Method, device, and computer program for training artificial intelligence model for predicting blood flow data

Abstract

According to an embodiment of the present disclosure, a method for generating an artificial intelligence model for predicting blood flow data may comprise the steps of: setting a structure of a blood flow data prediction model; generating blood vessel data in the form of a point cloud corresponding to a major blood vessel from a three-dimensional medical image, specifying a boundary region of the blood vessel data, obtaining a blood flow value of the boundary region, and preparing training data for a blood flow prediction model by using a position of the point cloud for an arbitrary region and the blood flow value at the boundary region; and training the blood flow prediction model by individually adjusting a learning rate of the blood flow prediction model for the training data according to a change amount of each parameter, and simultaneously setting an overall learning rate to gradually decrease in the form of a cosine function.

Description

Method, device, and computer program for training an artificial intelligence model to predict blood flow data

[0001] The present disclosure relates to a method, apparatus, and computer program for training an artificial intelligence model that predicts blood flow data.

[0002] In modern medicine, medical imaging is a crucial tool for the effective diagnosis of diseases and the treatment of patients. Furthermore, advancements in imaging technology have led to the generation of increasingly sophisticated medical image data. Consequently, the volume of data is becoming increasingly massive, making it difficult to analyze medical image data by relying solely on human vision. Therefore, over the past decade, clinical decision support systems and computer-assisted reading systems have played an essential role in the automated analysis of medical images.

[0003] Conventional clinical decision support systems or computer-assisted reading systems perform the function of detecting and displaying lesion areas or providing reading information to medical staff or medical professionals, etc. (hereinafter referred to as users).

[0004] For example, the 'Method and Apparatus for Calculating Disease Diagnosis Information Based on Medical Images' disclosed in Korean Published Patent No. 10-2017-0017614 includes the steps of detecting a region of interest in which an object to be analyzed is captured, calculating a coefficient of variation, creating a coefficient of variation image, and comparing it with a reference sample, and describes the effect of diagnosing the degree of a patient's disease by utilizing medical images obtained through a computed tomography device and an ultrasound imaging device.

[0005] Meanwhile, in recent years, Artificial Intelligence (AI) technology based on machine learning, such as deep learning, has been bringing about rapid advancements in the field of analyzing and processing medical images. Deep learning technology in medical imaging is applied to various tasks including image analysis, diagnosis, treatment, and prediction. For example, deep learning-based auxiliary diagnostic systems are being applied to diagnose cerebrovascular diseases, such as cerebral aneurysms.

[0006] One technical problem of the present disclosure is to provide a method for training an artificial intelligence model to predict blood flow data.

[0007] Another technical problem of the present disclosure is to provide a method for augmenting a training dataset to improve the performance of a blood flow data prediction model, and for the model to learn data with various structures by collecting the training dataset from different people.

[0008] Another technical objective of the present disclosure is to provide a method for reducing the waste of computational resources required for training the model and shortening the training time.

[0009] Another technical objective of the present disclosure is to provide a system for predicting blood flow data using a blood vessel graph and the artificial intelligence model, and a method for implementing the same.

[0010] Another technical problem of the present disclosure is to provide a method for generating a vascular graph containing coordinates and structural information for major blood vessels from a three-dimensional vascular image.

[0011] The technical problems of the present disclosure are not limited to the technical problems described above, and technical problems not mentioned will be clearly understood by those skilled in the art to which the present disclosure belongs from the present specification and the accompanying drawings.

[0012] This summary is provided to introduce the selection of the concept in a simplified form, which is further explained in the detailed description below. This summary is not intended to identify the principal or essential features of the claimed subject matter, nor is it intended to assist in determining the scope of the claimed subject matter.

[0013] A method for generating an artificial intelligence model for predicting blood flow data according to one embodiment of the present disclosure may include: a step of setting the structure of a blood flow data prediction model; a step of generating blood vessel data in the form of a point cloud corresponding to major blood vessels from a three-dimensional medical image, specifying a boundary region of the blood vessel data, obtaining a blood flow value of the boundary region, and preparing the location of the point cloud for an arbitrary region and the blood flow value of the region boundary as training data for the blood flow data prediction model; and a step of training the blood flow data prediction model by individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter, and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function.

[0014] A system for generating an artificial intelligence model for predicting blood flow data according to one embodiment of the present disclosure may include: a 3D medical image processing device that generates a 3D blood vessel image by analyzing a 3D medical image and extracting coordinates of blood vessel pixel points corresponding to blood vessels based on pixel values; a blood vessel graph generating device that generates a blood vessel graph from the 3D blood vessel image using the group of blood vessel pixel points as nodes; and an artificial intelligence model generating device that generates the blood flow data prediction model by setting the structure of the blood flow data prediction model, specifying a boundary area of ​​the blood vessel graph, obtaining blood flow values ​​of the boundary area, preparing the location of an arbitrary area and the blood flow values ​​of the area boundary as training data for the blood flow data prediction model, individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter, and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function.

[0015] A server for generating an artificial intelligence model for predicting blood flow data according to one embodiment of the present disclosure may include: a communication unit for receiving a three-dimensional medical image; and a control unit for setting the structure of a blood flow data prediction model, generating blood vessel data in the form of a point cloud corresponding to major blood vessels from the three-dimensional medical image, specifying a boundary area of ​​the blood vessel data, obtaining a blood flow value of the boundary area, preparing the location of the point cloud for an arbitrary area and the blood flow value of the boundary of the area as training data for the blood flow data prediction model, individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter, and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function to train the blood flow data prediction model and generate the blood flow data prediction model.

