Hemodynamic analysis of blood vessels using recurrent neural networks
Through the AI system and LSTM network, predicting hemodynamic parameters based on the vascular shape model and flow distribution, solving the problem of time-consuming and expensive CFD simulation, achieving rapid and accurate prediction of hemodynamic parameters, supporting clinical decision-making.
Patent Information
- Application Number
- CN202010517600.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2019-06-11
- Filing Date
- 2020-06-09
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2040-11-08
AI Technical Summary
Existing computational fluid dynamics (CFD) simulations are time-consuming and expensive to calculate hemodynamic parameters, difficult to integrate in clinical practice, and lack appropriate machine learning methods to quickly and accurately predict hemodynamic parameters.
Using artificial intelligence (AI) systems, especially long and short-term memory (LSTM) recurrent neural networks, predict hemodynamic parameters based on vascular shape model and flow distribution, and reduce the amount of input data through iterative prediction and using hidden states to achieve fast and accurate Hemodynamic parameter prediction.
It significantly reduces the calculation time and cost of hemodynamic parameter prediction, provides more efficient and accurate prediction of hemodynamic parameter, and supports clinical decision-making.
Smart Images

Figure CN112071419B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and an artificial intelligence (AI) system for predicting hemodynamic parameters of a target vessel and to a computer-implemented method for training an AI unit comprised by the AI system. Background Art
[0002] Coarctation of the aorta and aortic valve disease are the most common congenital heart defects. Treatment decision making for these diseases is a complex process and is highly dependent on the patient's condition. Treatment decisions can be significantly improved based on patient-specific hemodynamic parameters of the patient's aorta. The current clinical standard for assessing hemodynamic parameters is to perform a catheterization technique. However, catheterization techniques are an invasive procedure that poses risks to the patient (e.g., radiation, cost, complications). Non-invasive methods using different computational fluid dynamics (CFD) methods are being widely validated. These methods are increasingly being considered to replace invasive methods because they have lower risks and can provide more comprehensive and relevant information. However, these CFD methods are computationally expensive, which is a limiting factor for implementing such tools in the clinic.
[0003] To improve treatment decision making, hemodynamic analysis using computational fluid dynamics (CFD) simulations can be performed. CFD simulations calculate patient-specific hemodynamic parameters that can be used to better assess the patient's condition and optimize clinical decisions. However, CFD simulations are computationally demanding and can take up to a day to complete per scan. This prevents the integration of hemodynamic analysis based on CFD simulations in clinical practice.
[0004] Therefore, there is a need to provide an alternative way or method to more efficiently calculate hemodynamic parameters without the need for complex CFD simulations.
[0005] EP 3 270 308 A1 discloses a method for providing auxiliary parameters such as hemodynamic parameters inside a blood vessel to a clinical decision support (CDS) system. A shape dataset of the blood vessel is extracted from an image dataset and provided as an input dataset. A subsystem trained by machine learning (deep learning mechanism) is used to approximate the auxiliary parameters based on the input dataset.
[0006] The approximate auxiliary parameters or more precisely the hemodynamic parameters inside the blood vessels must be accurately determined, otherwise it may lead to misdiagnosis by the CDS system or clinicians based on the approximate hemodynamic parameters.
[0007] Furthermore, no appropriate machine learning approach has been proposed for this task, probably because big data is required to develop algorithms, which is not easy to obtain. Summary of the invention
[0008] The object of the present invention is to overcome or at least alleviate these problems by providing the following items: a method for predicting hemodynamic parameters of a target vessel according to the present application; and an AI system for predicting hemodynamic parameters of a target vessel; and a computer-implemented method for training an AI unit included in the AI system according to the present application.
[0009] According to a first aspect of the present invention, a method for predicting a hemodynamic parameter p of a target blood vessel is provided. k The method comprises the following steps:
[0010] -Receive a vessel shape model of the target vessel.
[0011] - receiving a corresponding flow profile of the target vessel.
[0012] - predicting, by an artificial intelligence (AI) unit, at least one hemodynamic parameter p based on the received vessel shape model and the received flow distribution k .
[0013] According to a second aspect of the present invention, a method for predicting a hemodynamic parameter (p) of a target blood vessel is provided. k ) is arranged and configured to perform the method according to the first aspect of the present invention. The AI system comprises a first interface, a second interface and an AI unit. The first interface is arranged and configured to receive a vessel shape model of a target vessel. The second interface is arranged and configured to receive a corresponding flow distribution of the target vessel. The AI unit is communicatively connected to the first interface and the second interface. The AI unit is arranged and configured to predict at least one hemodynamic parameter (p) based on the received vessel shape model and the received flow distribution. k ).
[0014] According to a third aspect of the present invention, a computer program comprises instructions which, when the program is executed by a computer, cause the computer to perform the steps of the method according to the first aspect of the present invention.
[0015] According to a fourth aspect of the present invention, a computer readable medium has stored thereon a computer program according to the third aspect of the present invention.
[0016] According to a fifth aspect of the present invention, a data processing system comprises means for performing the steps of the method according to the first aspect of the present invention.
[0017] The present invention provides a novel method and algorithm for hemodynamic parameter prediction using an AI unit and in particular using a recurrent neural network (RNN) such as a long short-term memory (LSTM) network, respectively.
[0018] At least one hemodynamic parameter p of the target vessel k It can be at least one of the following items: systolic blood pressure, diastolic blood pressure, arterial blood pressure, mean arterial pressure, systolic blood pressure change, pulse pressure change, stroke volume change, right atrial pressure, right ventricular pressure, pulmonary artery pressure, mean pulmonary artery pressure, pulmonary artery wedge pressure, left arterial pressure, cardiac output, cardiac index, stroke volume, systemic vascular resistance, systemic vascular resistance index, pulmonary vascular resistance, pulmonary vascular resistance index, left ventricular stroke work, left ventricular stroke work index, right ventricular stroke work, right ventricular stroke work index, coronary perfusion pressure, right ventricular end-diastolic volume, right ventricular end-systolic volume, right ventricular ejection fraction, etc.
[0019] The target vessel may be any vessel or a portion or part of such a vessel of the human or animal body being the subject of diagnosis or planned intervention. In particular, the target vessel may be the aorta or more precisely a portion of the aorta of the patient.
[0020] The vessel shape model of the target vessel is a first input for predicting at least one hemodynamic parameter of the target vessel. The vessel shape model of the target vessel can be a data set representing any kind of polygonal mesh. For example, the data set includes a set of vertices, triangles and / or polygons forming a two-dimensional (2D) or three-dimensional (3D) mesh. The vessel shape model or data set can also include a volume mesh formed by vertices, tetrahedrons, hexahedrons or other 3D elements. Preferably, the vessel shape model includes a segmentation of the surface of the 3D vessel, in particular the aorta, wherein the surface is obtained from an image data set. For example, the vessel shape model is extracted from an image data set showing the aorta of a patient. In addition, the vessel shape model can include a centerline extending from the entrance to the exit of the target vessel and a corresponding diameter of the target vessel, wherein the centerline runs through all center points along the target vessel.
[0021] The corresponding flow distribution of the target vessel is a second input for predicting at least one hemodynamic parameter of the target vessel. The corresponding flow distribution of the target vessel includes features of the blood flow in the target vessel. Preferably, the flow distribution includes at least three features and up to 100 features. Most preferably, the flow distribution includes 16 features. For example, the flow distribution may include blood flow velocity, features calculated based on the blood flow velocity such as secondary flow, boundary conditions such as input flow rate, output flow rate, etc. The flow distribution may include said features only for the inlet of the target vessel or more precisely for the vessel shape points at the inlet of the vessel shape model of the target vessel.
[0022] At least one hemodynamic parameter is predicted based on information and features included in the blood vessel shape model and the corresponding flow distribution of the target blood vessel (at the inlet). In particular, at least one hemodynamic parameter can be predicted at at least one blood vessel shape point of the blood vessel shape model based on three coordinates x, y, z of the at least one blood vessel shape point and the corresponding radius r of the target blood vessel and based on features of the flow distribution (e.g., blood flow velocity) at the inlet of the target blood vessel.
[0023] A vessel shape model of a target vessel is provided via a first interface of the AI system. A corresponding flow distribution of the target vessel is provided via a second interface of the AI system. The first input and the second input are forwarded to an AI unit of the AI system. At least one hemodynamic parameter is predicted by the AI unit. The AI unit, which may be an artificial neural network (ANN), is trained to predict the hemodynamic parameter based on the vessel shape model of the target vessel and the corresponding flow distribution (at the inlet).
[0024] Artificial neural network (ANN) is a system inspired by biological neural network that constitutes animal brain, especially computing system. ANN usually "learns" to perform tasks by considering (labeled) examples or training data without any task-specific rules for design. During the initial learning or training phase, ANN automatically generates recognition characteristics according to (labeled) training data. ANN includes a collection of connected nodes called artificial neurons, which roughly model neurons in biological brains. Each connection (synapse in biological brain) can transmit a signal from one node to another node. The node receiving the signal can process the signal and then send a signal to the subsequent neuron connected to the node. In the common ANN implementation, the signal at the connection between the nodes is a real number (e.g., 0, ..., 1), and the output of each artificial neuron is calculated by some nonlinear functions of the sum of its input (from other nodes). The connection between nodes is called "edge". The edges in ANN can each have a weight that is adjusted during the training of ANN. The weight increases or decreases the intensity of the signal at the corresponding edge. Nodes can each have a threshold so that they send a signal only if the aggregate signal exceeds that threshold. Typically, nodes are aggregated into layers. Different layers can perform different kinds of transformations on their inputs. The signal travels from the first or input layer to the last or output layer - possibly after traversing the layers multiple times.
[0025] In other words, an ANN is a network of simple elements (so-called nodes or artificial neurons) that receive input. Upon receiving the input, the node changes its internal state (activation) according to the input and produces an output according to the input and activation. The network is formed by connecting the outputs of certain nodes to the inputs of other nodes, thereby forming a directed weighted graph. The functions that calculate the activation of each node, as well as the weights, can be modified during the initial learning / training period, which is governed by a learning rule or paradigm.
