform noise interference processing on each candidate image feature of the medical image out-of-bag data, to obtain medical noise image out-of-bag data of each candidate image feature after noise interference, and determine a second error rate of development information prediction processing performed through each medical noise image out-of-bag data; for each candidate image feature, perform the following processing: determine a noise influence of any one decision tree, wherein the noise influence is a difference between the first error rate and the second error rate of the corresponding decision tree, and the second error rate corresponds to the decision tree and the candidate image feature; and average the differences of the plurality of decision trees to obtain a specificity degree of each candidate image feature.

[0143] In the above scheme, the time prediction processing is implemented through a second machine learning model including a plurality of decision trees; the prediction module 2554 is further configured to: map the target prediction image feature to event development information through each decision tree; and average the event development information obtained through each decision tree, and take the average processing result as the event development information.

[0144] In the foregoing scheme, the time prediction processing is implemented by the second machine learning model including a plurality of decision trees; the apparatus 255 further includes a training module 2555 configured to: before the event development information prediction processing based on the target prediction image features of the target event type is performed to obtain the event development information of the target event type of the target object, obtain a medical image sample with a pre-labeled event development information and a pre-labeled target event type; extract a region of interest sample corresponding to the pre-labeled target event type from the medical image sample; obtain a candidate image feature sample corresponding to the region of interest sample; perform feature selection processing on the candidate image feature sample of the region of interest to obtain a target prediction image feature sample corresponding to the medical image sample; and perform forward propagation on the target prediction image feature sample in the second machine learning model, and perform backward propagation on the pre-labeled event development information in the second machine learning model to update parameters of the second machine learning model.

[0145] The embodiment of the present application provides a computer program product or a computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the medical image processing method based on artificial intelligence provided in the embodiment of the present application.

[0146] The embodiment of the present application provides a computer readable storage medium storing executable instructions, wherein the executable instructions are stored in the computer readable storage medium. When the executable instructions are executed by a processor, the processor will execute the medical image processing method based on artificial intelligence provided in the embodiment of the present application, for example, as shown in the medical image processing method based on artificial intelligence. Figures 3-5 The medical image processing method based on artificial intelligence is shown.

[0147] In some embodiments, the computer readable storage medium can be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; or various devices including one or any combination of the above memories.

[0148] In some embodiments, the executable instructions can be in the form of programs, software, software modules, scripts or codes, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and can be deployed in any form, including being deployed as independent programs or being deployed as modules, components, subroutines or other units suitable for use in a computing environment.

[0149] By way of example, executable instructions can correspond to a file in a file system, can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub programs, or portions of code.

[0150] By way of example, executable instructions can be deployed to be executed on one computer, or on multiple computers of a distributed system at one location, or on multiple computers of a distributed system located at multiple locations including different geographical locations.

[0151] In summary, by performing the region of interest recognition and the effective feature extraction on the medical image through the embodiments of the present application, the effective feature information related to the disease type can be extracted, which lays a foundation for the precise development time prediction. In addition, by performing the selection processing on the features, the effective features can be obtained, which can save the calculation resource consumption and the feature engineering consumption of the prediction process.

[0152] The above merely illustrates the embodiments of the present application, but is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, and improvement within the spirit and scope of the present application shall be included in the protection scope of the present application.

Claims

1. An artificial intelligence-based medical image processing method, characterized by, The method comprises: obtaining a medical image processing request comprising a target object identifier; extracting a region of interest matching a target event type from a medical image corresponding to the target object identifier according to the medical image processing request; extracting candidate image features corresponding to the target event type from the region of interest; determining the importance of each candidate image feature for development information prediction processing, which is realized by a second machine learning model; sorting each candidate image feature in descending order based on the corresponding importance, and selecting the top candidate image features as candidate target prediction image features; when the number of candidate target prediction image features is not greater than a target prediction image feature threshold, determining the candidate target prediction image features as target prediction image features corresponding to the target event type; when the number of candidate target prediction image features is greater than the target prediction image feature threshold, updating the second machine learning model based on the candidate image features to continue determining the importance of each candidate target prediction image feature by the updated second machine learning model, and deleting the candidate target prediction image features with lower importance; performing development information prediction processing based on the target prediction image features corresponding to the target event type to obtain event development information corresponding to the target object identifier.