[0016] A non-transient computer-readable recording medium according to one embodiment of the present disclosure may have a computer program recorded thereon for executing a method to generate an artificial intelligence model that predicts blood flow data, which is executed by a processor.

[0017] The technical solution of the present disclosure is not limited to the technical solution described above, and technical solutions not mentioned will be clearly understood by those skilled in the art to which the present disclosure belongs from the present specification and the attached drawings.

[0018] According to the present disclosure, the performance of an artificial intelligence model for predicting blood flow data can be improved.

[0019] According to the present disclosure, there is an effect of reducing the resources required for training a blood flow data prediction model.

[0020] According to the present disclosure, a blood vessel graph containing coordinates and structural information for major blood vessels is generated from a three-dimensional blood vessel image, and blood flow data can be predicted using the blood vessel graph and a blood flow prediction artificial intelligence model.

[0021] The effects of the present disclosure are not limited to those described above, and unmentioned effects will be clearly understood by those skilled in the art from the present specification and the accompanying drawings.

[0022] FIG. 1 is a drawing for explaining a blood flow data prediction system according to one embodiment of the present disclosure.

[0023] FIG. 2 is a schematic diagram illustrating the operation of a blood flow data prediction system according to one embodiment of the present disclosure.

[0024] FIG. 3 is a diagram illustrating the structure and learning method of an artificial intelligence model for predicting blood flow data according to one embodiment of the present disclosure.

[0025] FIG. 4 is a flowchart illustrating a three-dimensional medical image processing method according to one embodiment of the present disclosure.

[0026] FIGS. 5 to 8 are illustrative drawings for explaining the operation of a three-dimensional medical image processing device according to one embodiment of the present disclosure.

[0027] FIG. 9 is a flowchart illustrating a method for generating a blood vessel graph according to one embodiment of the present disclosure.

[0028] FIGS. 10 to 14 are illustrative drawings for explaining the operation of a blood vessel graph generating device according to one embodiment of the present disclosure.

[0029] FIG. 15 is a flowchart illustrating a method for generating a blood flow data prediction model according to one embodiment of the present disclosure.

[0030] FIGS. 16 and 17 are drawings for explaining the structure and operation of a blood flow data prediction model according to one embodiment of the present disclosure.

[0031] FIG. 18 is a block diagram illustrating the configuration of a server that generates a blood flow data prediction model according to one embodiment of the present disclosure.

[0032] The technical problems, features, and advantages of the present disclosure described above will become more apparent from the following detailed description in conjunction with the accompanying drawings. However, as the present disclosure is subject to various modifications and may have various embodiments, specific embodiments are illustrated in the drawings and described in detail below.

[0033] Throughout the specification, identical reference numbers generally represent identical components. Additionally, components with identical functions within the same scope of concept appearing in the drawings of each embodiment are described using the same reference numeral, and redundant descriptions thereof are omitted.

[0034] If it is determined that a detailed description of known functions or configurations related to the present disclosure could unnecessarily obscure the essence of the present disclosure, such detailed description is omitted. Furthermore, numbers used in the description of this specification (e.g., First, Second, etc.) are merely identifiers to distinguish one component from another.

[0035] Furthermore, the suffixes "module" and "part" for components used in the following embodiments are assigned or used interchangeably solely for the ease of drafting the specification, and do not inherently possess distinct meanings or roles.

[0036] In the following examples, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0037] In the following embodiments, terms such as "include" or "have" mean that the features or components described in the specification are present, and do not preclude the possibility that one or more other features or components may be added.

[0038] In the drawings, the size of components may be exaggerated or reduced for convenience of explanation. For example, the size and thickness of each component shown in the drawings are arbitrarily depicted for convenience of explanation and the present disclosure is not necessarily limited to what is depicted.

[0039] Where an embodiment can be implemented differently, the order of a particular process may be performed differently from the order described. For example, two processes described consecutively may be performed substantially simultaneously or proceed in the reverse order of the description.

[0040] In the following embodiments, when components are described as being connected, the case includes not only instances where the components are directly connected but also instances where components are indirectly connected by interposing them in between.

[0041] For example, when it is stated in this specification that components, etc. are electrically connected, it includes not only cases where the components, etc. are directly electrically connected, but also cases where components, etc. are interposed in between and are indirectly electrically connected.

[0042] Hereinafter, with reference to FIGS. 1 to 18, a method, apparatus, and computer program for training an artificial intelligence model for predicting blood flow data according to an embodiment of the present disclosure will be described.

[0043] FIG. 1 is a drawing for explaining a blood flow data prediction system according to one embodiment of the present disclosure.

[0044] A blood flow data prediction system according to one embodiment of the present disclosure can acquire a three-dimensional medical image, extract blood vessel pixel points from the three-dimensional medical image to generate a three-dimensional blood vessel image, and generate a blood vessel graph using the three-dimensional blood vessel image. Subsequently, the blood flow data prediction system can specify a region of interest through user input and extract the region of interest from a previously generated blood vessel graph.

[0045] The blood vessel graph of the region of interest is in the form of a point cloud and can be configured as a 1D network connecting points (or nodes). Specifically, the blood vessel graph of the region of interest may include coordinate values ​​consisting of the x, y, and z coordinates of the nodes constituting the blood vessels, and radius information (r) of the nodes corresponding to the actual radius of the blood vessels.

[0046] Furthermore, the blood flow data prediction system can obtain actual blood flow data at the boundary of the region of interest. The actual blood flow data at the boundary of the region of interest may be, for example, blood pressure data at the boundary of the region of interest, which may be a measured value or a predicted value calculated by an algorithm. In the following description, the blood pressure data at the boundary of the region of interest may also be referred to as the 'boundary value (p)'.