[0026] A node that receives input from at least one predecessor neuron includes the following components: an activation, the state of the node, which depends on a discrete time parameter (optionally a threshold, which remains fixed unless changed by a learning / training function); an activation function (e.g., hyperbolic tangent function, sigmoid function, softmax function, rectifier function, etc.), which computes the new activation and net input at a given time; and an output function, which computes the output from the activation (usually, the output function is the identity function). An important property of an activation function is that it provides a smooth transition as the input value changes, i.e., a small change in the input produces a small change in the output.
[0027] The input node has no predecessor, but serves as the input interface of the entire ANN. Similarly, the output node has no successor, and therefore serves as the output interface of the entire ANN. The ANN includes edges / connections, each of which passes the output of a node (predecessor) to the input of another successor node (successor). In addition to the assigned weights, the edges can also have a bias term, which is added to the total weighted sum of the inputs to serve as a threshold for the mobile activation function. The propagation function calculates the input to the successor node (successor) based on the output of the previous node (predecessor) and may include a bias value.
[0028] Based on the forwarded vessel shape model of the target vessel and the flow distribution (at the inlet) as input, the AI unit predicts at least one hemodynamic parameter (at at least one vessel shape point).
[0029] Compared to existing methods using computational fluid dynamics (CFD) or common AI methods, the proposed method or rather the algorithm can be executed significantly faster. In particular, the additional information contained in the (inlet) flow profile, which is provided to the above-mentioned AI unit as an additional input in addition to the vessel shape model, facilitates a more accurate (in particular compared to common AI methods) and faster (in particular compared to methods using CFD) or even real-time prediction of at least one hemodynamic parameter. The predicted at least one hemodynamic parameter can help clinicians to perform complex treatment assessments for patients with, for example, aortic coarctation.
[0030] According to an improvement of the present invention, in the prediction step, at least one hemodynamic parameter is predicted for at least one blood vessel shape point of the received blood vessel shape model.
[0031] The vessel shape model includes at least one vessel shape point having a first coordinate in a first direction x, a second coordinate in a second direction y perpendicular to the first direction x, and a third coordinate in a third direction z perpendicular to the first direction x and the second direction y. The at least one vessel shape point of the vessel shape model may be located on a target vessel, or more precisely, on a centerline of the vessel shape model of the target vessel, and may include a radius r of the target vessel at a corresponding position of the vessel shape point along the centerline. In addition, the at least one vessel shape point may contain additional information, such as a change in the radius of the target vessel, a position along the centerline, etc.
[0032] Instead of predicting at least one hemodynamic parameter at each point of the target vessel, prediction is performed only for provided vessel shape points of the vessel shape model of the target vessel. For example, the vessel shape model may include a number of vessel shape points (e.g., 100 to 500) along a centerline, wherein each vessel shape point includes its coordinates x, y, z and a radius r of the target vessel at the corresponding vessel shape point. At least one hemodynamic parameter may be predicted by the AI unit at each given vessel shape point of the vessel shape model.
[0033] By predicting at least one hemodynamic parameter only at a given vessel shape point of the vessel shape model instead of at every node of the mesh of the vessel shape model, the computational effort can be significantly reduced. However, the information value is high such that a clinical decision (by a clinician or a CDS system) can be made based on the provided at least one hemodynamic parameter at a given vessel shape point.
[0034] According to a refinement of the present invention, the step of predicting is iteratively performed for each of at least two consecutive blood vessel shape points of the received blood vessel shape model.
[0035] At a first vessel shape point of the vessel shape model, at least one hemodynamic parameter is predicted by the AI unit based on the information comprised by the first vessel shape point as a first input and on the flow distribution (at the inlet of the target vessel) as a second input. At each subsequent vessel shape point after the first vessel shape point or in other words at the next vessel shape point downstream of the centerline, the second input to the AI unit is the internal information of the AI unit from the previous vessel shape point (particularly the hidden state of the layer) instead of the (inlet) flow distribution. Thus, each vessel shape point is iteratively stepped from the first vessel shape point to the last vessel shape point (along the centerline). For each vessel shape point, at least one hemodynamic parameter is predicted by the AI unit, wherein for the first vessel shape point the (inlet) flow distribution is used as the second input, and for any subsequent vessel shape point the internal information of the AI unit at the previous vessel shape point is used as the second input.
[0036] In particular, the AI unit may include a recurrent neural network, where internal information of the RNN, i.e., hidden states of layers of the RNN, are remembered and used in subsequent iterations.
[0037] Due to the iterative execution of the prediction step, the amount of input data that must be provided to the AI unit for predicting at least one hemodynamic parameter at different vessel shape points can be significantly reduced, in case the flow distribution is only required as a second input to the AI unit for a first vessel shape point.
[0038] According to an improvement of the present invention, the AI unit includes a recursive neural network RNN, and preferably includes a long short-term memory LSTM network.
[0039] RNN is a network with a loop in it, which allows information to persist. A layer of nodes in RNN receives values or more precisely hidden states at some inputs and outputs values or more precisely hidden states, while the corresponding loop allows information to be passed from one step to the next. LSTM network is a specific type of RNN that can learn long-term dependencies with the help of memory cells. All RNN and LSTM networks have the form of a chain of repeating modules of neural networks. In standard RNN, the repeating module or more precisely the layer of nodes has a simple structure (e.g., a single hyperbolic tangent (tanh) layer). The repeating module in the standard RNN contains a single layer. LSTM network also has a chain structure, but the repeating module has a different structure. Instead of having a single neural network layer, there are usually four interactive layers in the LSTM network, so that the memory unit can remember the cell state of each layer of the repeating module / node. LSTM network has the ability to remove or add information to the cell state in the memory cell via the gate of the memory cell. The gate is usually composed of a sigmoid layer and a point-by-point multiplication operation. The sigmoid layer outputs a number between 0 and 1 for weighting the output. LSTM networks typically have three of these gates to protect and control the cell state in the memory cell.
[0040] RNNs and especially LSTM networks improve efficiency because only one (inlet) flow distribution needs to be provided for the first vessel shape point of an iteration, while for subsequent vessel shape points, the corresponding remembered hidden states of previous iterations can be used to predict at least one hemodynamic parameter.
[0041] According to an improvement of the present invention, the step of predicting includes the following iterative steps:
[0042] - providing one vessel shape point of the received vessel shape model to the first input block of the AI unit at each iteration i.
[0043] - Providing the received flow distribution to the second input block of the AI unit at a first iteration i=1.
[0044] - based on the provided one vessel shape point and based on the provided flow distribution at the first iteration i=1 or at any iteration i>1 after the first iteration at least one previous hidden state h of the deductive block of the AI unit j,i-1 , generates at least one current hidden state h in the deduction block j,i .
[0045] - Based on at least one generated current hidden state h j,i The last generated current hidden state h in M,i , determining at least one hemodynamic parameter p in the output block of the AI unit k.
[0046] According to a further improvement of the present invention, the AI unit includes a first input block, a second input block, a deduction block and an output block. The first input block is communicatively connected to the first interface and is arranged and configured to obtain a provided vascular shape point of the received vascular shape model at each iteration i. The second input block is communicatively connected to the second interface and is arranged and configured to obtain a provided received flow distribution at the first iteration i=1. The deduction block is communicatively connected to the first input block and the second input block. The deduction block is arranged and configured to generate a vascular shape point based on the provided vascular shape point and based on the provided flow distribution at the first iteration i=1 or at least one previous hidden state h of the deduction block at any iteration i>1 after the first iteration. j,i-1 To generate at least one current hidden state h j,i The output block is communicatively connected to the deduction block and is arranged and configured to generate a current hidden state h based on at least one j,i The last generated current hidden state h in M,i To determine at least one hemodynamic parameter p k .
[0047] For each vessel shape point (1, ..., N) of the provided vessel shape model of the target vessel, at least one hemodynamic parameter p is predicted in an iteration i. k , where i=1,...,N.
[0048] The first input block of the AI unit may include at least one layer. The (one or more) layers of the first input block may be fully connected and preferably include 4 to 512 nodes, most preferably include 64 nodes. The second input block of the AI unit may include at least one layer. The (one or more) layers of the second input block may be fully connected and preferably include 4 to 512 nodes, most preferably include 64 nodes. The deductive block of the AI unit may include at least one deductive layer (M is the number of layers of the deductive block, M≥1). The (one or more) deductive layers of the deductive block may be fully connected and preferably include 4 to 512 nodes, most preferably include 64 nodes. The output block of the AI unit may include at least one layer, and preferably includes at least one cascade layer and an output layer. The layers of the output block may be fully connected and preferably include 4 to 512 (output) nodes, more preferably include 64 (output) nodes. Most preferably, at least one cascade layer includes 48 nodes, and the output layer includes as many output nodes as the hemodynamic parameters to be predicted by the AI unit.
[0049] For each vessel shape point of the provided vessel shape model of the target vessel, one iteration i has to be performed to predict each corresponding at least one hemodynamic parameter.
[0050] At the first input block, one vessel shape point is provided to the AI unit. The complete vessel shape model received at the first interface is stepped point by point so that at each iteration i, one vessel shape point is provided to the first input block. Starting from the first vessel shape point located closest to the inlet of the target vessel at the first iteration i=1, the vessel shape model is stepped until at the last iteration i=N, the last vessel shape point located closest to the outlet of the target vessel is provided to the first input block.
[0051] At the first iteration i=1 where the first vessel shape point (at the entrance of the target vessel) is provided to the first input block, the flow distribution received at the second interface is provided to the AI unit via the second input block. The flow distribution preferably includes information or features about the blood flow at the entrance of the target vessel.
