Prediction model training method, ffr data generation method, and electronic device for carrying out same

Abstract

A prediction model training method according to an embodiment comprises the steps of: generating a vascular model on the basis of a vascular image; determining a plurality of physical models corresponding to the vascular model, wherein each of the plurality of physical models includes a plurality of nodes; determining, by using an artificial neural network, a plurality of parameters on the basis of the vascular model; on the basis of the plurality of physical models and the plurality of parameters, calculating a pressure loss value at one node among the plurality of nodes; obtaining an actual loss value at the one node; and training the artificial neural network on the basis of the pressure loss value and the actual loss value.

Description

Predictive model training method, FFR data generation method, and electronic device for performing the same

[0001] The disclosure relates to a method for training a prediction model, a method for generating FFR data, and an electronic device for performing the same.

[0002] Coronary artery disease is a condition that can lead to serious health problems by reducing blood flow to the myocardium. However, relying solely on the degree of coronary artery stenosis (the ratio of narrowing) makes it difficult to accurately determine how much that stenosis actually affects myocardial blood flow. To address this, the concept of Fractional Flow Reserve (FFR) was introduced, which enables a quantitative assessment of blood flow resistance occurring at the site of stenosis. FFR represents the ratio of proximal to distal pressure under maximum blood flow conditions in a stenotic vessel and is used to evaluate the functional severity of the stenosis.

[0003] Traditional methods relied on catheter-based techniques that directly measured pressure at specific points within the coronary arteries; however, due to these invasive and high-cost natures, methods utilizing vascular physical models have recently been researched and utilized. Vascular physical models for FFR calculation simulate blood flow within the coronary arteries by applying the principles of fluid dynamics, and are non-invasive and cost-effective.

[0004] One embodiment aims to provide a prediction model training method, a FFR data generation method, and an electronic device for performing the same.

[0005] One embodiment aims to provide a method and apparatus for predicting personalized parameters that are difficult to extract as features based on an artificial neural network.

[0006] One embodiment aims to provide a method and apparatus for predicting hemodynamic indicators without using a lot of time and high-performance computing resources due to complex computational processes such as computational fluid dynamics (CFD) simulations.

[0007] One embodiment aims to provide a method and apparatus for obtaining a result close to the ground truth by allowing the predicted value of an artificial neural network to be applied to various types of pressure model equations and to participate in combining them.

[0008] One embodiment aims to provide a method and apparatus that enable interpretation by setting constraints on physical phenomena during the learning process, unlike conventional artificial neural networks.

[0009] However, the problems that the present invention aims to solve are not limited to those mentioned above, and may include problems that are not mentioned but can be clearly understood by those skilled in the art from the description below.

[0010] A method for training a prediction model according to one embodiment for solving such technical problems comprises the steps of: generating a plurality of vascular models based on a vascular image, wherein each of the plurality of vascular models includes a plurality of nodes; determining a plurality of coefficients to be applied to the plurality of vascular models based on the vascular image and an artificial neural network; calculating a pressure loss value at one of the plurality of nodes based on the plurality of vascular models and the plurality of coefficients; obtaining an actual loss value at the node; and training the artificial neural network based on the pressure loss value and the actual loss value.

[0011] A method for generating FFR data according to one embodiment comprises the steps of: acquiring a blood vessel image; determining a plurality of nodes in the blood vessel image; determining a plurality of node information of the plurality of nodes; determining a plurality of coefficients from the plurality of node information using an artificial neural network; generating a plurality of blood vessel models based on the blood vessel image; calculating a pressure loss value based on the plurality of blood vessel models and the plurality of coefficients; and generating FFR data based on the pressure loss value, wherein the artificial neural network takes node data as input data, takes prediction coefficients corresponding to the node data as output data, and is trained in a direction that minimizes the difference between the predicted loss value calculated based on the prediction coefficients and the plurality of blood vessel models and the actual loss value.

[0012] A method for training a prediction model according to one embodiment for solving such technical problems comprises: generating a blood vessel model based on a blood vessel image; determining a plurality of physical models corresponding to the blood vessel model, wherein each of the plurality of physical models includes a plurality of nodes; determining a plurality of parameters based on the blood vessel model using an artificial neural network; calculating a pressure loss value at one of the plurality of nodes based on the plurality of physical models and the plurality of parameters; obtaining an actual loss value at the one node; and training the artificial neural network based on the pressure loss value and the actual loss value.

[0013] A method for generating FFR data according to one embodiment comprises the steps of: acquiring a blood vessel image; generating a blood vessel model based on the blood vessel image; determining a plurality of physical models corresponding to the blood vessel model; determining a plurality of parameters based on the blood vessel model using an artificial neural network; and generating FFR data based on the plurality of physical models and the plurality of parameters, wherein the artificial neural network takes a blood vessel model for training as input data and a plurality of prediction parameters corresponding to the blood vessel model for training as output data, and is trained in a direction that minimizes the difference between a predicted loss value calculated based on the plurality of prediction parameters and the plurality of physical models and an actual loss value.

[0014] An electronic device according to one embodiment includes a processor and a memory connected to the processor, the memory is configured to store a program, the processor is configured to execute the program, and when the program is executed, the steps of the method according to the embodiments are implemented.

[0015] FIG. 1 is a schematic block diagram of a computing system according to one embodiment.

[0016] FIG. 2 is a block diagram of an electronic device according to one embodiment.

[0017] FIG. 3 is a flowchart of a method for training a prediction model according to one embodiment.

[0018] FIG. 4 is a diagram illustrating a training session of a prediction model according to one embodiment.

[0019] FIG. 5 is a block diagram illustrating a configuration for generating combined data according to one embodiment.

[0020] FIG. 6 is a flowchart of a method for generating FFR data according to one embodiment.

[0021] FIG. 7 is an example of a blood vessel image according to one embodiment.

[0022] FIG. 8 is an example of a blood vessel graph according to one embodiment.

[0023] FIG. 9 is a diagram illustrating an inference session of a prediction model according to one embodiment.

[0024] FIG. 10 is a diagram illustrating a learning session of an artificial neural network according to one embodiment.

[0025] The various embodiments described in this specification are illustrative for the purpose of clearly explaining the technical concept of this disclosure and are not intended to limit it to specific embodiments. The technical concept of this disclosure includes various modifications, equivalents, alternatives, and embodiments optionally combined from all or part of each embodiment described in this specification. Furthermore, the scope of the technical concept of this disclosure is not limited to the various embodiments presented below or the specific descriptions thereof.

[0026] Terms used in this specification, including technical or scientific terms, may have the meaning generally understood by those skilled in the art to which this disclosure pertains, unless otherwise defined.

[0027] Expressions used herein such as “comprising,” “may compose,” “possessing,” “possessing,” “having,” and “possessing” imply the existence of the subject feature (e.g., function, operation, or component, etc.) and do not exclude the existence of other additional features. That is, such expressions should be understood as open-ended terms implying the possibility of including a second embodiment.