[0047] A blood flow data prediction system can predict blood flow data for any arbitrary area by inputting a blood vessel graph of the region of interest and boundary values ​​(p) into a trained blood flow data prediction model. For example, the blood flow data prediction system can input boundary values ​​(p) into the blood flow data prediction model along with location information and radius information (x,y,z,r) of boundary nodes and non-boundary nodes (hereinafter referred to as 'non-boundary nodes') in the blood vessel graph of the region of interest. When an arbitrary point in the region of interest is identified as a query node, the blood flow data prediction model can predict the blood flow data (P) of the corresponding query node, and through this, the blood flow data prediction system can predict blood flow at any point, such as pressure, blood flow velocity, and flow rate.

[0048] More specifically, referring to FIG. 1, a blood flow data prediction system according to one embodiment of the present disclosure may include a three-dimensional medical image processing device (100), a blood vessel graph generation device (200), and a blood flow prediction device (300). Although FIG. 1 illustrates the blood flow data prediction system as including a plurality of devices, this is for convenience of explanation and the present disclosure should not be interpreted as being limited thereto. That is, in the blood flow data prediction system according to one embodiment of the present disclosure, modules performing each function may be executed in a single or plurality of devices or in a virtualized environment.

[0049] The three-dimensional medical image processing device (100) can generate a three-dimensional blood vessel image of a major blood vessel from a three-dimensional medical image.

[0050] A blood flow data prediction system according to one embodiment of the present disclosure can acquire a three-dimensional medical image, extract blood vessel pixel points from the three-dimensional medical image to generate a three-dimensional blood vessel image, and generate a blood vessel graph using the three-dimensional blood vessel image. Subsequently, the blood flow data prediction system can specify a region of interest through user input and extract the region of interest from a previously generated blood vessel graph.

[0051] More specifically, when the 3D medical image processing device (100) receives a 3D medical image, it can analyze the pixel values ​​of the 3D medical image to extract blood vessel pixel points and coordinates. For example, if any pixel value in the 3D medical image is greater than or equal to a preset threshold value, the 3D medical image processing device (100) determines that it corresponds to a blood vessel, identifies it as a blood vessel pixel point, and extracts coordinate information of the blood vessel pixel point.

[0052] In the above embodiment, the 3D medical image processing device (100) can extract the coordinates of a pixel if the pixel value is greater than or equal to a predetermined threshold value, and generate a blood vessel pixel coordinate table using the pixel coordinates. Subsequently, the 3D medical image processing device (100) can generate a coordinate system of all blood vessels by displaying blood vessel points on a coordinate system based on the blood vessel pixel coordinate table.

[0053] The 3D medical image processing device (100) can crop major blood vessel parts at blood vessel pixel points displayed on a coordinate system.

[0054] Subsequently, the 3D medical image processing device (100) can perform clustering for blood vessel pixel points and filter out noise and / or micro-blood vessel pixels.

[0055] According to an embodiment of the present disclosure, a three-dimensional medical image processing device (100) can group blood vessel pixel points into clusters by performing clustering based on one or more of the distance, density, and / or attribute information of blood vessel pixel points.

[0056] Subsequently, the 3D medical image processing device (100) can perform primary filtering by comparing the number of blood vessel pixel points included in any cluster with a threshold value and deleting clusters that are below the threshold value. Filtering can be performed by removing pixel points corresponding to clusters where the number of blood vessel points is below the threshold value from the blood vessel coordinate system. Through this, the 3D medical image processing device (100) can remove noise around major blood vessels.

[0057] A 3D medical image processing device (100) can group blood vessel pixel points into clusters by performing clustering again on primary filtered blood vessel pixel points based on one or more of their distance, density, and attribute information.

[0058] Subsequently, the 3D image processing device (100) can perform secondary filtering to delete the remaining clusters, excluding the blood vessel pixel points included in any cluster. For example, secondary filtering can be performed by removing pixel points corresponding to clusters other than the largest cluster from the blood vessel coordinate system. Through this, the 3D medical image processing device (100) can remove microvessels.

[0059] According to another embodiment of the present disclosure, a three-dimensional medical image processing device (100) may perform clustering and / or filtering of blood vessel pixel points using various algorithms, and the present disclosure is not to be interpreted as limiting the clustering and / or filtering algorithms.

[0060] For example, clustering can be performed by grouping blood vessel pixel points into K clusters, calculating the center of each cluster, and assigning the blood vessel pixel points to the cluster with the closest center. As another example, clusters can be formed using a similarity matrix for blood vessel pixel points. Subsequently, filtering of noise and / or microvessels surrounding the blood vessels can be performed by modifying or removing clusters.

[0061] According to an embodiment of the present disclosure, when a three-dimensional medical image processing device (100) performs filtering through clustering, blood vessel pixel points with noise and microvessels removed are formed in a blood vessel coordinate system, and thereby the three-dimensional medical image processing device (100) can generate a three-dimensional blood vessel image.

[0062] The blood vessel graph generating device (200) can generate a blood vessel graph based on a three-dimensional blood vessel image generated by a three-dimensional medical image processing device (100).

[0063] The blood vessel graph generating device (200) divides a three-dimensional blood vessel image into multiple layers and can check the coordinate values ​​of blood vessel pixel points for each layer. At this time, it is appropriate to generate the layers in two dimensions.

[0064] Subsequently, the blood vessel graph generating device (200) can group blood vessel pixel points for each layer. The group of blood vessel pixel points for each layer may correspond to a cross-section of an actual blood vessel.