[0052] The information contained in the blood vessel shape points provided to the first input block and the information contained in the (inlet) flow distribution provided to the second input block are converted into features by the corresponding input blocks and provided to the deduction block. The deduction block uses the features (transformed input information) from the first input block and the second input block to deduce at least one hidden state h j,i Based on the final hidden state h of the deduction block M,i , to predict at least one hemodynamic parameter at the provided vascular shape point. So far, the final hidden state h M,i is provided to the output block, which uses the last hidden state h M,i To determine at least one corresponding hemodynamic parameter at the provided blood vessel shape point.
[0053] In any iteration i>1 after the first iteration, no flow distribution is provided to the second input block. Instead of the flow distribution as the second input, in addition to a vessel shape point, the hidden state h of the deduction block at the previous iteration i-1 is provided j,i-1 As input to the deduction block.
[0054] The deduction block generates at least one hidden state h at each iteration i. j,i , where at least one of the hidden states is the last hidden state h M,i (M is the number of layers of the deduction block, M≥1) contains all the information of at least one hemodynamic parameter to be predicted, and then in the output block the information is concatenated and combined into at least one hemodynamic parameter.
[0055] For example, a vessel shape model of a target vessel (e.g., the aorta) including three vessel shape points (along the centerline) may be provided to a first interface and received at the first interface. In addition, an inlet flow distribution having characteristics of the blood flow at the inlet of the target vessel may be provided to a second interface and received at the second interface. The first vessel shape point may be located at the inlet of the target vessel. In order to predict at least one hemodynamic parameter at each vessel shape point, three iterations must be performed, one iteration for each vessel shape point. In the first iteration i=1, the first vessel shape point is provided to a first input block as a first input to the AI unit, and the inlet flow distribution is provided to a second input block as a second input to the AI unit. The two input blocks convert the information contained in the first vessel shape point and the inlet flow distribution into features of the deduction block, respectively. The deduction block deduces at least one hidden state h based on the features provided by the first input block and the second input block. 1,1 Finally, the output block is based on the (final) hidden state h provided by the deduction block. 1,1 The deduction block is based on the features or more precisely the transformed information of the second and last vessel shape points provided by the first input block and on the at least one hidden state h of the corresponding previous iteration. 1,1 and h 1,2 , to deduce at least one hidden state h at the second iteration i=2 1,2 and at least one hidden state h at the last iteration 1,3 The output blocks are respectively based on the (final) hidden state h provided by the deduction block. 2,1 and h 3,1 , to determine at least one hemodynamic parameter at the second blood vessel shape point and the last blood vessel shape point, respectively.
[0056] The two input blocks provide features to the deductive block based on the input data, and the output block determines at least one hemodynamic parameter according to the final hidden state of the deductive block in each iteration (i.e., for each vessel shape point), which facilitates particularly efficient prediction of the hemodynamic parameters of the target vessel.
[0057] According to an improvement of the present invention, the step of generating includes the following steps:
[0058] - based on a provided vessel shape point and based on a provided flow distribution at the first iteration i=1 or a provided first previous hidden state h of the first deductive layer of the AI unit at any iteration i>1 after the first iteration 1,i-1 , generating the first current hidden state h in the first deductive layer1,i .
[0059] - Optionally based on the current hidden state h of the previous deductive layer j-1,i and an initial state s of at least one further deductive layer based on the AI unit at the first iteration i=1 j,1 or at least one additional previous hidden state h provided by at least one additional deductive layer at any iteration i>1 after the first iteration j,i-1 , generating at least one additional current hidden state h in at least one additional deductive layer j,i .
[0060] - Based on the current hidden state h of the previous deductive layer M-1,i and the initial state s of the final deductive layer based on the AI unit M,1 or the last previous hidden state h provided by the last deductive layer M,i-1 , in the last deductive layer, the final current hidden state h is generated M,i .
[0061] According to a further improvement of the present invention, the AI unit comprises a first deductive layer, optionally at least one further deductive layer and a final deductive layer. The first deductive layer is communicatively connected to the first input block and the second input block. The first deductive layer is arranged and configured to provide a first previous hidden state h of the first deductive layer based on a provided vessel shape point and based on a provided flow distribution at a first iteration i=1 or any iteration i>1 after the first iteration. 1,i-1 To generate the first current hidden state h 1,i The optional at least one further deductive layer is communicatively connected to the first deductive layer. The optional at least one further deductive layer is arranged and configured to generate a current hidden state h of a previous deductive layer based on the current hidden state h of the previous deductive layer. j-1,i and an initial state s of at least one further deductive layer based on the first iteration i=1 j,1 or at least one additional previous hidden state h provided by at least one additional deductive layer at any iteration i>1 after the first iteration j,i-1 , to generate at least one additional current hidden state h j,i The last rendition layer is communicatively connected to the first rendition layer or the last of the at least one further rendition layer and is arranged and configured to generate a hidden state h based on the current hidden state h of the previous rendition layer. M-1,i and the initial state s of the last deductive layer based on the first iteration i=1 M,1 or the last previous hidden state h provided by the last deductive layer at any iteration i>1 after the first iteration M,i-1 , to generate the final current hidden state h M,i .
[0062] The interpretation block comprises at least one first interpretation layer and one final interpretation layer. Between the first interpretation layer and the final interpretation layer, the interpretation block may comprise further interpretation layers. Preferably, the interpretation block comprises a total of 3 to 10 interpretation layers (1 first layer, 1 to 8 further interpretation layers and 1 final interpretation layer, M=3, ..., 10).
[0063] A first one of the at least one additional rendition layer is communicatively connected to the first rendition layer, and a last one of the at least one additional rendition layer is communicatively connected to a last one of the at least one additional rendition layer. Each of the at least one additional rendition layer is communicatively connected to its adjacent rendition layer in succession.
[0064] Based on the current vessel shape point as a first input, the first input block provides information to the deductive block. Furthermore, the second input block provides information based on the (inlet) flow distribution at the first iteration i=1 or more precisely at the first vessel shape point of the provided vessel shape model of the target vessel to the deductive block. In the deductive block, the provided information is converted by the first deductive layer into at least one first hidden state h of the current iteration i. 1,i The first memory unit of the first deductive layer remembers the first current hidden state h of the first deductive layer 1,i In the subsequent iteration i+1, the first current hidden state h is used 1,i There may be further deductive layers communicatively connected continuously to the first deductive layer, which use the first current hidden state h 1,i To be used to deduce another current hidden state h j,i If there is no additional layer between the first layer and the last layer of the block (M=2), the last layer uses the first current hidden state h 1,i To get the final current hidden state h M,i , or more precisely, if there is an additional deductive layer between the first deductive layer and the last deductive layer of the deductive block (M>2), the last deductive layer uses the additional current hidden state h provided by the last additional deductive layer M-1,i The last one in the above example gets the final current hidden state h M,i .
[0065] Each deductive layer has a predefined initial state s at the beginning of the iteration. j,1 For each iteration i>1 after the first iteration i=1, the second input block does not receive the flow distribution as the second input of the deduction block. Instead, in addition to the current vessel shape point provided as the first input via the first input block, the deduction layer also uses its own corresponding previous hidden state h at the previous iteration i-1 j,i-1 As an additional input. The previous hidden state hj,i-1 is remembered by the corresponding memory unit of the corresponding deductive layer and replaces the corresponding initial state s after the first iteration i=1 j,1 At the first iteration, each memory cell of each deductive layer has a predetermined initial state s j,1 , which is used (except by the first deductive layer) to get the current hidden state h j,i These initial states of the memory cells of the deductive layer (except the first deductive layer) are j,1 can be seen as some kind of second input in later iterations i>1, such as the previous hidden state h j,i-1 .
[0066] Based on the final hidden state h of each iteration i M,i The output block determines at least one hemodynamic parameter p at the corresponding vascular shape point of the current iteration i k .
[0067] A deductive block having a first deductive layer and a last deductive layer and optionally further deductive layers between the first and last deductive layer provides a particularly precise and fast prediction of at least one hemodynamic parameter.
[0068] According to an improvement of the present invention, the vessel shape model includes a centerline along the main direction of the target vessel. In addition, each continuous vessel shape point includes continuous centerline points on the centerline and the radius of the target vessel at the corresponding centerline point.
[0069] The center line runs through the center of each cross section or slice of the target vessel. The radius r of the target vessel at each point along the center line is orthogonal to the center line. Each vessel shape point of the vessel shape model of the target vessel is located on the center line. Each vessel shape point may include its coordinates (x, y, z) in 3D Cartesian coordinates or a distance from the entrance along the center line. In addition, each vessel shape point includes the radius r of the target vessel at the corresponding vessel shape point.
[0070] The hemodynamic parameters at points along the centerline give the most relevant information about the condition of the target vessel. Thus, predicting at least one hemodynamic parameter only at vessel shape points along the centerline provides efficient determination of precise information about the condition of the target vessel that can be used by a clinician or a CDS system.
[0071] According to an improvement of the present invention, the blood vessel shape model includes a centerline along the main direction of the target blood vessel. In addition, each continuous blood vessel shape point includes a continuous centerline point on the centerline and a predetermined number k of surrounding points surrounding each centerline point in the form of a sphere, wherein the corresponding centerline point is located at the center of the corresponding sphere. The generating step includes the following steps:
[0072] - based on a provided blood vessel shape point including a corresponding centerline point and k surrounding points around the centerline point and based on a provided flow distribution at the first iteration i=1 or a provided first previous hidden state h of the first deductive layer of the AI unit at any iteration i>1 after the first iteration 1,i-1 、h 1,i-1 k , generating the first current hidden state h in the first deductive layer 1,i 、h 1,i k .
[0073] - Optionally based on the current hidden state h of the previous deductive layer j-1,i 、h j-1,i k and an initial state s of at least one further deductive layer based on the AI unit at the first iteration i=1 j,1 、s j,1 k or at least one additional previous hidden state h provided by at least one additional deductive layer 13.j at any iteration i>1 after the first iteration j,i-1 、h j,i-1 k , generating at least one additional set of current hidden states h in at least one additional deductive layer j,i 、h j,i k .