2. The method of claim 1, wherein, The method comprises: in response to a cropping operation on the medical image, collecting an event region corresponding to the target event type from the medical image corresponding to the target object identifier, and determining the event region as the region of interest corresponding to the target event type.

3. The method of claim 1, wherein, The method comprises: calling a first machine learning model to perform segmentation processing on the medical image corresponding to the target object identifier to obtain an event region corresponding to the target event type, and determining the event region as the region of interest corresponding to the target event type.

4. The method of claim 3, wherein, The first machine learning model comprises N cascaded down-sampling networks, a convolution network, and M cascaded up-sampling networks; N and M are integers greater than or equal to 2. The method comprises: performing down-sampling processing on the medical image by the N cascaded down-sampling networks to obtain a down-sampling processing result comprising features of different parts in the medical image; performing convolution processing on the down-sampling processing result by the convolution network to obtain convolution features corresponding to the medical image; performing up-sampling processing on the convolution features by the M cascaded up-sampling networks to obtain a region in the medical image corresponding to an event region comprising an event manifestation corresponding to the target event type; wherein the event manifestation comprises at least one of the following: a blood vessel manifestation, a bone manifestation, and a space-occupying manifestation.

5. The method of claim 1, wherein, The extracting the candidate image features corresponding to the target event type from the region of interest comprises: obtaining attribute data of the candidate image features corresponding to the target event type; The attribute data comprises at least one of the following: type of the candidate image features, dimension of the candidate image features; The method further comprises:

6. The method of claim 1, wherein, obtaining event history features corresponding to the target object identifier and the target event type; When the information amount of the event history features is greater than the information amount threshold, performing feature screening processing on the event history features and the candidate image features to obtain target prediction image features corresponding to the medical image; When the information amount of the event history features is not greater than the information amount threshold, performing feature screening processing on the candidate image features to obtain target prediction image features corresponding to the medical image. The method further comprises:

7. The method of claim 1, wherein, determining a correlation degree between each of the candidate image features and event development information corresponding to the target event type; The correlation degree comprises at least one of the following: Pearson coefficient between the candidate image features and the event development information; mutual information between the candidate image features and the event development information; analysis of variance ratio between the candidate image features and the event development information; determining an importance degree positively correlated with the correlation degree. The development information prediction processing is realized through a second machine learning model, and the second machine learning model comprises a plurality of decision trees.

8. The method of claim 1, wherein, The determining the importance degree of each of the candidate image features in the development information prediction processing comprises: determining a first error rate of each of the candidate image features through each of the decision trees in the second machine learning model, and determining the importance degree of each of the candidate image features in all the candidate image features based on the first error rate. The determining the importance degree of each of the candidate image features in the development information prediction processing comprises:

9. The method of claim 8, wherein, for each of the decision trees, performing the following processing: obtaining medical image out-of-bag data corresponding to the decision tree, determining a first error rate of development information prediction processing through the medical image out-of-bag data, performing noise interference processing on each of the candidate image features of the medical image out-of-bag data to obtain medical noise image out-of-bag data corresponding to each of the candidate image features after noise interference, and determining a second error rate of development information prediction processing through each of the medical noise image out-of-bag data; ​ The following processing is performed for each candidate image feature: determining a noise influence of any one of the decision trees, wherein the noise influence is a difference between a first error rate corresponding to the decision tree and a second error rate corresponding to the decision tree and the candidate image feature; and performing average processing on the differences of the plurality of decision trees to obtain a specificity degree of each candidate image feature.

10. The method of claim 1, wherein, The development information prediction processing is implemented by a second machine learning model including a plurality of decision trees. The development information prediction processing based on the target prediction image feature corresponding to the target event type is performed to obtain event development information corresponding to the target object identifier, including: mapping, by each decision tree, the target prediction image feature into event development information; performing average processing on the event development information obtained by each decision tree, and taking the average processing result as the event development information.