[0065] Subsequently, the blood vessel graph generation device (200) can calculate the center of gravity of each group. This is for generating nodes of the blood vessel graph, and the nodes of the blood vessel graph according to one embodiment of the present disclosure can be set as the center of gravity of the blood vessel pixel point group for each layer. In the blood vessel graph according to one embodiment of the present disclosure, the nodes may correspond to the center of the actual blood vessel, and it is appropriate to set the radius of the nodes in the blood vessel graph to the actual radius of the blood vessel.

[0066] Subsequently, the blood vessel graph generation device (200) can generate connectivity information for nodes in the blood vessel graph. More specifically, connectivity information can be generated for nodes in the first group and nodes in the second group based on whether there is an intersection between a first group of blood vessel pixel points of an arbitrary layer and a second group of blood vessel pixel points of another layer adjacent to said layer. This is because the blood vessel pixel point group of each layer corresponds to a cross-section of an actual blood vessel, and the fact that a blood vessel pixel point group of any one layer has an intersection with a blood vessel pixel point group of a layer adjacent to said layer can be interpreted as actual blood vessels being connected.

[0067] Through this, the blood vessel graph generation device (200) can generate node and connection information from a three-dimensional blood vessel image and generate a blood vessel graph. Furthermore, the blood vessel graph generation device (200) can calculate the actual radius of the blood vessel in the blood vessel graph and set it as the radius of the node. Through this, characteristics such as branching, connection, size, and location of the blood vessel can be reflected in the blood vessel graph. A specific method for this will be described later in the description of the attached drawings.

[0068] The blood flow prediction device (300) can predict blood flow data at any point based on a blood vessel graph. To this end, the blood flow prediction device (300) can generate a blood flow data prediction model that predicts the pressure value of any node using location information and radius information of each node of the blood vessel graph of the region of interest. A method for generating a blood flow data prediction model according to one embodiment of the present disclosure will be described later together with the description of the attached drawings.

[0069] When a blood flow data prediction model is prepared, the blood flow prediction device (300) can identify a blood vessel graph of the region of interest. The blood flow prediction device (300) distinguishes between a boundary node located at the end of the blood vessel graph of the region of interest and a non-boundary node located inside the blood vessel graph of the region of interest, obtains location information (x, y, z) and radius information (r) of the boundary node and the non-boundary node, obtains the pressure value (p) of the boundary node, and inputs the obtained values ​​into the blood flow data prediction model.

[0070] The blood flow data prediction model can predict and output the pressure of any node (e.g., query node), and the blood flow prediction device (300) can predict blood flow data at any point.

[0071] FIG. 2 is a schematic diagram illustrating the operation of a blood flow data prediction system according to one embodiment of the present disclosure.

[0072] A blood flow data prediction system according to one embodiment of the present disclosure can predict blood flow data of any area using a blood flow data prediction model.

[0073] More specifically, the blood flow data prediction system can extract a blood vessel graph from a three-dimensional medical image. The blood vessel graph is in the form of a point cloud and can be configured in the form of a 1D network connecting points (or nodes). Specifically, the blood vessel graph may include coordinate values ​​composed of the x, y, and z coordinates of the nodes constituting the blood vessel, and radius information (r) of the nodes corresponding to the actual radius of the blood vessel.

[0074] Furthermore, the blood flow data prediction system can input a blood vessel graph into a trained blood flow data prediction model. The training and structure of the blood flow data prediction model will be described in more detail later in the description of the attached drawings.

[0075] A blood flow data prediction system can predict blood flow data, such as blood pressure at an arbitrary point, through the output value of a blood flow data prediction model. For example, when an arbitrary area is specified in a blood vessel graph, the blood flow data prediction system can identify the boundary of the area, obtain location information (x, y, z) and radius information (r) of the boundary node and non-boundary node in the blood vessel graph, obtain the pressure value (p) of the boundary node through measurement or calculation, and input the obtained values ​​into a blood flow data prediction model.

[0076] The blood flow data prediction model can predict and output the pressure of an arbitrary node (e.g., a query node), and through this, the blood flow data prediction system can predict blood flow data at an arbitrary point.

[0077] FIG. 3 is a diagram illustrating the structure and learning method of a blood flow data prediction model according to one embodiment of the present disclosure.

[0078] A blood flow data prediction model according to one embodiment of the present disclosure illustrated in FIG. 3 is a model trained to predict blood flow data based on a Graph Neural Network (GNN) architecture optimized for graph value prediction. The blood flow data prediction model can provide high prediction accuracy for blood flow volume and pressure.

[0079] In Fig. 3, n represents node features, e represents edge features, f represents the ground truth value of the flow rate, and p represents the ground truth value of the pressure. f' and p' represent the intermediate predicted flow rate and pressure, respectively. And, and represents the final predicted flow rate and pressure, respectively.

[0080] Meanwhile, although not separately illustrated in Fig. 3, the architecture of the blood flow data prediction model may include three components: an encoder, a processor, and a decoder.

[0081] The encoder converts the physical characteristics of nodes and the geometric information of edges into latent representations. The processor exchanges information between nodes through a message-passing mechanism and can perform the function of effectively propagating physical and geometric information within the network. The decoder can perform the function of predicting blood flow and pressure at the next time point by inferring changes in physical quantities from latent vectors.

[0082] The blood flow data prediction model of Fig. 3 can learn a large amount of graph data consisting of nodes and edges. Information such as the diameter of blood vessels, blood flow volume, and blood pressure can be stored in each node of the graph.

[0083] Vascular network data input to the blood flow data prediction model is collected from Magnetic Resonance Angiography (MRA) datasets. Vascular network data can be converted into a graph format, for example, a NetworkX graph, using the NetworkX library. The converted NetworkX graph can represent structural information of blood vessels as nodes and edges. Next, the converted NetworkX graph can be reconstructed using the Deep Graph Library (DGL). DGL is a framework designed to efficiently process large-scale graph data and can be used to extract features of nodes and edges. The dataset processed in this way can be input to the blood flow data prediction model.