[0028] In this specification, singular expressions include plural expressions unless the context clearly specifies them as singular. Additionally, plural expressions include singular expressions unless the context clearly specifies them as plural. Throughout the specification, when a part is described as including a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0029] Additionally, the terms 'module' or 'part' as used in the specification refer to software or hardware components, and the 'module' or 'part' performs certain roles. However, the meaning of 'module' or 'part' is not limited to software or hardware. The 'module' or 'part' may be configured to reside in an addressable storage medium or configured to run on one or more processors. Thus, as an example, the 'module' or 'part' may include components such as software components, object-oriented software components, class components, and task components, and at least one of processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, or variables. The components and the functions provided within the 'module' or 'part' may be combined into a smaller number of components and 'modules' or 'parts', or further separated into additional components and 'modules' or 'parts'.

[0030] According to one embodiment of the present disclosure, a ‘module’ or ‘part’ may be implemented as a processor and memory. The term ‘processor’ should be broadly interpreted to include a general-purpose processor, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a controller, a microcontroller, a state machine, etc. In some environments, the term ‘processor’ may refer to an application-specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), etc. The term ‘processor’ may also refer to a combination of processing devices, such as, for example, a combination of a DSP and a microprocessor, a combination of multiple microprocessors, a combination of one or more microprocessors combined with a DSP core, or any other combination of such configurations. Additionally, the term ‘memory’ should be broadly interpreted to include any electronic component capable of storing electronic information. 'Memory' may refer to various types of processor-readable media, such as Random Access Memory (RAM), Read-Only Memory (ROM), Non-Volatile Random Access Memory (NVRAM), Programmable Read-Only Memory (PROM), Erasable-Programmable Read-Only Memory (EPROM), Electrically Erasable PROM (EEPROM), Flash Memory, Magnetic or Optical Data Storage Devices, Registers, etc. If a processor can read information from memory and / or write information to memory, the memory is said to be in an electronic communication state with the processor. Memory integrated into a processor is in an electronic communication state with the processor.

[0031] Expressions such as "first," "second," or "first," "second" as used in this specification are used to distinguish one object from another when referring to a plurality of objects of the same kind, unless otherwise indicated in the context, and do not limit the order or importance of said objects.

[0032] Expressions used herein such as “A, B, and C,” “A, B, or C,” “A, B, and / or C,” or “at least one of A, B, and C,” “at least one of A, B, or C,” “at least one of A, B, and / or C,” “at least one selected from A, B, and C,” “at least one selected from A, B, or C,” “at least one selected from A, B, and / or C,” etc., may mean each of the listed items or all possible combinations of the listed items. For example, “at least one selected from A and B” may refer to (1) A, (2) at least one of A, (3) B, (4) at least one of B, (5) at least one of A and at least one of B, (6) at least one of A and B, (7) at least one of B and A, and (8) all of A and B.

[0033] As used herein, the expression “based on” is used to describe one or more factors affecting an act or action of a decision or judgment described in the phrase or sentence containing such expression, and such expression does not exclude additional factors affecting said act or action of a decision or judgment.

[0034] As used in this specification, the expression that a certain component (e.g., a first component) is "connected" or "connected" to another component (e.g., a second component) may mean that the said certain component is not only directly connected or connected to the said other component, but is also connected or connected through a new other component (e.g., a third component).

[0035] As used herein, the expression "configured to" may have meanings such as "set to," "capable of," "modified to," "made to," or "capable of." Such expression is not limited to the meaning of "specifically designed in hardware," and, for example, a processor configured to perform a specific operation may mean a generic-purpose processor capable of performing that specific operation by executing software.

[0036] Various embodiments of the present disclosure will be described below with reference to the accompanying drawings. In the accompanying drawings and the description thereof, identical or substantially equivalent components may be given the same reference numerals. Furthermore, in the description of the various embodiments below, the description of identical or corresponding components may be omitted, but this does not mean that such components are not included in the embodiments.

[0037]

[0038] FIG. 1 is a schematic block diagram of a computing system according to one embodiment, and FIG. 2 is a block diagram of an electronic device according to one embodiment.

[0039] Referring to FIG. 1, a computing system (10) according to one embodiment may acquire a blood vessel image of a first user, perform calculations based on the acquired blood vessel image, and display a processed image. The computing system (10) may provide the processed image to a second user. For example, the first user may be a patient, and the second user may be a medical professional.

[0040] In one embodiment, the computing system (10) may obtain a cardiovascular image by photographing the first user's cardiovascular system and perform calculations based on the cardiovascular image. For example, the computing system may determine pressure losses at each node of the blood vessel through calculations and determine the blood vessel FFR (Fractional Flow Reserve) value based on the pressure losses. According to an embodiment, the computing system (10) may generate three-dimensional data by performing image processing on the cardiovascular image. However, the embodiment is not necessarily limited thereto, and the computing system (10) may photograph blood vessels subject to angiography, such as cerebral blood vessels and gastrointestinal blood vessels.

[0041] A computing system (10) according to one embodiment includes a shooting device (100) and an electronic device (200). The shooting device (100) may be a shooting device configured to capture a first user and acquire an image. For example, the shooting device (100) may be an X-ray imaging device (e.g., C-arm X-ray device), angiography imaging device (e.g., angio device), optical coherence tomography (OCT) device, computed tomography (CT) device, magnetic resonance imaging (MRI), magnetic resonance angiography (MRA) device, etc., but the embodiment is not necessarily limited thereto and may be implemented as various devices configured to capture the blood vessels of the first user.

[0042] The imaging device (100) can photograph a first user at multiple imaging points and acquire multiple images. In one embodiment, the imaging device (100) can photograph the first user while rotating around the first user. In another embodiment, the imaging device (100) can rotate the first user and photograph the rotating first user. The imaging device (100) can transmit the acquired multiple images to an electronic device (200). Here, the images acquired and transmitted by the imaging device (100) may include X-ray images, ultrasound (or sonography) images, CT (Computed Tomography) images, PET (Positron Emission Tomography) images, MRI (Magnetic Resonance Imaging) images, fMRI (functional Magnetic Resonance Imaging) images, digital pathology WSI (WSI) images, DBT (Digital Breast Tomosynthesis) images, etc.

[0043] The electronic device (200) may be a computing device configured to calculate an FFR value based on an image received from a shooting device (100). The electronic device (200) may be an artificial intelligence (AI) device, a personal computer (PC), a laptop computer, a mobile phone, a smartphone, a tablet PC, a wearable device, medical imaging equipment, a healthcare device, etc.

[0044] In some embodiments, the computing system (10) may further include a server such as an AI server or a data center, and the electronic device (200) may be implemented to communicate with the server. That is, the electronic device (200) may use the artificial neural network of the server.

[0045] Referring to FIG. 2, an electronic device (200) according to one embodiment may include a processor (210) and a memory (220) connected to the processor (210). The memory (220) may be configured to store a program. The processor (210) may be configured to execute a program in the memory (220). When the program is executed, steps of the method according to the embodiments may be implemented.

[0046] The processor (210) can perform calculations based on an image received from the imaging device (100) and calculate an FFR value. The processor (210) can perform calculations using an artificial neural network and a vascular model. A vascular model may refer to a physical model for calculating pressure loss values ​​at each node of a blood vessel. Pressure loss values ​​may be factors for calculating the FFR value. The processor (210) may generate a plurality of vascular models based on the image. The plurality of vascular models may include at least one of a friction model, an expansion model, a contraction model, a vortex model, a dissection model, and a branching model, but the embodiment is not necessarily limited thereto, and the plurality of vascular models may be implemented to include a wider variety of physical models.