[0074] - Based on the current hidden state h of the previous deductive layer M-1,i 、h M-1,i k and the initial state s of the final deductive layer based on the AI unit M,1 、s M,1 k or the last previous hidden state h provided by the last deductive layer M,i-1 、h M,i-1 k , in the last deductive layer, the final current hidden state h is generated M,i 、h M,i k .
[0075] In the determined step, based on the last generated current hidden state h M,i 、h M,i k At least one hemodynamic parameter p is determined in the output block of the AI unit k .
[0076] According to a further improvement of the present invention, the first input block is arranged and configured to obtain a provided blood vessel shape point, the blood vessel shape point comprising a continuous centerline point along the centerline of the main direction of the target blood vessel and a predetermined number k of surrounding points surrounding each centerline point in the form of a sphere, wherein the corresponding centerline point is located at the center of the corresponding sphere at each iteration i. The first deductive layer is arranged and configured to generate a first previous hidden state h of the first deductive layer based on the provided blood vessel shape point comprising the corresponding centerline point and the k surrounding points surrounding the centerline point and based on the provided flow distribution at the first iteration i=1 or any iteration i>1 after the first iteration. 1,i-1 、h 1,i-1 k , to generate the first current hidden state h 1,i 、h 1,i k Optionally, at least one further deductive layer is arranged and configured to be based on the current hidden state h of the previous deductive layer. j-1,i 、h j-1,i k and an initial state s of at least one further deductive layer based on the first iteration i=1 j,1 、s j,1 k or at least one additional previous hidden state h provided by at least one additional deductive layer 13.j at any iteration i>1 after the first iteration j,i-1 、h j,i-1 k , to generate at least one additional set of current hidden states h j,i 、h j,i k The final deductive layer is arranged and configured to be based on the current hidden state h of the previous deductive layer. M-1,i 、h M-1,i k and the initial state s of the last deductive layer based on the first iteration i=1 M,1 、s M,1 k or any iteration i>1 after the first iteration, providing the last previous hidden state h of the last deductive layer M,i-1 、h M,i-1 k , to generate the final current hidden state h M,i 、h M,i k The output block is arranged and configured to be based on at least the last generated current hidden state h M,i 、h M,i k To determine at least one hemodynamic parameter p k .
[0077] The k surrounding points form a sphere around each centerline point along the centerline. The sphere has a predetermined radius c from each centerline point to the (spherical) surface of the sphere. Each surrounding point may have the same distance between itself and its neighboring surrounding points or more precisely the corresponding centerline point. Each centerline point is preferably surrounded by 50 to 100 surrounding points (k∈[50; 100]).
[0078] The first input block provides information to the deductive block based on the current vessel shape point including the current centerline point and k surrounding points around the current centerline point as a first input. In addition, the second input block provides information based on the (inlet) flow distribution to the deductive block at the first iteration i=1 or more precisely at the first vessel shape point of the provided vessel shape model of the target vessel including the current centerline point and k surrounding points around the current centerline point. In the deductive block, the provided information is converted into the first hidden state h of the current iteration i by the first deductive layer. 1,i 、h 1,i k The first memory unit of the first deductive layer remembers the first current hidden state h of the first deductive layer 1,i 、h 1,i k These first current hidden states h 1,i 、h 1,i k For the subsequent iteration i+1. There may be further deductive layers continuously communicatively connected to the first deductive layer, which use the first current hidden state h 1,i 、h 1,i k To be used to deduce another current hidden state h j,i 、h j,i k If there is no additional layer between the first layer and the last layer of the block (M=2), the last layer uses the first current hidden state h 1,i 、h 1,i k To get the final current hidden state h M,i 、h M,i k , or more precisely, if there is an additional deductive layer between the first deductive layer and the last deductive layer of the deductive block (M>2), the last deductive layer uses the additional current hidden state h provided by the last additional deductive layer M-1,i 、h M-1,i k The last one in the above example gets the final current hidden state h M,i 、h M,i k .
[0079] Each deductive layer has a predefined initial state s at the beginning of the iteration. j,1 、s j,1 k For each iteration i>1 after the first iteration i=1, the second input block does not receive the flow distribution as the second input of the deduction block. Instead, in addition to the current vessel shape point including the current centerline point and k surrounding points around the current centerline point provided as the first input via the first input block, the deduction layer also uses its own corresponding previous hidden state h at the previous iteration i-1 j,i-1 、h j,i-1 k As an additional input. The previous hidden state h j,i-1 、h j,i-1 k is remembered by the corresponding memory unit of the corresponding deductive layer and replaces the corresponding initial state s after the first iteration i=1 j,1 、s j,1 k At the first iteration, each memory cell of each deductive layer has a predetermined initial state s j,1 、s j,1 k , which is used (except by the first deductive layer) to get the current hidden state h j,i 、h j,i k These initial states of the memory cells of the deductive layer (except the first deductive layer) are j,1 、s j,1 k can be seen as some kind of second input, such as the previous hidden state h in the following iteration i>1 j,i-1 、h j,i-1 k .
[0080] Based on the final hidden state h of each iteration i M,i 、h M,i k The output block determines at least one hemodynamic parameter p at the corresponding vascular shape point of the current iteration i k .
[0081] A deductive block using a vessel shape point comprising a centerline point and k surrounding points, having a first deductive layer and a final deductive layer and optional additional deductive layers between the first and final deductive layers provides the most accurate prediction of at least one hemodynamic parameter along a target vessel and its branches.
[0082] According to an improvement of the present invention, the flow distribution includes an inlet blood flow velocity u in a direction x of the corresponding image data set, an inlet blood flow velocity v in a direction y of the image data set orthogonal to the direction x, and an inlet blood flow velocity w in a direction z of the image data set orthogonal to the directions x and y. The flow distribution also includes optional additional information and optional boundary conditions such as input flow rate, output flow rate, etc., which additional information is obtained from at least one of the inlet blood flow velocities u, v, w.
[0083] The flow distribution, in particular the flow distribution at the entrance of the target vessel, can be obtained from a corresponding image data set, which may have been recorded by a medical imaging system (e.g., a computed tomography (CT) device or a magnetic resonance imaging (MRI) device). The image data set is a three-dimensional (3D) data set with an inherent Cartesian coordinate system (which has orthogonal main directions x, y, z). The (entrance) flow distribution used as a second input to the AI unit includes three blood flow velocities u, v, w in the main directions x, y, z.
[0084] The three blood flow velocities u, v, w or more precisely the three Cartesian directions x, y, z can be transformed into different coordinate systems. For example, the Cartesian directions x, y, z can be transformed into a coordinate system having a path s along the center line, a radius r from the center line and an angle The radial coordinates of the axis, and the velocities u, v, w can be converted into longitudinal velocities, radial velocities and tangential velocities accordingly.
[0085] Furthermore, further information, such as secondary flow rates, can be derived from the velocities u, v, w and included in the (inlet) flow profile.
[0086] The (inlet) flow profile provides initial conditions for predicting at least one hemodynamic parameter in a very efficient manner.
[0087] According to a refinement of the invention, the vessel shape model is derived from an image dataset preferably generated by a medical imaging system, and most preferably from a 4D MRI flow dataset generated by a four-dimensional 4D magnetic resonance imaging MRI system.
[0088] The radius r and the coordinates x, y, z of the target vessel can be provided directly from the image dataset without any prior transformation. Thus, a vessel shape model and in particular a vessel shape point can be provided very efficiently.
[0089] According to a refinement of the invention, the flow distribution is derived from an image dataset preferably generated by a medical imaging system, and most preferably from a 4D MRI flow dataset generated by a 4D MRI system.
[0090] In particular, from the 4D MRI flow dataset, the blood flow velocities u, v, w of the (inlet) flow distribution can be provided directly without any previous conversion. The (inlet) flow distribution or more precisely the blood flow velocities u, v, w are measured directly at the patient and are therefore more accurate than any derived values. Furthermore, the higher the quality of the 4D MRI flow dataset (the higher the quality of the control image), the more accurate the (inlet) flow distribution / the more accurate the blood flow velocities u, v, w. Therefore, the inlet flow distribution can be provided very efficiently with a high degree of accuracy.
[0091] According to a further improvement of the present invention, the output block includes: at least one cascade layer and an output layer. The at least one cascade layer is communicatively connected to the deduction block. The output layer is communicatively connected to the last cascade layer in the at least one cascade layer and the deduction block. The at least one cascade layer and the output layer are jointly arranged and configured to generate a current hidden state h based on at least one generated j,i To determine at least one hemodynamic parameter p k .
[0092] At least one cascade layer may be a dense layer, a convolutional layer, etc. At least one cascade layer may include 4 to 256 nodes, and preferably includes 48 nodes. The output block may include 1 to 10 cascade layers.
[0093] The output layer may include one output node for each predicted hemodynamic parameter, and preferably includes four output nodes.
[0094] For example, at least one cascade layer is communicatively connected to the last deductive layer of the deductive block, wherein preferably all nodes of the cascade layer are connected to nodes of the last deductive layer. The output layer is communicatively connected to the last cascade layer of the at least one cascade layer and the last deductive layer of the deductive block, wherein all output nodes of the output layer are preferably connected to nodes of the last cascade layer of the at least one cascade layer and also to nodes of the last deductive layer. Thus, the output layer receives information from the last cascade layer and furthermore directly from the deductive block. In other words, the at least one cascade layer and the output layer are jointly arranged and configured to generate information based at least on the last generated current hidden state h M,i To determine at least one hemodynamic parameter p k , the last generated current hidden state h M,i implicitly contains at least one hemodynamic parameter p k All information.
[0095] An output block having at least one cascade layer and one output layer enables fast, stable and accurate deduction of at least one hemodynamic parameter from the last hidden state of the deduction block.
[0096] According to a sixth aspect of the invention, a computer-implemented method of training an AI unit (preferably a recurrent neural network (RNN) and most preferably a long short-term memory (LSTM) network) as comprised by an artificial intelligence (AI) system according to the second aspect of the invention comprises the following steps:
[0097] - receiving an input training dataset comprising a training vessel shape model and a corresponding training flow distribution.