[0084] In particular, according to an embodiment of the present disclosure, training data for a blood flow data prediction model can be augmented. For example, if a dataset of 10 cerebrovascular graphs is obtained, data augmentation can be performed by randomly modifying the radius and boundary conditions of their nodes to generate 40 modified graphs per original graph, thereby constructing a total of 10 × 40 = 400 graph data.

[0085] Once the dataset is prepared, the blood flow data prediction model can be trained, for example, on 320 cerebrovascular graph data points from 8 patients. Each cerebrovascular graph data point represents the cerebrovascular network of an individual patient as a graph and may include various characteristics of nodes and edges. During the training process, the blood flow data prediction model will use the blood flow or pressure values ​​of the given data as input to predict blood flow and pressure based on them.

[0086] The predicted values ​​can be refined through iterative learning using the Mean Absolute Error (MAE) as the loss function, and finally, accurate blood flow and pressure values ​​can be calculated.

[0087] In particular, according to one embodiment of the present disclosure, the time and resources required for model training can be optimized by gradually reducing the learning rate of the model for a dataset. For example, training can be set to automatically stop if the result of the loss function, which is the difference between the predicted value and the actual value, does not improve for a certain number of times.

[0088] More specifically, according to one embodiment of the present disclosure, Stochastic Gradient Descent (SGD) is applied to the model's dataset training, and the learning rate can be adaptively adjusted for each parameter using estimates of the first moment (average of the gradient) and the second moment (variance of the gradient). Through this setting, the loss function can be rapidly minimized during the initial training phase to reduce the total training time, and sensitivity to hyperparameters such as the learning rate can be reduced.

[0089] According to another embodiment of the present disclosure, the learning rate of a model can be gradually reduced in the form of a cosine function. That is, the model can be configured to start with a large learning rate initially to rapidly explore the global structure of the loss function and significantly reduce the loss value, and then switch to a small learning rate as training progresses to fine-tune the model parameters. As the learning rate decreases, the model captures minute changes in the loss function to reach optimal performance. Through this configuration, the time required for the entire training is reduced, and resource usage can be optimized by reducing unnecessary computations.

[0090] According to another embodiment of the present disclosure, individual learning rate adjustments may be performed by considering the change amount of each parameter, while simultaneously providing global scheduling of the overall learning rate. This allows for the simultaneous achievement of fast initial convergence and stable final convergence.

[0091] Subsequently, to evaluate the performance of the trained model, it can be tested using, for example, 80 cerebrovascular graph data from 2 individuals. This is intended to verify the model's generalization ability and prediction accuracy.

[0092] FIG. 4 is a flowchart illustrating a three-dimensional medical image processing method according to one embodiment of the present disclosure. FIGS. 5 to 8 are illustrative diagrams illustrating the operation of a three-dimensional medical image processing device according to one embodiment of the present disclosure.

[0093] Referring to FIG. 4, a three-dimensional medical image processing device (100) can analyze a three-dimensional medical image and extract pixel coordinates corresponding to blood vessels based on pixel values ​​(S410).

[0094] In one embodiment of S410, a three-dimensional medical image processing device (100) analyzes a three-dimensional medical image and determines that if any pixel value is greater than or equal to a predetermined threshold value, it is a blood vessel, identifies it as a blood vessel pixel point, and can extract coordinate information of the blood vessel pixel point. Subsequently, the three-dimensional medical image processing device (100) can generate a coordinate system of all blood vessels by displaying the blood vessel points on a coordinate system based on the coordinate information.

[0095] For example, a three-dimensional medical image processing device (100) can analyze the three-dimensional medical image of FIG. 5 (a), extract pixel coordinates corresponding to blood vessels based on pixel values ​​as in FIG. 5 (b), and then generate a pixel coordinate table as in FIG. 6 (a). Next, the three-dimensional medical image processing device (100) can generate a coordinate system of all blood vessels based on the pixel coordinate table as in FIG. 6 (b). That is, a coordinate system of all blood vessels can be generated by forming blood vessel pixel points on the coordinate system.

[0096] Referring again to FIG. 4, the 3D medical image processing device (100) can crop a major blood vessel portion from a blood vessel pixel point displayed on a coordinate system (S420). For example, the 3D medical image processing device (100) can identify a major blood vessel portion to be the subject of image processing from a total blood vessel pixel point displayed on a coordinate system as in FIG. 7 (a), as in FIG. 7 (b).

[0097] Referring again to FIG. 4, the 3D medical image processing device (100) can remove noise and / or microvascular pixels by performing clustering and filtering on the vascular pixel points of the crop area (S430). Through this, the 3D medical image processing device (100) can generate a 3D vascular image (S440).

[0098] For example, a three-dimensional medical image processing device (100) can group blood vessel pixel points into clusters by performing clustering based on one or more of the distance, density, and / or attribute information of the blood vessel pixel points. Subsequently, the three-dimensional medical image processing device (100) can perform filtering by comparing the number of blood vessel points included in any cluster with a threshold value and deleting clusters where the number of blood vessel points is less than or equal to the threshold value. Filtering can be performed by removing pixel points corresponding to clusters where the number of blood vessel points is less than or equal to the threshold value from the blood vessel coordinate system. Through this, the three-dimensional medical image processing device (100) can remove noise around major blood vessels.

[0099] Furthermore, the 3D medical image processing device (100) can group the blood vessel pixel points into clusters by performing clustering again on the first filtered blood vessel pixel points based on one or more of their distance, density, and attribute information. Subsequently, the 3D medical image processing device (100) can perform a second filtering to delete the remaining clusters, excluding the blood vessel pixel points included in any cluster. For example, the second filtering can be performed by removing pixel points corresponding to clusters other than the largest cluster from the blood vessel coordinate system. Through this, the 3D medical image processing device (100) can remove microvessels.