[0047] The processor (210) may use an artificial neural network trained to output coefficients for application to a blood vessel model. For example, the artificial neural network may be implemented as a Graph Neural Network (GNN), Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Generative Adversarial Network (GAN), Multi-Layer Perceptron (MLP), Transformer, Residual Network (ResNet), U-Net, a hybrid model, etc., but the embodiments are not necessarily limited thereto. A hybrid model may be composed of a combination of a GNN and a CNN, a combination of a GNN and a Transformer, etc., to analyze blood vessel branches and connectivity. In the embodiments, the electronic device (200) may be configured to train the artificial neural network. According to the embodiment, the artificial neural network may be stored in a device outside the electronic device (100) or in memory (220).

[0048] The processor (210) can calculate the FFR value based on the pressure loss value obtained using a vascular model and the coefficient obtained using an artificial neural network. In this way, the electronic device (200) can achieve high FFR value prediction performance while using relatively little data by using a vascular model and an artificial neural network.

[0049] In FIG. 2, for convenience of explanation, the electronic device (200) is depicted as including a processor (210) and a memory (220), but the embodiment is not necessarily limited thereto, and the electronic device (200) may be implemented to include at least one additional component. For example, the electronic device (200) may further include components such as an input / output interface, a communication module, and a display. The input / output interface may be a component for an interface between an input / output device and the electronic device (200). The communication module may be a component for communication with a device outside the electronic device (200) (e.g., a camera (100), a server, etc.). The display may be a component for displaying the output of the processor (210) (e.g., a processed image, an FFR value, etc.).

[0050]

[0051] FIG. 3 is a flowchart of a method for training a prediction model according to one embodiment.

[0052] Referring to FIG. 3, a method for training a prediction model according to one embodiment may be performed by an electronic device (e.g., 200 in FIG. 1). The prediction model may be understood to include a blood vessel model and an artificial neural network. The electronic device may generate a plurality of blood vessel models based on images of blood vessels (BV) (S310). The electronic device may model blood vessels by generating a graph from the blood vessel images. Generating a graph may be understood as generating nodes and edges connecting the nodes. In one embodiment, the electronic device may use a GNN to generate a graph including nodes and edges. The blood vessel model is a physical model for calculating pressure loss within blood vessels through mathematical modeling or computational fluid dynamics analysis, and may utilize graph-based blood vessel extraction algorithms, hierarchical structure analysis techniques, segmentation techniques, blood vessel centerline extraction algorithms, etc.

[0053] Vascular models can be classified into friction models, dilation models, constriction models, vortex models, dissection models, branching models, etc., depending on the characteristics of the blood vessels. For example, a friction model models pressure changes caused by viscous friction between the blood vessel wall and the blood flow, a dilation model models pressure changes that occur when a blood vessel dilates, a constriction model models pressure changes that occur when a blood vessel narrows, a vortex model models pressure changes that occur as vortices form within the blood vessel, a dissection model models pressure changes that occur when the boundary layer of the blood flow is dissected after the bending or stenosis of the blood vessel, and a branching model models pressure changes when a blood vessel branches. An electronic device can generate at least one of these models as a plurality of vascular models.

[0054] In one embodiment, the electronic device can determine a target blood vessel in a blood vessel image. Determining the target blood vessel may be performed based on user selection or automatically based on the computing of the electronic device. For example, the electronic device may segment regions containing blood vessel structures in the blood vessel image. The electronic device may select a target blood vessel from the segmented blood vessel regions based on certain criteria. For example, certain criteria may include hemodynamic importance, presence of lesions, anatomical location, etc. The electronic device may generate a plurality of blood vessel models based on the target blood vessel.

[0055] Each of the multiple blood vessel models may include multiple nodes. In one embodiment, the electronic device may determine multiple nodes based on at least one of a bifurcation region, a line region, and a lesion region in the target blood vessel. The bifurcation region refers to a point where the blood vessel branches, and the electronic device may determine nodes based on the branching point and angle of the blood vessel in the bifurcation region. The line region refers to a section where the blood vessel centerline extends, and the electronic device may determine multiple nodes based on at least one of node length, cross-sectional area, and curvature in the line region. The lesion region refers to a pathological deformation site that affects blood flow characteristics, and the electronic device may determine nodes based on pathological characteristics that affect blood flow in the lesion region.

[0056] The electronic device can determine multiple coefficients to be applied to multiple vascular models based on vascular images and an artificial neural network (S320). The electronic device can acquire node information based on vascular images. The node information may include anatomical information of the blood vessels, such as shape information, image information, hemodynamic information, and location information.

[0057] The electronic device can generate shape information for each node from data extracted from a blood vessel image. The shape information may include quantifiable physical features such as the diameter, curvature, length, degree of stenosis, and changes in cross-sectional area of ​​the blood vessel. The electronic device can derive image information based on the shooting angle and blood vessel type acquired through the blood vessel image. The image information is used to define the relative positions between nodes and the characteristics of the blood vessels, and may include the brightness, contrast, and edge characteristics of pixels of the node and adjacent nodes. The electronic device can calculate pressure loss and blood flow characteristics at each node by applying various hemodynamic models such as friction, vortex, dissection, and branching. The electronic device can generate position information for each node from the blood vessel image, and the position information may include connectivity between nodes and the cumulative distance from a specific point within the blood vessel.

[0058] In the embodiments, the electronic device may further acquire physiological information. The physiological information may include clinical information such as age, gender, height, weight, medical history (e.g., underlying diseases, drug history, etc.), blood pressure, and heart rate. For example, a second user may input physiological information into the electronic device. According to the embodiments, the electronic device may retrieve physiological information stored in memory (or a database).

[0059] The electronic device can input node information into an artificial neural network. The artificial neural network can be configured to output coefficients for application to each blood vessel model. The artificial neural network can generate multiple coefficients from the node information received from the electronic device. In the embodiments, the electronic device further inputs clinical information corresponding to the blood vessel image into the artificial neural network, and the artificial neural network can generate multiple coefficients from the node information and the clinical information.

[0060] The electronic device can calculate a pressure loss value at one of a plurality of nodes based on a plurality of blood vessel models and a plurality of coefficients (S330). The electronic device can calculate a pressure loss value at a node based on a first loss value in a first model among the plurality of blood vessel models, a first coefficient among the plurality of coefficients, a second loss value in a second model among the plurality of blood vessel models, and a second coefficient among the plurality of coefficients.

[0061] For example, the electronic device can generate a friction model as a first model, an expansion model as a second model, and a contraction model, a vortex model, a separation model, and a branching model. The electronic device can determine a first coefficient applied to the friction model, a second coefficient applied to the expansion model, a third coefficient applied to the contraction model, a fourth coefficient applied to the vortex model, a fifth coefficient applied to the separation model, and a sixth coefficient applied to the branching model.

[0062] The electronic device can calculate a pressure loss value based on a first loss value, a first coefficient, a second loss value, a second coefficient, a third loss value and a third coefficient in a contraction model, a fourth loss value and a fourth coefficient in a vortex model, a fifth loss value and a fifth coefficient in a separation model, a sixth loss value in a branching model, and a sixth coefficient. In one embodiment, the electronic device can calculate a pressure loss value from a plurality of loss values ​​and a plurality of coefficients using a plurality of weights. The form of the plurality of weights is not particularly limited, such as linear or exponential.