[0098] - receiving an output training data set corresponding to the received training input data set, the output training data set comprising corresponding training hemodynamic parameters.
[0099] - training the AI unit to predict at least one hemodynamic parameter p of the target vessel using the received training vessel shape model and the corresponding training flow distribution and the received set of corresponding training outputs k .
[0100] According to a seventh aspect of the present invention, a computer program comprises instructions which, when the program is executed by a computer, cause the computer to perform the steps of the method according to the sixth aspect of the present invention.
[0101] According to an eighth aspect of the present invention, a computer readable medium has stored thereon a computer program according to the seventh aspect of the present invention.
[0102] According to a ninth aspect of the present invention, a data processing system comprises means for performing the steps of the method according to the sixth aspect of the present invention.
[0103] The learning rule or paradigm or more precisely the training rule or paradigm that is the basis of the current computer-implemented method of training AI units is an algorithm that modifies the parameters of the corresponding artificial neural network (ANN) such as the AI unit so as to produce a favorable output for a given input to the ANN. This training is usually equivalent to modifying the weights and thresholds of the variables within the ANN. Given the specific task and function class to be solved, learning means using a set of observations to find a function in the function class that solves the task in some optimal sense. This requires defining a cost function so that for the optimal solution, the cost is minimized and no other solution has a cost lower than the cost of the optimal solution. The cost function is an important concept in learning because it is a measure of the distance between a specific solution and the best solution to the problem to be solved. The learning algorithm searches in the solution space to find the function with the minimum possible cost. For applications where the solution is data-dependent, the cost must be a function of the observation, otherwise the model will be independent of the data. It is often defined as a statistic that can only be approximated. Any cost function can be defined, however, a particular cost function may be used because it has desirable properties (eg, convexity) or because it arises naturally from a particular form of the problem.
[0104] ANNs can be trained discriminatively using a standard back-propagation algorithm. Back-propagation is a method of calculating the gradient of a loss function relative to the weights in an ANN (producing the cost associated with a given state). The weight update of back-propagation can be accomplished via stochastic gradient descent. The choice of cost function depends on factors such as the type of learning (e.g., supervised, unsupervised, reinforcement, etc.) and the activation function. Typically, the activation function and cost function are the softmax function and the cross entropy function, respectively.
[0105] In other words, training an ANN essentially means selecting a model that minimizes the cost from a set of allowed models (or in a Bayesian framework, determining the distribution of the set of allowed models). Some form of gradient descent is usually deployed, which uses back propagation to calculate the actual gradient. This is done by simply taking the derivative of the cost function with respect to the network parameters and then changing these parameters in the direction associated with the gradient. Back propagation training algorithms are divided into three categories: steepest descent (with variable learning rate and momentum, elastic back propagation), quasi-Newton (Broyden-Fletcher-Goldfarb-Shanno, one-step secant), Levenberg-Marquardt, and conjugate gradient (Fletcher-Reeves update, Polak-Ribiére update, Powell-Beale restart, scaled conjugate gradient).
[0106] Common training paradigms include supervised learning, unsupervised learning, and reinforcement learning. Supervised learning uses a set of example pairs, and the purpose is to find a function in a class of allowed functions that matches the example. In other words, the mapping implied by the data is inferred; the cost function is related to the mismatch between the mapping of the data and the ANN, and it implicitly contains prior knowledge about the problem domain. The cost can be the mean square error that attempts to minimize the mean square error between the output of the ANN and the target value on all example pairs. The class of ANNs called multilayer perceptrons (MLPs) uses gradient descent to minimize this cost, resulting in a back propagation algorithm for training ANNs. In unsupervised learning, some data is given, and the cost function is minimized, which can be any function of the output of the data and the ANN. The cost function depends on the task and any prior assumptions (e.g., implicit attributes or parameters of the model, observed variables, etc.). In reinforcement learning, data is usually not given, but generated by the interaction of the agent with the environment. At each point in time, the agent performs an action, and the environment generates observations and instantaneous costs according to some (usually unknown) dynamics. The goal is to discover a policy for selecting actions that minimizes some measure of long-term cost (e.g., expected cumulative cost). The dynamics of the environment and the long-term cost of each policy are usually unknown, but can also be estimated. The environment is usually modeled as a Markov decision process (MDP) with states and actions, where the states and actions have the following probability distributions: instantaneous cost distribution, observation distribution, and transitions; while the policy is defined as the conditional distribution over actions given the observations. Together, the two then define a Markov chain (MC). The goal is to discover a policy (i.e., MC) that minimizes the cost.
[0107] The AI unit is trained based on the input training data set and the corresponding output data set. In order to train the AI unit sufficiently, it can be based on the hidden state h of the deductive block containing the actual AI network (such as RNN or LSTM network) j,i Reliably predict at least one hemodynamic parameter p k , the AI unit must be trained using at least about 1,000 different input training data sets and corresponding output data sets.
[0108] The received input training data set includes a training vessel shape distribution and a corresponding training flow distribution for a vessel that is the same type as a future target vessel for which at least one hemodynamic parameter is predicted. For example, the input training data set may include a training vessel shape model and a corresponding flow distribution for the aorta, such that the trained AI unit is arranged and configured to predict at least one hemodynamic parameter p of the aorta as a target vessel. kThe training vessel shape distribution and the corresponding training flow distribution may be obtained from an image dataset acquired by a medical imaging system, preferably from a 4D MRI flow dataset acquired by a 4D MRI system.
[0109] The received output training data set includes each training vessel shape distribution and the corresponding at least one hemodynamic parameter of the corresponding training flow distribution of the input training data set. The at least one training hemodynamic parameter may be obtained from an image data set acquired by a medical imaging system that is the same as the image data set used to obtain the training vessel shape distribution and the corresponding training flow distribution of the input training data set, preferably from a 4D MRI flow data set acquired by a 4D MRI system.
[0110] The AI unit is continuously provided with a training vessel shape distribution and a corresponding training flow distribution of the input training data set until all pairs of training vessel shape distributions and a corresponding training flow distribution of the input training data set have been used to train the AI unit. During training, internal weights of the layers of the AI unit are adapted based on differences in a current output of the AI unit compared to corresponding training hemodynamic parameters of the output training data set.
[0111] According to an improvement of the present invention, the computer-implemented method comprises the following initial steps:
[0112] - Determining the shape variability and geometric variability of blood vessels within a given population by applying statistical shape analysis to a set of true vessel shape models of the blood vessels.
[0113] - synthesizing at least one synthetic vessel shape model of the blood vessels based on the determined shape variability and geometric variability of the blood vessels within a given population.
[0114] - determining a corresponding training flow distribution for at least one synthetic vessel shape model and / or the set of real vessel shape models.
[0115] - Aggregating the input training data set with the set of true shape models and optionally at least one synthetic shape model and with the determined corresponding training flow distribution.
[0116] - calculating a corresponding at least one training hemodynamic parameter of an output training set based on the set of real shape models and / or at least one synthetic vessel shape model and a corresponding flow distribution of the input training data set.
[0117] The shape variability and geometric variability of blood vessels within a given population are determined via a data-driven approach. The statistical shape analysis receives a set of true blood vessel shape models of the blood vessels, which may be derived from or based on image datasets of blood vessels from different subjects of a given population generated using some or different medical imaging systems (such as computed tomography (CT) or MRI / 4D MRI imaging systems). Via the statistical shape analysis, the average and variation of the shapes of the blood vessels in the given population are calculated. Each variation can be interpreted as any human-defined shape definition (e.g., size, curvature, aortic length, width, height, etc.). In order to determine the shape variability and geometric variability of blood vessels within a given population, preferably 50 to 500 true blood vessel shape models of the blood vessels are used, and more preferably about 200 true blood vessel shape models. For example, 194 true blood vessel shape models of the blood vessels (based on image datasets of 194 different subjects of a given population) are used to determine the shape variability and geometric variability of blood vessels (e.g., aorta) within a given population.
[0118] The at least one synthesized vessel shape model is later used to train the AI unit. Preferably, 500 to 10,000 synthesized vessel shape models are synthesized based on the determined shape variability and geometric variability of the vessels within a given population, and most preferably, about 800 to 1,000 synthesized vessel shape models are synthesized.
[0119] Optionally, at least one synthesized vessel shape model of each of at least two different vessel shape model types is synthesized. The at least two different vessel shape model types may in particular be a mitral valve inlet model type and a tricuspid valve inlet model type, preferably of the aorta. Each of the at least one synthesized vessel shape model of each of the at least two different vessel shape model types is synthesized based on corresponding different shape variability and geometric variability of the vessels within a given population (i.e., separate image datasets of the mitral and tricuspid vessels or more precisely the aorta in the given population). The (same) AI unit may be trained using all synthesized vessel shape models of all at least two different vessel shape model types (and also all real vessel shape models of each of the at least two different vessel shape model types), such that the trained AI unit may predict at least one hemodynamic parameter for all corresponding at least two different vessel (shape) types (i.e., mitral and tricuspid vessels / aorta).
[0120] The corresponding training flow distribution of at least one synthetic vessel shape model and / or a set of real vessel shape models (optionally, of each of the at least two different shape model types) may be determined by statistical methods such as interpolation and / or extrapolation or by matching the derived shape parameters of the real vessel shape model and / or the synthetic vessel shape model with the shape parameters of a given vessel shape model in the target population. The flow distribution of at least a part of the set of real vessel shape models may be determined directly from an image dataset, preferably generated by a medical imaging system, and most preferably from a 4D MRI flow dataset, generated by a 4D MRI system.
[0121] The input training dataset may be aggregated based on at least one synthetic vessel shape model (of each of at least two different shape model types) and a set of real vessel shape models, such that a total of preferably 500 to 10,000 and most preferably about 1,000 (real and / or synthetic) vessel shape models are included in the input training dataset. For each (real / synthetic) vessel shape model, a determined corresponding flow distribution is also included in the input training dataset.