[0100] For example, a three-dimensional medical image processing device (100) can remove noise and microvascular pixels from a blood vessel that is the subject of image processing as in (a) of FIG. 8 and generate a three-dimensional blood vessel image as in (b) of FIG. 8.

[0101] FIG. 9 is a flowchart illustrating a method for generating a blood vessel graph according to one embodiment of the present disclosure. FIGS. 10 to 14 are illustrative diagrams illustrating the operation of a blood vessel graph generating device according to one embodiment of the present disclosure.

[0102] Referring to FIG. 9, the blood vessel graph generating device (200) can divide a three-dimensional blood vessel image into multiple layers and check the coordinate values ​​of blood vessel pixel points for each layer. At this time, it is appropriate to generate the layers in two dimensions (S910). For example, the blood vessel graph generating device (200) can divide a three-dimensional blood vessel image into multiple layers as in FIG. 10 (a) and check the coordinate values ​​of blood vessel pixel points for each layer as in FIG. 10 (b).

[0103] Referring again to FIG. 9, after S910, the blood vessel graph generating device (200) can group blood vessel pixel points for each layer (S920). Furthermore, the blood vessel graph generating device (200) can calculate the centroid of each group and generate nodes of the blood vessel graph based on the calculated centroid of each group (S930). Specifically, the blood vessel graph generating device (200) can set the centroid of the group of blood vessel pixel points for each layer as nodes of the blood vessel graph.

[0104] For example, the blood vessel graph generating device (200) can group blood vessel pixel points in any layer as in FIG. 11 and form the centroid of each group as a node of the blood vessel graph. In the example of FIG. 11, the group of blood vessel pixel points corresponds to a cross-section of an actual blood vessel, and the centroid of the group corresponds to the center of a cross-section of an actual blood vessel. Therefore, it is appropriate to set the radius of the node of the blood vessel graph to the actual radius of the blood vessel.

[0105] Referring again to FIG. 9, after S930, the blood vessel graph generating device (200) can generate connectivity information of the nodes of the blood vessel graph (S940).

[0106] For example, the blood vessel graph generating device (200) checks the blood vessel pixel point group of the i-th layer and the blood vessel pixel point group of the i+1-th layer adjacent to the i-th layer as shown in FIG. 12. As a result of the check, if there is an intersection between the blood vessel pixel point group of the i-th layer and the blood vessel pixel point group of the i+1-th layer as shown in FIG. 13, connection information between the nodes can be generated. The blood vessel pixel point group of each layer corresponds to a cross-section of an actual blood vessel, and the fact that the blood vessel pixel point group of one layer has an intersection with the blood vessel pixel point group of the layer adjacent to that layer can be interpreted as the actual blood vessels being connected.

[0107] Referring again to FIG. 9, after S940, the blood vessel graph generating device (200) can calculate the actual radius of the blood vessel in the blood vessel graph and set it as the radius of the node (S950).

[0108] For example, referring to Fig. 14, when a 3D blood vessel image is divided into 2D layers, the radius of the cross-section corresponds to the radius (R0) of the blood vessel pixel point group. That is, since the radius of the cross-section is not the actual radius of the blood vessel but the radius (R0) of the blood vessel pixel point group, it is necessary to calculate the actual radius (r) of the blood vessel. To do this, the radius (R0) in the blood vessel pixel point group can be transformed by cosθ, and the actual radius (r) can be calculated by the following mathematical formula 1.

[0109]

[0110] In mathematical formula 1, r represents the actual radius of the blood vessel, and R0 represents the radius of the blood vessel pixel point group. And cosθ is the direction of connection information from the corresponding layer to the adjacent layer ( ) and the radial direction of the blood vessel pixel point group of the corresponding layer ( It is calculated using ). That is, cosθ can be calculated by the following mathematical formula 2.

[0111]

[0112] : Length direction calculated from connection information, : Radial direction of the blood vessel cross-section, or

[0113] : Perpendicular to the length direction calculated from the connection information, : Vertical of the blood vessel pixel point group

[0114] Referring again to FIG. 9, after S950, the blood vessel graph generating device (200) can generate a blood vessel graph using nodes, connection information, and the radius of the nodes (S960). The blood vessel graph is a visual representation of a blood vessel structure, in which information regarding characteristics such as branching, connection, size, and location of the blood vessels is compressed, and parameters such as the length, diameter, angle, branching pattern, and shape of the blood vessels are included.

[0115] FIG. 15 is a flowchart illustrating a method for generating a blood flow data prediction model according to one embodiment of the present disclosure. FIG. 16 and 17 are drawings illustrating the structure and operation of a blood flow data prediction model according to one embodiment of the present disclosure.

[0116] Referring to FIG. 15, the blood flow data prediction system can collect training data for training a blood flow data prediction model (S1501).

[0117] The training data may be a vascular graph for any region as illustrated in FIG. 16. The vascular graph may be in the form of a point cloud and may be configured as a 1D network connecting points (or nodes). Specifically, the vascular graph may include boundary nodes (1601) located at the boundary of the region and non-boundary nodes (1602) located inside the region. Furthermore, the vascular graph may include coordinate values ​​consisting of the x, y, and z coordinates of each node, and radius information (r) of the node corresponding to the actual radius of the blood vessel. Furthermore, the training data may further include pressure values ​​(p) of the boundary nodes (1601), and the pressure values ​​(p) of the non-boundary nodes may be set as labels or masks.