[0063] In one embodiment, the electronic device can calculate a pressure loss value from a plurality of loss values ​​and a plurality of coefficients using a machine learning model. The machine learning model includes parameters applied to the plurality of loss values ​​and a plurality of coefficients, and can output a pressure loss value based on the parameters.

[0064] In one embodiment, the electronic device can calculate the pressure loss value at one node based on Equation 1.

[0065]

[0066] Here, ΔP node is the pressure loss value at the one node, and ΔP friction is the loss value at one node of the friction model, and ΔP expansion is the loss value at one node of the extension model, and ΔP contraction is the loss value at one node of the contraction model, and ΔP vortex is the loss value at one node of the vortex model, and ΔP separation is the loss value at one node of the peeling model, and ΔP bifurcation is the loss value at one node of the branch model, a is the first coefficient, b is the second coefficient, c is the third coefficient, d is the fourth coefficient, e is the fifth coefficient, and f is the sixth coefficient. In this way, the electronic device can estimate the vascular pressure distribution from various perspectives by combining various types of vascular models.

[0067] The electronic device can obtain the actual loss value at the node (S340). The electronic device can obtain FFR (Fractional Flow Reverse) measurement data. For example, FFR measurement data is obtained by measuring pressure within the coronary artery through a guide wire and a pressure sensor, and may include pullback curve data representing the pressure distribution between nodes. The electronic device can obtain the actual loss value at the node from the FFR measurement data. The actual loss value may be the ground truth used for training the artificial neural network.

[0068] The electronic device can train an artificial neural network based on pressure loss values ​​and actual loss values ​​(S330). The artificial neural network can update weights (or parameters) and biases using backpropagation algorithms and optimization techniques (e.g., Stochastic Gradient Descent (SGD), Adam optimization) by defining the difference between the pressure loss values ​​and the actual loss values ​​as a loss function. Through this, a prediction model specialized for blood vessel image-based hemodynamic analysis can be efficiently trained, and the parameters of the artificial neural network included in the prediction model can be adjusted to more accurately predict the pressure characteristics of a specific node by utilizing blood vessel models and coefficients when the same or similar blood vessel images are input later.

[0069]

[0070] FIG. 4 is a diagram illustrating a training session of a prediction model according to one embodiment.

[0071] Referring to FIG. 4, a training session of a prediction model (400) according to one embodiment is described. Through the training session, the prediction model (400) can have weights (or parameters) for outputting appropriate and precise prediction information (i.e., pressure loss value (PRS_LOSS)). That is, the prediction model (400) can be trained to calculate the pressure loss value (PRS_LOSS) from node information (or node data).

[0072] The prediction model (400) may include an artificial neural network (410) and a plurality of vascular models (420). The artificial neural network (410) may include an input layer, a hidden layer, and an output layer. The artificial neural network (410) may receive node information. The node information may include shape information, image information, hemodynamic information, and location information, etc. The node information may be obtained from a vascular image. According to an embodiment, the artificial neural network (410) may further receive clinical information along with the node information. Since the contents of FIG. 3 can be applied identically to the node information and clinical information, redundant descriptions are omitted.

[0073] Node information can be input to each node of the input layer as input data. The node information of the blood vessel and the nodes of the layer of the artificial neural network (410) can be understood as different. In one embodiment, an electronic device (e.g., 200 in FIG. 1) can generate combined data by combining shape information, image information, hemodynamic information, location information, etc. of the node information.

[0074] In one embodiment, the electronic device may combine shape information, image information, hemodynamic information, location information, etc. using an attention mechanism. For example, the electronic device may learn the importance (influence) of shape information, image information, hemodynamic information, and location information on the pressure loss value (PRS_LOSS), and generate combined data from these data based on the learned parameters. The configuration of the electronic device using the attention mechanism will be described later with reference to FIG. 5. The electronic device may input the combined data into a prediction model (400).

[0075] In one embodiment, the electronic device can generate combined data by combining node information and clinical information. The combined data can be vectorized before being input to the input layer.

[0076] Input data can be computed based on weights and passed to the hidden layer. The computing result of the hidden layer is passed to the output layer, and the output layer can output coefficients (COEEFS) based on an activation function. For example, the activation function can be implemented as softmax, ReLU, sigmoid, hyperbolic tangent (tanH), etc.

[0077] The artificial neural network (410) can transmit coefficients (COEEFS) to a plurality of vascular models (420). The plurality of vascular models (420) can output a loss value for each model. The prediction model (400) can calculate a pressure loss value (PRS_LOSS) based on the model loss value and coefficients (COEEFS) output by each model. For example, the prediction model (400) can calculate the pressure loss value (PRS_LOSS) using Equation 1.

[0078] Back propagation can be performed to update the weights of the hidden layer of the artificial neural network (410) in a direction that reduces the error between the pressure loss value (PRS_LOSS) and the actual loss value (ACT_LOSS) (or ground truth). The error can be calculated using a loss function.

[0079] According to an embodiment, the error may be calculated through a custom loss function that reflects physical-based constraints. Physical-based constraints include continuity equation constraints, Poiseuille's law constraints, etc., and may increase the error of the loss function if the constraints are violated. The continuity equation constraint may mean that the inflow and outflow at each node of the blood vessel must be balanced. The electronic device may reflect the degree of deviation from the continuity equation constraint in the error as a penalty. The Poiseuille's law constraint may mean that a preset relationship between the predicted pressure difference and the flow rate must be satisfied for each edge of the blood vessel. The electronic device may reflect the degree of deviation from the Poiseuille's law constraint in the error as a penalty.

[0080] The artificial neural network (410) trained through such a training session has weights as parameters, and when node information (or node information and clinical information) is input, the artificial neural network (410) can output coefficients (COEFFS) based on the weights. Accordingly, the prediction model (400) can predict an accurate loss value from a vascular image by using the optimal coefficients (COEFFS) output by the artificial neural network (410) and the precise model loss values ​​output by a plurality of vascular models (420).

[0081]

[0082] FIG. 5 is a block diagram illustrating a configuration for generating combined data according to one embodiment.

[0083] Referring to FIG. 5, a combiner (500) according to one embodiment may be configured to generate combined data (COMB_DATA) from node information. The node information may include shape information (shape info), image information (image info), hemodynamic information (hemodynamic(HMDN) info), and location information (location(LCTN) info). However, the embodiment is not necessarily limited thereto, and the node information may be implemented to include at least one of shape information, image information, hemodynamic information, and location information. According to the embodiment, the combiner (500) may be configured to generate combined data (COMB_DATA) from node information and clinical information.

[0084] The combiner (500) may include a QKV generator (510), an attention module (520), and a weighted adder (530). The QKV generator (510) may generate a Q vector, a K vector, and a V vector using a plurality of weight matrices. QKV refers to a query (Q) vector, a key (K) vector, and a value (V) vector, respectively, and may refer to matrices rather than vectors depending on the embodiment.

[0085] For example, the QKV generator (510) may use a first weight matrix for generating a Q vector, a second weight matrix for generating a K vector, and a third weight matrix for generating a V vector. The QKV generator (510) may use the first weight matrix to generate a first Q vector (Q1) of shape information, a second Q vector (Q2) of image information, a third Q vector (Q3) of hemodynamic information, and a fourth Q vector (Q4) of position information. The QKV generator (510) may use the second weight matrix to generate a first K vector (K1) of shape information, a second K vector (K2) of image information, a third K vector (K3) of hemodynamic information, and a fourth K vector (K4) of position information. The QKV generator (510) can generate a first V vector (V1) of shape information, a second V vector (V2) of image information, a third V vector (V3) of hemodynamic information, and a fourth V vector (V4) of position information using a third weight matrix.