[0122] Based on the aggregated input data set, at least one corresponding training hemodynamic parameter of each real and / or synthetic blood vessel shape model with the corresponding flow distribution of the input data set is determined, preferably by physical calculation. The at least one training hemodynamic parameter of each (real / synthetic) blood vessel shape model and the corresponding flow distribution of the input training data set form an output training data set.
[0123] Since there are only a limited number of image datasets available for training of the blood vessels (e.g., the aorta) for which the AI unit is to be trained, the total number of blood vessel shape models in the training input dataset can be substantially increased by deriving synthetic blood vessel shape models based on available real blood vessel shape models. Thus, the AI unit can be trained with a sufficient number of real blood vessel shape models and synthetic blood vessel shape models so that the trained AI unit can predict at least one hemodynamic parameter of the target blood vessel with an accuracy sufficient to deduce a diagnosis.
[0124] According to an improvement of the present invention, computational fluid dynamics (CFD) simulation is used to calculate the training hemodynamic parameters of the output training set.
[0125] Through CFD simulation, the training hemodynamic parameters of the output training set are determined with high accuracy, so that the AI unit trained based on these training hemodynamic parameters can predict at least one hemodynamic parameter of the target blood vessel with high accuracy. This enables a reliable diagnosis by a clinician and / or a CDS system based on the predicted at least one hemodynamic parameter. BRIEF DESCRIPTION OF THE DRAWINGS
[0126] The present invention and its technical field are then further described in detail by the exemplary embodiments shown in the accompanying drawings. The exemplary embodiments are only useful for a better understanding of the present invention and should not be interpreted as limiting the scope of the present invention under any circumstances. In particular, if not explicitly described differently, various aspects of the subject matter described in the accompanying drawings can be extracted and combined with other components and findings of this specification or the accompanying drawings. The same reference numerals refer to the same objects, so that the descriptions from other drawings can be used supplementarily.
[0127] Figure 1 A schematic flow chart showing an embodiment of a method for predicting hemodynamic parameters according to the first aspect of the present invention.
[0128] Figure 2 A schematic flow chart showing a further embodiment of the method for predicting hemodynamic parameters according to the first aspect of the present invention.
[0129] Figure 3 A schematic flow chart showing a further embodiment of the method for predicting hemodynamic parameters according to the first aspect of the present invention.
[0130] Figure 4A A schematic diagram of an exemplary vessel shape model of the aorta of a human subject is shown.
[0131] Figure 4B A schematic diagram of another exemplary vessel shape model of the aorta of a human subject is shown.
[0132] Figure 5 A schematic diagram showing an embodiment of an artificial intelligence (AI) system for predicting hemodynamic parameters according to the second aspect of the present invention is shown.
[0133] Figure 6 A schematic diagram showing an improved implementation of an AI system for predicting hemodynamic parameters according to the second aspect of the present invention is shown.
[0134] Figure 7 Another schematic diagram shows an embodiment of an AI system for predicting hemodynamic parameters according to the second aspect of the present invention.
[0135] Figure 8 A schematic diagram showing an exemplary algorithm of a computer program according to the third aspect of the invention.
[0136] Fig. 9 A schematic diagram showing an embodiment of a computer readable medium according to a fourth aspect of the invention is shown.
[0137] Fig.10A schematic diagram showing an embodiment of a data processing system according to a fifth aspect of the present invention is shown.
[0138] Fig.11 A schematic flow chart of an embodiment of a computer-implemented method of training an artificial intelligence (AI) unit according to a sixth aspect of the present invention is shown.
[0139] Fig.12 A schematic diagram showing an exemplary algorithm of a computer program according to the seventh aspect of the invention.
[0140] Fig.13 A schematic diagram showing an embodiment of a computer readable medium according to an eighth aspect of the present invention is shown.
[0141] Fig.14 A schematic diagram showing an embodiment of a data processing system according to a ninth aspect of the present invention is shown. DETAILED DESCRIPTION
[0142] exist Figure 1 , an embodiment of the method for predicting hemodynamic parameters according to the first aspect of the present invention is exemplarily depicted. The method comprises the following steps: receiving 1 a blood vessel shape model; receiving 2 a corresponding flow distribution; and predicting 3 at least one hemodynamic parameter p k .
[0143] In the step of receiving 1, a vessel shape model of a target vessel is received as a first input, the target vessel being exemplarily the aorta of a human subject here. The vessel shape model has been extracted from an image dataset generated by a medical imaging system showing the aorta of the subject (e.g., a 4D MRI flow dataset generated by a 4D MRI system). The vessel shape model includes a centerline extending from an inlet to an outlet of the aorta. The centerline runs through all center points along the aorta. The vessel shape model includes a large number of N=100 vessel shape points, each of which has a first coordinate in a first direction x, a second coordinate in a second direction y perpendicular to the first direction x, and a third coordinate in a third direction z perpendicular to the first direction x and the second direction y. The vessel shape points of the vessel shape model are located on the centerline of the aorta and include the radius r of the aorta at the corresponding position of the vessel shape point along the centerline. In addition, each vessel shape point may contain additional information (e.g., a change in the radius of the aorta, a position along the centerline, etc.).
[0144] In the step of receiving 2, a corresponding inlet flow distribution of the aorta is received as a second input. The corresponding flow distribution of the aorta includes blood flow velocities u, v, w in the x, y, z directions at the inlet of the aorta, secondary flow degrees obtained from the blood flow velocities u, v, w, and boundary conditions (i.e., input flow rate and output flow rate at the inlet of the aorta). The inlet flow distribution is obtained from the same image data set as the blood vessel shape model.
[0145] In the prediction 3 step, four hemodynamic parameters p1, p2, p3, and p4 of the aorta (for example, systolic pressure, diastolic pressure, arterial blood pressure, mean arterial pressure, systolic pressure variation, pulse pressure variation, stroke volume variation, right atrial pressure, right ventricular pressure, pulmonary artery pressure, mean pulmonary artery pressure, pulmonary artery wedge pressure, left arterial pressure, cardiac output, cardiac index, stroke volume, systemic vascular resistance, systemic vascular resistance index, pulmonary vascular resistance, pulmonary vascular resistance index, left ventricular stroke work, left ventricular stroke work index, right ventricular stroke work, right ventricular stroke work index, coronary perfusion pressure, right ventricular end-diastolic volume, right ventricular end-systolic volume, right ventricular ejection fraction, etc.) are predicted by the AI unit at one of the vascular shape points of the vascular shape model.
[0146] The step of prediction 3 is iteratively performed for each consecutive vessel shape point of the received vessel shape model of the aorta, starting from the inlet to the outlet of the aorta. In other words, for each vessel shape point (1, ..., N, N = 100) of the provided vessel shape model of the aorta, four hemodynamic parameters p1, p2, p3, p4 are predicted in one iteration i, where i = 1, ..., N, and N = 100. The received vessel shape model of the aorta is used to predict the hemodynamic parameters p1, p2, p3, p4 of the aorta. k The first input of (here k=1, ..., 4) is the corresponding inlet flow distribution of the aorta received, which is used to predict the aortic hemodynamic parameter p k(here k=1, ..., 4). Four hemodynamic parameters p1, p2, p3, p4 are predicted based on the information and features contained in the received vascular shape model and the corresponding inlet flow distribution of the aorta. In particular, at each vascular shape point in the vascular shape model, the AI unit predicts the four hemodynamic parameters based on the three coordinates x, y, z of the vascular shape point and the corresponding radius r of the aorta and based on the features of the flow distribution at the inlet of the aorta (blood flow velocities u, v, w, the obtained secondary flow rate, and boundary conditions). Here, the AI unit is exemplarily an artificial neural network (ANN), which is trained to predict the hemodynamic parameters p1, p2, p3, p4 of the aorta based on the vascular shape model and the corresponding inlet flow distribution (of the aorta). k (k=1, ..., 4).
[0147] Each vessel shape point is iteratively stepped from the first vessel shape point of the received vessel shape model of the aorta to the last vessel shape point (along the centerline). For each vessel shape point, four hemodynamic parameters p1, p2, p3, p4 are predicted by the AI unit, where for the first vessel shape point (at the first iteration i=1), the inlet flow distribution is used as the second input, and for any subsequent vessel shape point (at any other iteration i>1), the internal information of the AI unit at the previous vessel shape point is used as the second input.
[0148] exist Figure 2 , further embodiments of the method for predicting hemodynamic parameters according to the first aspect of the present invention are exemplarily described. The method comprises: Figure 1 In addition, the step of determining 3 includes the following iterative (sub) steps: providing 4 a blood vessel shape point; providing 5 the received flow distribution; generating 6 at least one current hidden state h j,i ; and determining 7 at least one hemodynamic parameter p k .
[0149] The steps of providing 4, providing 5, generating 6 and determining 7 are iteratively performed for each blood vessel shape point 1, . . . , N (N=100) of the received blood vessel shape model of the aorta.
[0150] At each iteration i=1, ..., N (N=100), in the step of providing 4, one (continuous) blood vessel shape point of the received blood vessel shape model is provided to the first input block of the AI unit as a first input.
[0151] Only at the first iteration i=1, in the step of providing 5, the received inlet flow profile is provided to the second input block of the AI unit as a second input.
[0152] Information contained in the blood vessel shape points supplied to the first input block and information contained in the inlet flow distribution supplied to the second input block are converted into features by the respective input blocks and supplied to the deduction block.
[0153] At each iteration i=1, ..., N (N=100), in the step of generating 6, two current hidden states h are exemplarily generated in the deduction block of the AI unit. 1,i 、h M,i (M=2). The deductive block here exemplarily comprises two deductive layers j=1, ..., N, N=2. Based on a provided vessel shape point or more precisely from the features provided as a first input of the first input block and at the first iteration i=1 based on the provided flow distribution or more precisely from the provided features of the second input block or at any iteration i>1 after the first iteration based on two corresponding previous hidden states h of the deductive block as a second input 1,i-1 、h M,i-1 , to generate the two current hidden states h of the two deductive layers of the deductive block 1,i 、h M,i .