[0118] Referring again to FIG. 15, after S1501, the blood flow data prediction system can set the input layer of the blood flow data prediction model (S1502).

[0119] According to an embodiment, the input layer of the blood flow data prediction model can be set to a first channel, a second channel, and a third channel, as shown in 1701 of FIG. 17. The first channel receives coordinate values ​​and radius information (x,y,z,r) of a query node in a blood vessel graph, the second channel receives coordinate values ​​and radius information (x,y,z,r) of a boundary node in a blood vessel graph, and the third channel receives pressure values ​​(p) of a boundary node.

[0120] Referring again to FIG. 15, after S1502, the blood flow data prediction system can set the feature extraction layer of the blood flow data prediction model (S1503).

[0121] According to an example, the feature extraction layer of the blood flow data prediction model can be referenced to 1702 in FIG. 17.

[0122] The feature extraction layer (1702) can separate the pressure value (p) of the boundary node input from the third channel into four values ​​corresponding to the coordinate value of the boundary node and the radius information (x,y,z,r) in order to reflect the correlation between the coordinate value of the boundary node and the radius information (x,y,z,r). Furthermore, the feature extraction layer (1702) can pass the (x,y,z,r) input from the first channel, the (x,y,z,r) input from the second channel, or the four separated pressure values ​​(p) to a fully connected layer (e.g., FC1, FC2) that shares weights.

[0123] Through this, the feature extraction layer (1702) can form local features (1705) that reflect the characteristics of the query node and boundary features (1706) that reflect the characteristics of the boundary node. At this time, it is appropriate to apply a max pooling layer (1707) to remove the input order information of the boundary node from the boundary features (1706).

[0124] Referring again to FIG. 15, after S1503, the blood flow data prediction system can set a feature stitching layer (S1504).

[0125] According to an embodiment, the feature stitching layer of the blood flow data prediction model may be structured to combine local features (1705) and boundary features (1706), as shown in 1703 of FIG. 17, and pass the combined features to a fully connected layer (e.g., FC1, FC2). This is so that the blood flow data prediction model can understand global features for the entire channel.

[0126] Referring again to FIG. 15, after S1504, the blood flow data prediction system can set the output layer of the blood flow data prediction model (S1505).

[0127] According to an embodiment, the output layer of the blood flow data prediction model may be structured to pass features output from a feature stitching layer to a fully connected layer (e.g., FC5, FC6), as shown in 1704 of FIG. 17, and may be configured to predict and output pressure values ​​of query nodes.

[0128] Referring again to FIG. 15, when training data for a blood flow data prediction model is prepared (S1501) and the structure of the model is set up (S1502 to S1505), the blood flow data prediction system can train the blood flow data prediction model (S1506).

[0129] Once the creation of the blood flow data prediction model is complete, the blood flow data prediction system can specify a blood vessel graph of the region of interest, distinguish between a boundary node located at the end of the specified blood vessel graph and a non-boundary node located inside the specified blood vessel graph, obtain location information (x,y,z) and radius information (r) of the boundary node and the non-boundary node, obtain the pressure value (p) of the boundary node, and input the obtained values ​​into the trained blood flow data prediction model. The blood flow data prediction model can predict the pressure value (P) of an arbitrary node (e.g., a query node) and output the predicted pressure value (P). Through this, the blood flow data prediction system can predict blood flow data (P) at an arbitrary point.

[0130] FIG. 18 is a block diagram illustrating the configuration of a server that generates a blood flow data prediction model according to one embodiment of the present disclosure.

[0131] As illustrated in FIG. 18, a server (1800) according to one embodiment of the present disclosure may include a communication unit (1810), a storage unit (1830), and a control unit (1820), and may further include an input unit and a display unit, although not illustrated in FIG. 18.

[0132] In FIG. 18, the server (1800) is depicted as including a communication unit (1810), a storage unit (1830), and a control unit (1820), but each block may exist physically separated. For example, the storage unit (1830) may exist in a virtualized data center and may be connected to the control unit (1820) of the server (1800) through the communication unit (1810).

[0133] The communication unit (1810) performs the function of transmitting and receiving data for wired and wireless communication of the server (1800), receives data through a wired and wireless channel and outputs it to the control unit (1820), and performs the function of transmitting the data output from the control unit (1820) through a wired channel. In particular, the communication unit (1810) can perform the function of receiving three-dimensional medical images.

[0134] The storage unit (1830) performs the role of storing programs and data necessary for the operation of the server (1800) and can be divided into a program area and a data area. The data area of ​​the storage unit (1830) can store coordinates of blood vessel pixel points, 3D blood vessel images and blood vessel graphs, and blood flow data (e.g., pressure values) at any point of the blood vessel.

[0135] Furthermore, the program area of ​​the storage unit (1830) can store a blood flow data prediction model, which is an artificial intelligence model that predicts blood flow data at any point in a blood vessel image. The blood flow data prediction model can perform the function of predicting the pressure value of any node by using the location information and radius information of each node of the blood vessel graph of the area of ​​interest.

[0136] Furthermore, the program area of ​​the storage unit (1830) can store a computer program that executes a process to generate a blood vessel graph on a server. The computer program can perform functions such as analyzing a three-dimensional medical image to extract coordinates of blood vessel pixel points corresponding to blood vessels based on pixel values, removing noise around major blood vessels, and generating a three-dimensional blood vessel image, and dividing the three-dimensional blood vessel image into multiple layers, setting groups of blood vessel pixel points for each layer as blood vessel nodes, and generating connection information between blood vessel nodes to generate a blood vessel graph.

[0137] The control unit (1820) controls the overall operation of each component of the server (1800). In particular, the control unit (1820) can set the structure of the blood flow data prediction model and generate data for training the blood flow data prediction model.