[0086] The attention module (520) can determine attention scores (ATSC1~ATSC4) based on the first to fourth Q vectors (Q1~Q4) and the first to fourth K vectors (K1~K4). The attention scores (ATSC1~ATSC4) are implemented as vectors or matrices and may include a first attention score (ATSC1) of shape information, a second attention score (ATSC2) of image information, a third attention score (ATSC3) of hemodynamic information, and a fourth attention score (ATSC4) of position information. The attention module (520) can determine the attention scores (ATSC1~ATSC4) by performing an inner product operation on the first to fourth Q vectors (Q1~Q4) and the first to fourth K vectors (K1~K4), and dividing the inner product operation by a scaling value. The scaling value is expressed as the square root of the dimension of the K vector and can be used for stability. The attention module (520) can prioritize matching the dimensions of the first to fourth K vectors (K1 to K4) if their dimensions do not match.

[0087] The weighted adder (530) can generate combined data (COMB_DATA) based on the first to fourth attention scores (ATSC1~ATSC4) and the first to fourth V vectors (V1~V4). For example, the weighted adder (530) can generate attention weights using a softmax function to convert the first to fourth attention scores (ATSC1~ATSC4) into probabilities. The weighted adder (530) can generate combined data (COMB_DATA) by applying the first to fourth V vectors (V1~V4) to the attention weights.

[0088] The combiner (500) can transmit combined data (COMB_DATA) to a prediction module (e.g., 400 in FIG. 4). The prediction module can generate a pressure loss value based on the combined data (COMB_DATA). A loss value of a loss function can be calculated based on the error between the pressure loss value output by the prediction module and the actual loss value. As an example, the Mean Square Error (MSE) can be used as the loss function. The loss value can be used to update the weights (parameters) of the first to fourth weight matrices. That is, the weights of the first to fourth weight matrices can be updated through a backpropagation algorithm to minimize the loss value. In a training session, an optimizer such as SGD or Adam may be used to optimize the model by adjusting hyperparameters such as the learning rate.

[0089] In this way, the weighted adder (530) can make the inference of the prediction module more sophisticated by generating combined data (COMB_DATA) that combines different types of data according to importance. That is, V vectors with high weights have a significant impact on the combined data (COMB_DATA) to emphasize information of important data types, while V vectors with low weights have less impact on the combined data (COMB_DATA) to reduce the influence of noise or less important information. In addition, by efficiently integrating information of various data types, complex information required for prediction can be expressed as a single vector.

[0090]

[0091] FIG. 6 is a flowchart of a method for generating FFR data according to one embodiment, FIG. 7 is an example of a blood vessel image according to one embodiment, and FIG. 8 is an example of a blood vessel graph according to one embodiment.

[0092] Referring to FIG. 6, a method for generating FFR data according to one embodiment may be performed by an electronic device (e.g., 200 in FIG. 1). The electronic device may acquire a blood vessel image (S510). The blood vessel image is one that is captured by a imaging device (e.g., 100 in FIG. 1) and transmitted to the electronic device, and the image (BVIG) of FIG. 7 may be illustrated as an example of a blood vessel image. The image (BVIG) may be one in which the imaging device captures the blood vessel of the first user while a contrast agent is injected into the blood vessel of the first user. Since the description in FIG. 1 applies equally to the blood vessel image, redundant descriptions are omitted.

[0093] The electronic device can determine multiple nodes in a blood vessel image (S520). For example, the electronic device can determine multiple nodes by determining a target blood vessel in a blood vessel image and converting the target blood vessel into a first graph. The electronic device can model the target blood vessel in the blood vessel image in the form of a graph, and the graph (BVGP) of FIG. 8 can be shown as an example of the first graph. The graph (BVGP) may include multiple nodes (N1 to N16).

[0094] The electronic device can determine multiple node information of multiple nodes (S530). At this time, the first graph and multiple nodes can satisfy Equation 2.

[0095]

[0096] Here, G is a set representing the first graph, V is information about multiple nodes of the first graph, R is a set of all real matrices of size kxl, k is the total number of multiple nodes included in the first graph, l is the dimension of the information possessed by the nodes, and C may be the first edge information.

[0097] In addition, each node of the first graph can satisfy Equation 3.

[0098]

[0099] Here, i is the node index, which is an integer greater than 0 and less than or equal to k, and V i is the i-th node among a plurality of nodes of the first graph, and I is node information, which may include at least one of shape information, image information, hemodynamic information, and location information. That is, V i It may include l node information, and the node information may be applied in the same way as described with reference to FIG. 3.

[0100] The electronic device can determine multiple coefficients from multiple node information using an artificial neural network (S540). In one embodiment, the artificial neural network can generate a second graph from a first graph of Equation 2 and multiple nodes. The second graph output by the artificial neural network can satisfy Equation 4.

[0101]

[0102] Here, G is a set representing the first graph, θ is the weight (or parameter) of the artificial neural network, j is the index of the blood vessel model, and F(G,θ) j is an artificial neural network that transforms a first graph (G) using weights (θ), Y is information of multiple nodes of a second graph, R is a set of all real matrices of size kxl, k is the total number of multiple nodes included in the first graph, m is the total number of multiple blood vessel models, and C' may be second edge information. According to an embodiment, the first edge information and the second edge information may be the same or different. For example, if pressure loss in a side branch affects the main branch, the edge information may be modified so that the first edge information and the second edge information become different.

[0103] In addition, each node of the second graph can satisfy Equation 5.

[0104]

[0105] Here, i is the node index, which is an integer greater than 0 and less than or equal to k, and Y i is the i-th node among multiple nodes of the second graph, and w i1 inside w im can be a coefficient applied to m blood vessel models. That is, each node of the second graph can contain m coefficients applied to m blood vessel models.

[0106] An artificial neural network may be trained to output prediction coefficients when receiving a first graph (including multiple nodes and information about multiple nodes) as input. The artificial neural network may be trained to minimize the difference between the predicted loss value calculated based on the prediction coefficients and multiple vascular models and the actual loss value.

[0107] The electronic device can generate a plurality of vascular models based on a vascular image (S550). The plurality of vascular models may include at least one of a friction model, an expansion model, a contraction model, a vortex model, a separation model, and a branching model.

[0108] The electronic device can calculate a pressure loss value based on a plurality of blood vessel models and a plurality of coefficients (S560). The electronic device can calculate a pressure loss value of a node based on a first loss value in a first model among the plurality of blood vessel models, a first coefficient among the plurality of coefficients, a second loss value in a second model among the plurality of blood vessel models, and a second coefficient among the plurality of coefficients. For example, the electronic device can calculate a pressure loss value using Equation 6.

[0109]

[0110] Here, ΔP i is the pressure loss value of the i-th node among multiple nodes, j is the index of the vascular model, m is the total number of multiple vascular models, and wij is a coefficient applied to the j-th vascular model, and E j can be the loss value output by the j-th blood vessel model.