[0154] At each iteration i=1, ..., N (N=100), in step 7, the four hemodynamic parameters p1, p2, p3, p4 are determined by the output block of the AI unit. In each iteration, the final hidden state h of the second / final deductive layer M,i (here M = 2) contains the hemodynamic parameters p at the current vascular shape point k All information of (k=1, ..., 4). The final hidden state h 2,i The information is concatenated in the output block and merged into four hemodynamic parameters p1, p2, p3, and p4 at the current blood vessel shape point.
[0155] At the first iteration i=1, for the first blood vessel shape point, the first deductive layer j=1 receives features from the first input block (the transformed information of the first blood vessel shape point) and features from the second input block (the transformed information of the inlet flow distribution). The first deductive layer j=1 generates a first current hidden state h based on the features from the first input block and the second input block. 1,1 The second or here final deductive layer has a predefined initial state s M,1 (Here, exemplary, all nodes of the deductive layer are set to 0). Based on the current hidden state h of the first deductive layer 1,1 and the initial state s M,1 , the second / final deductive layer generates the second or more precisely the final current hidden state h 2,1 The final hidden state h M,1The information contained in is concatenated and merged by the output layer into four hemodynamic parameters p1, p2, p3, p4 at the first vessel shape point.
[0156] At each iteration i=2, ..., N (N=100) after the first iteration, the steps of providing 4, providing 5, generating 6 and determining 7 are performed in the same manner, except that each deductive layer of the deductive block receives the corresponding previous hidden state h 1,i-1 、h M,i-1 as the second input.
[0157] exist Figure 3 , further embodiments of the method for predicting hemodynamic parameters according to the first aspect of the present invention are exemplarily described. The method comprises: Figure 1 and Figure 2 In addition, the step of generating 6 includes the following iterative (sub) steps: generating 6.1 the first current hidden state h 1,i ; Generate 6.j at least one additional current hidden state h j,i ; and generate the final current hidden state h of 6.M M,i .
[0158] The steps of generating 6.1, generating 6.j and generating 6.M are iteratively performed for each blood vessel shape point 1, ..., N (N=100) of the received blood vessel shape model of the aorta.
[0159] The deduction block of the AI unit here exemplarily includes three deduction layers, namely a first deduction layer, an additional deduction layer and a final deduction layer.
[0160] At each iteration i=1, ..., N (N=100), in step 6.1, the first deductive layer generates a first current hidden state h of the first deductive layer. 1,i The first current hidden state h generated 1,i Based on a provided blood vessel shape point or more precisely, a feature provided by the first input block (the provided transformation information of the blood vessel shape point) and based on a provided inlet flow distribution or more precisely, a feature provided by the second input block at the first iteration i=1 (the provided transformation information of the inlet flow distribution) or at any iteration i>1 after the first iteration, a first previous hidden state h of the first deductive layer is provided 1,i-1 .
[0161] At each iteration i=1, ..., N, N=100, in the step of generating 6.j (where j=2), an additional deductive layer generates an additional current hidden state h of an additional deductive layer. j,i (here j=2). The generated additional current hidden state hj,i Based on the first current hidden state h of the first deductive layer 1,i and a predefined initial state s based on another deductive layer at the first iteration i=1 (here exemplarily, all nodes of the deductive layer are set to 0) j,1 (here j=2) or any iteration i=2, ..., N after the first iteration, the first previous hidden state h of one additional deductive layer provided j,i-1 .
[0162] At each iteration i=1, ..., N (N=100), in the step of generating 6.M (where M=3), the last deductive layer generates the last current hidden state h of the last deductive layer M,i (Here M = 3). The final current hidden state h generated M,i Based on another current hidden state h of another deductive layer j,i and a predefined initial state s based on the last deductive layer at the first iteration i=1 (here exemplarily, all nodes of the deductive layer are set to 0) M,1 (here M=2) or any iteration i=2, ..., N after the first iteration, the first previous hidden state h of one additional deductive layer provided j,i-1 .
[0163] exist Figure 4A , an exemplary vessel shape model 21 of an aorta 20 of a human subject is schematically depicted.
[0164] The blood vessel shape model 21 includes continuous blood vessel shape points 21.i (here, exemplarily, i=1, ..., 13 or 17, respectively) from the inlet (i=1) of the aorta 20 to the outlet (here, exemplarily, i=13 or 26, respectively) of the aorta 20. The blood vessel shape model 21 can be used only for a predefined portion of the aorta 20 of a human subject. The blood vessel shape model 21 can preferably include 50 to 500, most preferably 100, continuous blood vessel shape points (i=N, preferably, N∈[50; 500]], most preferably, N=100) from the inlet (i=1) of the aorta 20 to the outlet of the aorta 20. The aorta 20 has a centerline CL that runs through all the centerline points CLP.i of the aorta. Each blood vessel shape point 21.i includes one of the centerline points CLP.i with corresponding coordinates x, y, z, and also includes a radius ri of the aorta 20 that is orthogonal to the centerline CL at the corresponding centerline point CLP.i.
[0165] The vessel shape model 21 may be used for a target vessel 20, such as an aorta having more than one inlet and more than one outlet. Such a vessel shape model 21 may include vessel shape points 21.i, which include centerline points CLP.i along a plurality of centerlines CL, which connect the inlet and outlet of the target vessel 20 (e.g., the aorta) and extend uniformly in a cross section of the target vessel 20 or, more precisely, in a cross section of the vessel shape model 21.
[0166] exist Figure 4B , another exemplary vessel shape model 21 of an aorta 20 of a human subject is schematically depicted.
[0167] The vessel shape model 21 includes continuous vessel shape points 21.i (here exemplarily, i=1, ..., 13 or 17, respectively) from the inlet (i=1) of the aorta 20 to the outlet (here exemplarily, i=13 or 26, respectively) of the aorta 20. The vessel shape model 21 can be used only for a predefined portion of the aorta 20 of a human subject. The vessel shape model 21 can preferably include 50 to 500, most preferably 100, continuous vessel shape points (i=N, preferably, N∈[50; 500]], most preferably, N=100) from the inlet (i=1) of the aorta 20 to the outlet of the aorta 20 and optionally also along the branches of the aorta. The aorta 20 has a centerline CL that runs through all centerline points CLP.i of the aorta and optionally also through the centerline of each branch. Each blood vessel shape point 21.i includes one of the centerline points CLP.i having corresponding coordinates x, y, z and also includes a predetermined number k of surrounding points SP.i at the corresponding centerline point CLP.i. k The k surrounding points SP.i surround the corresponding center line point CLP.i in the form of a sphere Si, wherein the corresponding center line point CLP.i is located at the center of the sphere Si. The sphere Si of surrounding points has a radius c from the corresponding center line point CLP.i to the surface of the sphere Si.
[0168] The vessel shape model 21 may be used for a target vessel 20, such as the aorta having more than one inlet and more than one outlet. Such a vessel shape model 21 may include vessel shape points 21.i, which include centerline points CLP.i along a plurality of centerlines CL, which connect the inlets and outlets of the target vessel 20 (e.g., the aorta and the branches of the aorta) and extend uniformly in a cross section of the target vessel 20 or, more precisely, in a cross section of the vessel shape model 21.
[0169] exist Figure 5, an embodiment of an artificial intelligence (AI) system 30 for predicting hemodynamic parameters according to the second aspect of the present invention is schematically depicted. The AI system 30 is arranged and configured to perform the method for predicting hemodynamic parameters according to the first aspect of the present invention. The AI system 30 includes a first interface 31, a second interface 32 and an AI unit 10.
[0170] The first interface 31 and the second interface 32 are communicatively connected to the AI unit 10. A vessel shape model (VSM) 21 of a target vessel (e.g., the aorta of a human subject) is provided as a first input via the first interface 31 of the AI system 30. A corresponding flow profile (FP) 22 of the target vessel is provided as a second input via the second interface 32 of the AI system 30. The first input (VSM) and the second input (FP) are forwarded to the AI unit 10 of the AI system 30. Based on the forwarded vessel shape model as the first input and the flow profile as the second input, the AI unit 10 predicts at least one hemodynamic parameter p k (For example, four hemodynamic parameters p1, p2, p3, p4).
[0171] exist Figure 6 , an improved embodiment of an AI system 30 for predicting hemodynamic parameters according to the second aspect of the present invention is schematically depicted. Figure 6 The AI systems depicted include Figure 5 The same elements of the AI system.
[0172] The vessel shape model (VSM) 21 of the target vessel 20 can be obtained from a medical image obtained by a medical imaging system such as 4D MRI, computed tomography (CT), etc. The vessel shape model is a 3D point cloud here. The 3D point cloud 21 may include 10,000 to 100,000 points, preferably 40,000 points.
[0173] Based on the 3D point cloud, the centerline CL along the target vessel, in particular along the main vessel and its branches, can be calculated. For example, a surface model of the target vessel 20 can first be obtained from the 3D point cloud. Then, the centerline CL of the target vessel (i.e., the aorta and its branches) can be determined using the surface model.
[0174] The center line CL includes a plurality of consecutive center line points CLP.i. Each center line point CLP.i includes three coordinates x, y, z in the blood vessel shape model, i.e., the 3D point cloud.
[0175] The vessel shape model 21 includes a plurality of vessel shape points (VSP) 21.i, for each of which at least one hemodynamic parameter p is predicted. k .
[0176] As in Figure 6 The upper branch of Figure 4A ), each vessel shape point 21.i includes one of the consecutive centerline points CLP.i and the corresponding radius r of the target vessel at the corresponding centerline point CLP.i. The vessel shape point 21.i (i.e., i multiplied by the coordinates x, y, z of the centerline point CLP.i and the corresponding radius r) is iteratively provided as input to the first input block 11 of the system 30 for predicting the hemodynamic parameters (reference Figure 7 ).