11. An artificial intelligence-based medical image processing apparatus, characterized by, including: a region module configured to obtain a medical image processing request including a target object identifier; extracting, from a medical image corresponding to the target object identifier, a region of interest matching a target event type according to the medical image processing request; a feature module configured to extract, from the region of interest, candidate image features corresponding to the target event type; a selection module configured to determine an importance degree of each candidate image feature for development information prediction processing, wherein the development information prediction processing is implemented by a second machine learning model; performing descending order sorting of each candidate image feature based on the corresponding importance degree, taking a plurality of candidate image features ranked at the front as candidate target prediction image features; when the number of the candidate target prediction image features is not greater than a target prediction image feature number threshold, determining the candidate target prediction image features as target prediction image features corresponding to the target event type; when the number of the candidate target prediction image features is greater than the target prediction image feature number threshold, updating the second machine learning model based on the candidate image features, so as to continue to determine the importance degree of each candidate target prediction image feature by using the updated second machine learning model, and delete the candidate target prediction image features ranked at the back according to the importance degree; a prediction module configured to perform development information prediction processing based on target prediction image features corresponding to the target event type to obtain event development information corresponding to the target object identifier.

12. The apparatus of claim 11, wherein, The region module is further configured to: in response to a cropping operation on the medical image, collect an event region corresponding to the target event type from a medical image corresponding to the target object identifier, and determine the event region as the region of interest corresponding to the target event type.

13. The apparatus of claim 11, wherein, The region module is further configured to: invoke a first machine learning model to perform segmentation processing on a medical image corresponding to the target object identifier to obtain an event region corresponding to the target event type, and determine the event region as the region of interest corresponding to the target event type.

14. The apparatus of claim 13, wherein, The first machine learning model comprises N cascaded down-sampling networks, a convolution network, and M cascaded up-sampling networks; N and M are integers greater than or equal to 2; the region module is further configured to: perform down-sampling processing on the medical image through the N cascaded down-sampling networks to obtain a down-sampling processing result comprising features of different parts in the medical image; perform convolution processing on the down-sampling processing result through the convolution network to obtain convolution features corresponding to the medical image; and perform up-sampling processing on the convolution features through the M cascaded up-sampling networks to obtain a region in the medical image corresponding to an event manifestation comprising the target event type, as a corresponding event region; wherein the event manifestation comprises at least one of a blood vessel manifestation, a bone manifestation, and a space-occupying manifestation.

15. The apparatus of claim 11, wherein, The feature module is further configured to: obtain attribute data of a candidate image feature corresponding to the target event type; wherein the attribute data comprises at least one of a type of the candidate image feature and a dimension of the candidate image feature; and extract at least one of the following features of the event manifestation corresponding to the target event type from a region of interest of the medical image: a histogram intensity candidate image feature, a shape candidate image feature, a texture candidate image feature, and a wavelet candidate image feature.

16. The apparatus of claim 11, wherein, The selection module is further configured to: obtain an event history feature corresponding to the target object identifier and the target event type; when an information amount of the event history feature is greater than an information amount threshold, perform feature screening processing on the event history feature and the candidate image feature to obtain a target prediction image feature corresponding to the medical image; and when the information amount of the event history feature is not greater than the information amount threshold, perform feature screening processing on the candidate image feature to obtain the target prediction image feature corresponding to the medical image.

17. The apparatus of claim 11, wherein, The selection module is further configured to: determine a correlation degree between each of the candidate image features and event development information corresponding to the target event type; wherein the correlation degree comprises at least one of a Pearson coefficient between the candidate image feature and the event development information, mutual information between the candidate image feature and the event development information, and an analysis of variance ratio between the candidate image feature and the event development information; and determine an importance degree positively correlated with the correlation degree.

18. An electronic device, comprising: comprises: a memory configured to store executable instructions; a processor configured to execute the executable instructions stored in the memory to implement the medical image processing method based on artificial intelligence according to any one of claims 1 to 10.

19. A computer-readable storage medium, characterized in that, executable instructions stored in the memory, and configured to be executed by a processor to implement the medical image processing method based on artificial intelligence according to any one of claims 1 to 10.

20. A computer program product comprising computer instructions, characterized in that, The computer instructions are executed by the processor to implement the medical image processing method based on artificial intelligence according to any one of claims 1 to 10.

Images