[0111] According to an embodiment, the electronic device can calculate the pressure loss value using Equation 1.

[0112] The electronic device can generate FFR data based on pressure loss values ​​(S570). In one embodiment, the electronic device can calculate multiple pressure loss values ​​for multiple nodes. For example, the electronic device can calculate a first pressure loss value of a first node based on multiple first coefficients and multiple vascular models corresponding to a first node among multiple nodes. The electronic device can calculate a second pressure loss value of a second node based on multiple second coefficients and multiple vascular models corresponding to a second node among multiple nodes.

[0113] The electronic device can generate FFR data based on multiple pressure loss values ​​and proximal pressure. Proximal pressure may refer to the average blood pressure measured prior to (or above) the lesion site of the blood vessel (e.g., stenosis site).

[0114] In one embodiment, the electronic device can generate FFR data using Equation 7.

[0115]

[0116] Here, FFR is the FFR data calculated from the target vessel, i is the node index, k is the total number of multiple nodes, j is the index of the vessel model, m is the total number of multiple vessel models, and w ij is a coefficient applied to the j-th vascular model, and P j is the pressure loss value output by the j-th blood vessel model, and P a It may be proximal pressure.

[0117] In some embodiments, the electronic device may generate FFR data from a plurality of pressure loss values ​​using a machine learning model. The machine learning model includes parameters applied to the plurality of pressure loss values ​​and can output FFR data based on the parameters.

[0118]

[0119] FIG. 9 is a diagram illustrating an inference session of a prediction model according to one embodiment.

[0120] Referring to FIG. 9, a prediction model (800) according to one embodiment may be configured to output a pressure loss value (PRS_LOSS) when receiving node data (or node information). Node data may be obtained from blood vessel images (BV IMAGES). According to an embodiment, the prediction model (800) may be implemented by further obtaining clinical information along with the node information.

[0121] The prediction model (800) may include an artificial neural network and a plurality of vascular models. The prediction model (800) may output a pressure loss value (PRS_LOSS) based on a plurality of coefficients output by the artificial neural network and a plurality of loss values ​​output by the plurality of vascular models.

[0122] The artificial neural network (800) may be trained using the learning method described with reference to the embodiments. The artificial neural network (800) may output a pressure loss value (PRS_LOSS) from the node data using weights (parameters) obtained through training. The pressure loss value (PRS_LOSS) output by the artificial neural network (800) may be used to determine FFR data. For example, the electronic device may calculate the pressure loss values ​​of the nodes of the target blood vessel in the blood vessel image and determine FFR data based on the pressure loss values. In this way, by using a blood vessel model and an artificial neural network, the electronic device can achieve high FFR value prediction performance while using relatively little data.

[0123]

[0124] FIG. 10 is a diagram illustrating a learning session of an artificial neural network according to one embodiment.

[0125] Referring to FIG. 10, a learning session of an artificial neural network (1010) according to one embodiment is described. Through the learning session, the artificial neural network (1010) may have weights (or parameters) for outputting parameters (PRMS) and coefficients (COFS) for calculating appropriate and precise FFR values. That is, the artificial neural network (1010) may be trained to output multiple parameters (PRMS) and multiple coefficients (COFS) applied to multiple physical models (1020) from a vascular model (or geometric features, node information, or node data). The multiple parameters (PRMS) and multiple coefficients (COFS) may be used by the FFR module (1030) to calculate an FFR value (P_FFR).

[0126] A vascular model can be generated by an electronic device from contrast images (e.g., vascular images). For example, the electronic device can construct a three-dimensional vascular model from multiple contrast images or generate a two-dimensional vascular model from a single contrast image.

[0127] The electronic device can determine a plurality of physical models (1020) corresponding to a blood vessel model. The plurality of physical models (1020) may include at least one of a friction model, an expansion model, a contraction model, a vortex model, a dissection model, and a branching model. Each of the blood vessel model and the plurality of physical models (1020) may include a plurality of nodes. For example, the electronic device may determine a target blood vessel in a blood vessel image and generate a plurality of physical models based on the target blood vessel.

[0128] In one embodiment, the electronic device can determine a plurality of nodes based on at least one of a branching region, a line region, and a lesion region in a target blood vessel. The electronic device can determine a plurality of nodes based on at least one of node length, cross-sectional area, and curvature in the line region.

[0129] The electronic device can extract geometric features of the blood vessel based on the contrast image. The geometric features of the blood vessel may include coordinate values ​​of each point along the blood vessel center point, the blood vessel diameter and cross-sectional area of ​​each point, the tangent vector or curvature of each point, etc., but the embodiments are not necessarily limited thereto, and the electronic device can generate various geometric features based on a blood vessel model. The electronic device can use the geometric features as input to an artificial neural network (1010).

[0130] An artificial neural network (1010) may be configured to output a plurality of parameters involved in each physical model (1020) for calculating an FFR value (P_FFR). In one embodiment, the plurality of parameters may determine a plurality of first parameters applied to a plurality of physical models and a plurality of second parameters which are coefficients applied to a plurality of loss values. An electronic device may calculate a loss value, a pressure loss value, etc., using at least one of the plurality of first parameters and the plurality of second parameters. That is, the plurality of first parameters and the plurality of second parameters may be used individually or together. In this case, the plurality of first parameters may be parameters (PRMS) (or physical parameters), and the plurality of second parameters may indicate coefficients (COFS) between physical models (1020). That is, the parameters (PRMS) may be applied to the physical model (1020), and the coefficients (COFS) may be applied to the pressure loss (PLOS).

[0131] In one embodiment, the electronic device may acquire node information based on a blood vessel image and transmit it to an artificial neural network (1010). The artificial neural network (1010) may generate a plurality of parameters (PRMS) and a plurality of coefficients (COFS) from the node information. The node information may include at least one of shape information, image information, and hemodynamic information.

[0132] In one embodiment, the electronic device may further acquire clinical information corresponding to a blood vessel image. The electronic device may transmit the clinical information to an artificial neural network (1010). At this time, the artificial neural network (1010) may generate a plurality of parameters (PRMS) and a plurality of coefficients (COFS) from the node information and the clinical information.

[0133] The artificial neural network (1010) can transmit parameters (PRMS) to the physical model (1020) and coefficients (COFS) to the FFR module (1030). The physical model (1020) may include multiple models.

[0134] A physical model (1020) can calculate a pressure loss (PLOS) at one of a plurality of nodes based on parameters (PRMS). In one embodiment, the plurality of physical models (1020) can calculate a pressure loss (PLOS) at one node based on a first loss value in a first model among the plurality of physical models (1020), a first parameter among the plurality of parameters (PRMS), a second loss value in a second model among the plurality of physical models (1020), and a second parameter among the plurality of parameters (PRMS).

[0135] For example, the physical model (1020) can calculate the pressure loss (PLOS) using mathematical formula 8.

[0136]

[0137] Here, f m is the pressure loss predicted by the m-th physical model, and k mis a parameter corresponding to the m-th physical model, ρ is the density of blood, A is a coefficient related to the cross-sectional area of ​​the blood vessel, V is the blood flow velocity, and n may be an exponent applied to the blood flow velocity.