[0138] The control unit (1820) can generate blood vessel data in the form of a point cloud corresponding to major blood vessels from a three-dimensional medical image, specify a boundary area of ​​the blood vessel data, obtain a blood flow value of the boundary area, and specify the location of the point cloud for an arbitrary area and the blood flow value of the area boundary to generate training data for a blood flow data prediction model.

[0139] The control unit (1820) can set the structure of the blood flow data prediction model, and the blood flow data prediction model may include an input layer, a feature extraction layer, and an output layer. As an example, the control unit (1820) can set the input layer of the blood flow data prediction model to include a first channel into which location information and radius information of a non-boundary node (e.g., a query node) in a blood vessel graph of a region of interest are input, a second channel into which location information and radius information of a boundary node in a blood vessel graph of a region of interest are input, and a third channel into which blood flow values ​​of the boundary node are input. As another example, the control unit (1820) can set the feature extraction layer to form a local feature that reflects the characteristics of a non-boundary node (e.g., a query node), and a boundary feature that reflects the characteristics of a boundary node and has the input order information of the boundary node removed.

[0140] After going through the process described above, training data for the blood flow data prediction model is prepared and the structure setting of the model is completed, the control unit (1820) can train the blood flow data prediction model.

[0141] Furthermore, the control unit (1820) analyzes a three-dimensional medical image to extract coordinates of blood vessel pixel points corresponding to blood vessels based on pixel values, removes noise around major blood vessels, generates a three-dimensional blood vessel image, divides the three-dimensional blood vessel image into multiple layers, sets blood vessel pixel point groups for each layer as blood vessel nodes, generates connection information between blood vessel nodes to generate a blood vessel graph, and can control the communication unit (1810) to provide the blood vessel graph.

[0142] Although the present disclosure has been described by limited embodiments and drawings, it is not limited to the above embodiments, and various modifications and variations are possible from this description by those skilled in the art to which the present disclosure pertains. Accordingly, the concept of the present disclosure should be understood only by the claims set forth below, and all equivalent or analogous variations thereof shall be considered to be within the scope of the concept of the present disclosure.

[0143] As described above, the method for predicting brain disease risk, the learning method for the brain disease risk analysis model, and the brain disease risk analysis device can be applied to the medical service field that provides medical services.

Claims

1. In a method for generating an artificial intelligence model that predicts blood flow data on a server, Step of establishing the structure of the blood flow data prediction model; A step of generating vascular data in the form of a point cloud corresponding to major blood vessels from a 3D medical image, specifying a boundary region of the vascular data, obtaining a blood flow value of the boundary region, and preparing the location of the point cloud for an arbitrary region and the blood flow value of the region boundary as training data for the blood flow data prediction model; and A method comprising the step of training the blood flow data prediction model by individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter, and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function. Method for generating artificial intelligence models.

2. In Paragraph 1, The step of preparing the training data for the above blood flow data prediction model is, A step comprising performing data augmentation by modifying node radius information and boundary conditions in a blood vessel graph for an arbitrary region, Method for generating artificial intelligence models.

3. In Paragraph 1, The step of establishing the structure of the above blood flow data prediction model is, A step comprising setting a feature extraction layer of the blood flow data prediction model to form local features reflecting the characteristics of the other nodes and boundary features reflecting the characteristics of the boundary nodes. Method for generating artificial intelligence models.

4. In Paragraph 3, The step of setting the feature extraction layer of the above blood flow data prediction model is, A step comprising forming the boundary feature by removing the input order information of the boundary node. Method for generating artificial intelligence models.

5. A 3D medical image processing device that analyzes a 3D medical image, extracts the coordinates of blood vessel pixel points corresponding to blood vessels based on pixel values, and generates a 3D blood vessel image; A blood vessel graph generating device that generates a blood vessel graph from a three-dimensional blood vessel image using the blood vessel pixel point group as a node; and An artificial intelligence model generation device comprising: establishing the structure of a blood flow data prediction model; specifying a boundary region of the blood vessel graph; obtaining blood flow values ​​of the boundary region; preparing the location of an arbitrary region and the blood flow values ​​of the region boundary as training data for the blood flow data prediction model; individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter; and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function to train the blood flow data prediction model and generate the blood flow data prediction model. Artificial intelligence model generation system.

6. In a server that generates an artificial intelligence model for predicting blood flow data, A communication unit for receiving a 3D medical image; and A control unit comprising: establishing the structure of a blood flow data prediction model; generating vascular data in the form of a point cloud corresponding to major blood vessels from the three-dimensional medical image; specifying a boundary region of the vascular data; obtaining blood flow values ​​of the boundary region; preparing the location of the point cloud for an arbitrary region and the blood flow values ​​of the region boundary as training data for the blood flow data prediction model; individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter; and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function to train the blood flow data prediction model and generate the blood flow data prediction model. AI model generation server.

7. A non-transient computer-readable recording medium having a computer program that is executed by a processor to execute a method for generating an artificial intelligence model that predicts blood flow data, The above method is, Step of establishing the structure of the blood flow data prediction model; A step of generating vascular data in the form of a point cloud corresponding to major blood vessels from a 3D medical image, specifying a boundary region of the vascular data, obtaining a blood flow value of the boundary region, and preparing the location of the point cloud for an arbitrary region and the blood flow value of the region boundary as training data for the blood flow data prediction model; and A method comprising the step of training the blood flow data prediction model by individually adjusting the learning rate of the blood flow data prediction model for the training data according to the amount of change of each parameter, and simultaneously setting the overall learning rate to gradually decrease in the form of a cosine function. Non-transient computer-readable recording medium.

Images