[0138] The physical model (1020) can transmit the pressure loss (PLOS) to the FFR module (1030). The FFR module (1030) can calculate the FFR value (P_FFR), which is the pressure loss value at one node, based on the pressure loss (PLOS) and the coefficient (COFS). For example, the FFR module (1030) can calculate the FFR value (P_FFR) by weighting the pressure loss (PLOS) using Equation 9.

[0139]

[0140] Here, i is the index within the blood vessel, and ΔP i is the pressure loss at the i-th point (or time point), and f j is the pressure loss output by the j-th physical model, and c ij is the weighting coefficient in which the j-th physical model contributes to the pressure loss at the i-th point, j is the index of the physical model, and k may be the total number of physical models.

[0141] In one embodiment, the electronic device may generate a friction model as a first model, an expansion model as a second model, and a contraction model, a vortex model, a separation model, and a branching model. The electronic device may use an artificial neural network (1010) to obtain a plurality of parameters (PRMS) applied to a plurality of physical models (1020). The electronic device may determine a first parameter applied to the friction model, a second parameter applied to the expansion model, a third parameter applied to the contraction model, a fourth parameter applied to the vortex model, a fifth parameter applied to the separation model, and a sixth parameter applied to the branching model.

[0142] Additionally, the electronic device can use an artificial neural network (1010) to obtain multiple coefficients (COFS) to be applied to the pressure loss (PLOS) output by multiple physical models (1020). The electronic device can determine a first coefficient applied to a first loss value, a second coefficient applied to a second loss value, a third coefficient applied to a third loss value in the shrinkage model, a fourth coefficient applied to a fourth loss value in the peeling model, a fifth coefficient applied to a fifth loss value in the branching model, and a sixth coefficient applied to a sixth loss value in the branching model.

[0143] The electronic device can calculate the FFR value (P_FFR), which is the pressure loss value at one node, based on the first loss value, the first coefficient, the second loss value, the second coefficient, the third loss value, the third coefficient, the fourth loss value, the fourth coefficient, the fifth loss value, the fifth coefficient, the sixth loss value, and the sixth coefficient through the FFR module (1030).

[0144] The electronic device can calculate the loss based on the FFR value (P_FFR) and the correct answer data (W_FFR). The electronic device can obtain the correct answer data (W_FFR), which is the actual loss value, using an invasive method involving a wire, but the embodiments are not necessarily limited thereto.

[0145] In the embodiments, the electronic device may enable the training of the artificial neural network (1010) by using a single-node measured FFR value as ground truth data (W_FFR). The electronic device may calculate the predicted losses of multiple nodes using multiple coefficients (COFS) and multiple parameters (PRMS) output by the artificial neural network (1010). The electronic device may calculate the FFR value (P_FFR) by accumulating the predicted losses through the FFR module (1030) and calculate the loss by comparing it with the ground truth data (W_FFR) of the corresponding node. That is, the ground truth data (W_FFR) used for training may be sufficient with only the FFR value measured at a single node, for example, the distal node of the blood vessel. Therefore, in clinical environments where it is difficult to obtain continuous full-back data because FFR measurement is performed at only one node, the artificial neural network (1010) may be trained using only the measured FFR value of a single node.

[0146] According to an embodiment, when continuous pullback data is obtained and FFR measurements exist at multiple nodes, the electronic device may further define a loss for each node in the same manner and improve learning performance by summing or averaging them.

[0147] In one embodiment, the electronic device acquires FFR measurement data and can obtain an actual loss value at one node from the FFR measurement data. For example, the electronic device can obtain an actual loss value by calculating the cumulative pressure loss up to one node.

[0148] The electronic device may calculate the loss using a loss function. In one embodiment, the electronic device may use a loss function that represents the similarity between a predicted FFR value (P_FFR) and ground truth data (W_FFR), which is the actual measured pressure loss. P_FFR and W_FFR may be FFR values ​​for each time point of the blood vessel or FFR values ​​at a specific location.

[0149] In another embodiment, the electronic device may use a loss function that causes parameters involved in a specific physical model to follow specific physical equations. For example, the electronic device may set the flow velocity to satisfy physical equations such as the conservation of mass (continuity equation) or the conservation of momentum (Navier-Stokes equation).

[0150] The artificial neural network (1010) can be trained based on loss through backpropagation. The artificial neural network (1010) trained through such a training session has weights as parameters, and when a vascular model or geometric feature is input, the artificial neural network (1010) can output parameters (PRMS) and coefficients (COFS) based on the weights. Accordingly, by using the optimal coefficients (COFS) output by the artificial neural network (1010) and the pressure loss (PLOS) output by the physical model (1020), it becomes possible to predict an accurate FFR value (P_FFR) from the vascular model.

[0151] An artificial neural network that has completed a training session can be used in an inference session. For example, an electronic device can acquire a blood vessel image, generate a blood vessel model based on the blood vessel image, and determine a plurality of physical models (1020) corresponding to the blood vessel model.

[0152] The electronic device can output multiple parameters based on a vascular model using a trained artificial neural network. For example, the electronic device can determine multiple parameters (PRMS) as multiple first parameters and multiple coefficients (COFS) as multiple second parameters. The electronic device can calculate multiple pressure losses (PLOS) based on multiple physical models (1020) and multiple parameters (PRMS).

[0153] As described in the learning session, the artificial neural network (1010) may be trained to minimize the difference between the actual loss value and the predicted loss value calculated based on the multiple prediction coefficients and multiple prediction parameters and the multiple physical models (1020), with the learning blood vessel model as input data and the multiple prediction coefficients and multiple prediction parameters as output data.

[0154] The electronic device can generate FFR data based on a plurality of pressure losses (PLOS) and a plurality of coefficients (COFS). In one embodiment, the electronic device can calculate a plurality of pressure losses (PLOS) for a plurality of nodes and generate FFR data based on the plurality of pressure losses (PLOS) and proximal pressure.

[0155] For example, the electronic device may calculate a first pressure loss value of a first node based on a plurality of first parameters and a plurality of physical models (1020) corresponding to a first node among a plurality of nodes, and may calculate a second pressure loss value of a second node based on a plurality of second parameters and a plurality of physical models (1020) corresponding to a second node among a plurality of nodes. The electronic device may generate FFR data based on the first and second pressure loss values.

[0156]

[0157] It is obvious that each step or operation of the method according to the embodiments of the present disclosure may be performed by a computer comprising one or more processors in accordance with the execution of a computer program stored in a computer-readable recording medium.

[0158] The computer-executable instructions stored on the aforementioned recording medium can be implemented through a computer program programmed to perform each corresponding step, and such a computer program can be stored on a computer-readable recording medium and executed by a processor. The computer-readable recording medium may be a non-transitory readable medium. In this case, a non-transitory readable medium refers to a medium that stores data semi-permanently and is readable by a device, rather than a medium that stores data for a short moment, such as a register, cache, or memory. Specifically, programs for performing the various methods described above may be provided by being stored on a non-transitory readable medium, such as semiconductor memory devices including erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), and flash memory devices; magnetic disks such as internal hard disks and removable disks; optical-magnetic disks; and non-volatile memory including CD-ROMs and DVD-ROMs.

[0159] Methods according to the various examples disclosed in this document may be provided by being included in a computer program product. The computer program product may be distributed in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)) or online through an application store (e.g., Play Store™). In the case of online distribution, at least a portion of the computer program product may be temporarily stored or temporarily created on a storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0160] As explained above, a person skilled in the art to which this disclosure pertains will understand that this disclosure may be implemented in other specific forms without altering its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of this disclosure is defined by the claims set forth below rather than by the detailed description, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts should be interpreted as being included within the scope of this disclosure.

[0161] The features and advantages described herein are not all included, and in particular, many additional features and advantages will become apparent to those skilled in the art by considering the drawings, the specification, and the claims. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes and may not be chosen to describe or limit the subject matter of this disclosure.

[0162] The foregoing description of the embodiments of the present disclosure is provided for illustrative purposes only. It is not intended to limit the present disclosure to the exact form disclosed or to make it incomplete. Those skilled in the art will understand that many modifications and variations are possible in light of the foregoing disclosure.

[0163] Therefore, the scope of the present disclosure is not limited by the detailed description but by any of the claims of the application based thereon. Accordingly, the disclosure of embodiments of the present disclosure is illustrative and does not limit the scope of the present disclosure as set forth in the following claims.

Claims

1. A step of generating a vascular model based on a vascular image; A step of determining a plurality of physical models corresponding to the above-mentioned blood vessel model - each of the plurality of physical models includes a plurality of nodes -; A step of determining a plurality of parameters based on the above-mentioned blood vessel model using an artificial neural network; A step of calculating a pressure loss value at one of the plurality of nodes based on the plurality of physical models and the plurality of parameters; A step of obtaining the actual loss value at the above-mentioned node; and Step of training the artificial neural network based on the pressure loss value and the actual loss value A prediction model training method including 2. In Paragraph 1, The step of determining the above plurality of physical models is, A step of determining a target blood vessel in the above blood vessel image; and A step of generating a plurality of physical models based on the above-mentioned target blood vessels A prediction model training method including 3. In Paragraph 2, The step of determining a plurality of physical models based on the above-mentioned target blood vessels is, A step of determining the plurality of nodes based on at least one of a bifurcation region, a line region, and a lesion region in the above-mentioned target blood vessel A prediction model training method including 4. In Paragraph 3, The step of determining the above plurality of nodes is, A step of determining the plurality of nodes based on at least one of node length, cross-sectional area, and curvature in the above line area A prediction model training method including 5. In Paragraph 1, The step of determining the above plurality of parameters is, A step of obtaining node information based on the above-mentioned blood vessel image; and The artificial neural network generates the plurality of parameters from the node information. Includes, The above node information is, Comprising at least one of shape information, image information, and hemodynamic information, Predictive model training method.

6. In Paragraph 5, Step of further acquiring clinical information corresponding to the above blood vessel image Includes more, The step of generating the above plurality of parameters is, The artificial neural network generates the plurality of parameters from the node information and the clinical information. A prediction model training method including 7. In Paragraph 1, The step of obtaining the actual loss value above is, A step of acquiring FFR (Fractional Flow Reverse) measurement data; and Step of obtaining the actual loss value at the one node from the above FFR measurement data A prediction model training method including 8. In Paragraph 1, The step of determining the above plurality of parameters is, A step of determining a plurality of first parameters applied to the plurality of physical models and a plurality of second parameters which are coefficients applied to the plurality of loss values. Includes, The step of calculating the above pressure loss value is, A step of calculating a plurality of loss values ​​at one node among the plurality of nodes based on the plurality of physical models and the plurality of first parameters; and A step of calculating a pressure loss value at the one node based on the plurality of loss values ​​and the plurality of second parameters. A prediction model training method including 9. In Paragraph 8, The step of calculating the above plurality of loss values ​​is, A step of calculating the pressure loss value of the one node based on a first loss value in a first model among the plurality of physical models, a first physical parameter among the plurality of first parameters, a second loss value in a second model among the plurality of physical models, and a second physical parameter among the plurality of first parameters. A prediction model training method including 10. In Paragraph 9, The step of determining the above plurality of physical models is, A step of generating at least one of a friction model, an expansion model, a contraction model, a vortex model, a separation model, and a branching model. A prediction model training method including 11. In Paragraph 10, The step of generating at least one of the above friction model, expansion model, contraction model, vortex model, separation model, and branching model is, Step of generating a friction model as the first model, an expansion model as the second model, and a contraction model, a vortex model, a separation model, and a branching model. Includes, The step of determining a plurality of first parameters applied to the plurality of physical models and a plurality of second parameters, which are coefficients applied to the plurality of loss values, is A step of determining, as the plurality of first parameters, a first physical parameter applied to the friction model, a second physical parameter applied to the expansion model, a third physical parameter applied to the contraction model, a fourth physical parameter applied to the vortex model, a fifth physical parameter applied to the separation model, and a sixth physical parameter applied to the branching model; and A step of determining, as a plurality of second parameters, a first coefficient applied to the first loss value, a second coefficient applied to the second loss value, a third coefficient applied to the third loss value in the shrinkage model, a fourth coefficient applied to the fourth loss value in the peeling model, a fifth coefficient applied to the fifth loss value in the branching model, and a sixth coefficient applied to the sixth loss value in the branching model. Includes, The step of calculating the above pressure loss value is, A step of calculating the pressure loss value based on the first loss value, the first coefficient, the second loss value, the second coefficient, the third loss value, the third coefficient, the fourth loss value, the fourth coefficient, the fifth loss value, the fifth coefficient, the sixth loss value, and the sixth coefficient. including, Predictive model training method.

12. Step of acquiring a blood vessel image; A step of generating a blood vessel model based on the above blood vessel image; A step of determining a plurality of physical models corresponding to the above-mentioned blood vessel model; A step of determining a plurality of parameters based on the above-mentioned blood vessel model using an artificial neural network; A step of generating FFR data based on the plurality of physical models and the plurality of parameters. Includes, The above artificial neural network is, A training vascular model is used as input data, and a plurality of prediction parameters corresponding to the training vascular model are used as output data; the training is performed in a direction that minimizes the difference between the predicted loss value and the actual loss value calculated based on the plurality of prediction parameters and the plurality of physical models. FFR Data Generation Method 13. In Paragraph 12, The step of determining the above plurality of parameters is, A step of determining a plurality of first parameters applied to the plurality of physical models and a plurality of second parameters which are coefficients applied to the plurality of loss values. Includes, The step of generating the above FFR data is, A step of calculating a plurality of pressure loss values ​​based on the plurality of physical models and the plurality of first parameters; and A step of generating FFR data based on the plurality of pressure loss values ​​and the plurality of second parameters. A method for generating FFR data including 14. In Paragraph 12, The step of generating the above FFR data is, A step of calculating a plurality of pressure loss values ​​for the plurality of nodes; and A step of generating the FFR data based on the plurality of pressure loss values ​​and proximal pressure. A method for generating FFR data including 15. In Paragraph 14, The step of calculating multiple pressure loss values ​​for the above multiple nodes is, A step of calculating a first pressure loss value of a first node based on a plurality of first physical parameters corresponding to a first node among the plurality of nodes and the plurality of physical models; and A step of calculating a second pressure loss value of a second node based on a plurality of second physical parameters corresponding to a second node among the plurality of nodes and the plurality of physical models. A method for generating FFR data including 16. A processor and a memory connected to the processor, and The above memory is configured to store a program, and The above processor is configured to execute the above program, and When the above program is executed, the steps of the method of any one of claims 1 to 15 are implemented, Electronic device.

Images