[0177] Alternatively, if Figure 6 The lower branch of Figure 4B ), each blood vessel shape point 21.i includes one of the continuous centerline points CLP.i and a plurality of surrounding points SP.i surrounding each of the centerline points CLP.i. The surrounding points SP.i of the centerline point CLP.i k The center line point CLP.i is surrounded by a sphere Si having a radius c. The vessel shape point 21.i (i.e., i multiplied by the center line point CLP.i and the corresponding surrounding point SP.i) k k+1 coordinates x, y, z) of the system 30 are iteratively provided as input to the first input block 11 of the system 30 for predicting the hemodynamic parameters (control Figure 7 ).
[0178] exist Figure 7 , another schematic diagram of an embodiment of an AI system 30 for predicting hemodynamic parameters according to the second aspect of the present invention is schematically depicted. Figure 7 The AI system depicted in Figure 5 and Figure 6 The AI unit 10 of the AI system 30 includes a first input block 11, a second input block 12, a deduction block 13, and an output block 15. Here, the first input block 11 exemplarily includes a fully connected layer with 64 nodes. The blood vessel shape points 21.i provided to the first input block 11 each include a centerline point CLP.i and k surrounding points SP.i k In the case of , the first input block exemplarily includes k+1 times 64 nodes - that is, the center line point CP.i and each surrounding point SP.i kThe second input block 12 here exemplarily includes a fully connected layer with 64 nodes. Here, the deductive block 13 exemplarily includes two (j=1, ..., M, M=2) fully connected deductive layers 13.1, 13.2, each with 64 nodes. The blood vessel shape points 21.i provided to the first input block 11 each include a center line point CLP.i and k surrounding points SP.i k In the case of , the deductive layers 13.1, 13.2 each exemplarily include k+1 times 64 nodes, namely, the center line point CP.i and each surrounding point SP.i k The output block 15 here exemplarily includes a fully connected cascade layer 16 with 48 nodes and a fully connected output layer 17 with four nodes. Each hemodynamic parameter p to be predicted by the AI unit 10 k (k=1, ..., 4) corresponds to one node. In the case where the deductive layers 13.1, 13.2 each include k+1 times 64 nodes, the output block includes one fully connected cascade layer 16 with k+1 times 48 nodes and one fully connected output layer 17 with four nodes.
[0179] The first interface 31 is communicatively connected to the first input block 11 of the AI unit 10, and the second interface 32 is communicatively connected to the second input block 12 of the AI unit. The first input block 11 and the second input block 12 are communicatively connected to the first deductive layer 13.1 of the deductive block 13. The first deductive layer 13.1 is communicatively connected to the second or more precisely the last deductive layer 13.2 of the deductive block 13. The second / last deductive layer 13.2 is communicatively connected to the cascade layer 16 and the output layer 17 of the output block 15. The cascade layer 16 is communicatively connected to the output layer 17 of the output block.
[0180] exist Figure 7 In FIG. 1 , three iterations i=1, ..., N, N=3 are exemplarily depicted.
[0181] At each iteration i=1, ..., N (here N=3), the first input block 11 receives one continuous vessel shape point (VSP) 21.i of the vessel shape model 21 of the target vessel (e.g., the aorta of a human subject) from the first interface 31 starting from the inlet of the target vessel. The first input block 11 converts the information of the provided vessel shape points 21.i into features each represented by one node of a layer of the first input block 11.
[0182] At the first iteration i=1, the second input block 12 receives the (inlet) flow distribution 22 from the second interface 32. The received (inlet) flow distribution corresponds to the first vessel shape point 21.1 (at the inlet of the target vessel) received by the first input block 11. The second input block 12 converts the information of the provided (inlet) flow distribution 22 into features each represented by a node of a layer of the second input block 12. At any subsequent iteration i=2, ..., N (here N=3), no flow distribution is provided to the second input block.
[0183] At each iteration i=1, ..., N (here N=3), the first deductive layer 13.1 receives features (transformed information of the current vessel shape point 21.i) from the first input block 11. Only at the first iteration i=1, the first deductive layer 13.1 receives features (transformed information of the (inlet) flow distribution 22) from the second input block 12. At any subsequent iteration i=2, ..., N (here N=3), the first deductive layer 13.1 receives the previous hidden state h of the first deductive layer 13.1 from the previous iteration i-1. 1,i-1 The first deductive layer 13.1 transforms the provided features into the current first hidden state h 1,i , or more precisely, each of the provided blood vessel shape points VSP.i includes a centerline point CLP.i and k surrounding points SP.i k In the case of , the provided features are converted into a set of current first hidden states h of the first deductive layer 13.1 1,i 、h 1,i k .
[0184] At each iteration i=1, ..., N (here N=3), the second / final deductive layer 13.2 receives the current first hidden state h 1,i , or more precisely, each of the provided blood vessel shape points VSP.i includes a centerline point CLP.i and k surrounding points SP.i k In the case of receiving a set of current first hidden states h from the first deductive layer 13.1 1,i 、h 1,i k The second deductive layer 13.2 transforms the following into the current second / final hidden state h of the second / final deductive layer 13.2 2,i or more precisely a set of current second / last hidden states h 2,i 、h 2,i k : The current first hidden state h provided 1,i Or more precisely, a set of current first hidden states h 1,i 、h 1,ik , and at the first iteration i=1 is a predefined initial state s 2,1 , or more precisely, each of the provided blood vessel shape points VSP.i includes a centerline point CLP.i and k surrounding points SP.i k The second / final deductive layer of the case 13.2 has a set of predefined initial states s 2,1 、s 2,1 k , or at any subsequent iteration i=2, ..., N (here N=3) the previous second or more precisely the last hidden state h 2,i-1 Or more precisely the set of previous second or more precisely last hidden states h of the second / last deductive layer 13.2 from the previous iteration i-1 2,i-1 、h 2,i-1 k .
[0185] At each iteration i=1, ..., N (here N=3), the cascade layer 16 of the output block and also the output layer 17 receive the current second / last hidden state(s) h from the second / last deductive layer 13.2 2,i (h 2,i k ). The cascade layer is used to calculate the current second / last (one or more) hidden states h 2,i (h 2,i k ) are cascaded (64 features become 48 features, each node of the cascade layer corresponds to one feature). The output layer 17 receives the current second / last (one or more) hidden states h from the second / last deductive layer 13.2 2,i (h 2,i k ) and the current second / last (one or more) hidden states of the cascade from the cascade layer 16 (each corresponding to 48 features), and obtain at least one hemodynamic parameter p for the current vessel shape point 21.i k (Here, four hemodynamic parameters p1, p2, p3, p4 are exemplified).
[0186] exist Figure 8 In FIG. 1 , an exemplary algorithm of a computer program according to the third aspect of the present invention is schematically depicted. The computer program implements the method for predicting hemodynamic parameters according to the first aspect of the present invention.
[0187] Based on the first instruction I.1, the computer program causes the computer system to perform the following steps: receiving a vessel shape model of a target vessel (eg, an aorta of a human subject).
[0188] Based on the second instruction I.2, the computer program causes the computer system to perform the following steps: receiving a corresponding (inlet) flow profile of the target vessel.
[0189] Based on the third instruction I.3, the computer program causes the computer system to perform the following steps: predicting at least one hemodynamic parameter p based on the received blood vessel shape model and the received flow distribution by the AI unit; k .
[0190] exist Fig. 9 In FIG. 4 , an embodiment of a computer readable medium 40 according to the fourth aspect of the present invention is schematically depicted.
[0191] Here, illustratively, a computer readable storage disc 40 such as a compact disc (CD), a digital video disc (DVD), a high-definition DVD (HD DVD) or a Blu-ray disc (BD) has stored thereon the data according to the third aspect of the present invention and as described in Figure 7 However, the computer readable medium 40 may also be a data storage device, such as a magnetic storage device / memory (e.g., a core memory, a magnetic tape, a magnetic card, a magnetic stripe, a bubble storage device, a drum storage device, a hard disk drive, a floppy disk, or a removable storage device), an optical storage device / memory (e.g., a holographic memory, an optical tape, a Tesa tape, a laser disc (Laserdisc), a Phasewriter (Phasewriter Dual, PD) or an ultra-density optical (UDO)), a magneto-optical storage device / memory (e.g., a MiniDisc or a magneto-optical disc (MO-Disk)), a volatile semiconductor / solid-state memory (e.g., a random access memory (RAM), a dynamic RAM (DRAM) or a static RAM (SRAM)), a non-volatile semiconductor / solid-state memory (e.g., a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrical EPROM (EEPROM), a flash-EEPROM (e.g., a USB-stick), a ferroelectric RAM (FRAM), a magnetoresistive RAM (MRAM) or a phase-change RAM).
[0192] exist Fig.10 , an embodiment of a data processing system 50 according to the fifth aspect of the present invention is schematically depicted.
[0193] The data processing system 50 may be a personal computer (PC), a laptop computer, a tablet computer, a server, a distributed system (e.g., a cloud system), etc. The data processing system 50 includes a central processing unit (CPU) 51, a memory having a random access memory (RAM) 52 and a non-volatile memory (MEM, such as a hard disk) 53, a human-machine interface device (HID, such as a keyboard, a mouse, a touch screen, etc.) 54, and an output device (MON, such as a monitor, a printer, a speaker, etc.) 55. The CPU 51, the RAM 52, the HID 54, and the MON 55 are communicatively connected via a data bus. The RAM 52 and the MEM 53 are communicatively connected via another data bus. The data processing system 50 may be configured to be a computer system according to the third aspect of the present invention and to ... Figure 7 The computer program schematically depicted in FIG. 5 is loaded from the MEM 53 or another computer readable medium 40 into the RAM 52. According to the computer program, the CPU 51 executes the first aspect of the invention and Figures 1 to 3 5. The execution may be initiated and controlled by a user via HID 54. The status and / or results of the executed computer program may be indicated to the user by MON 55. The results of the executed computer program may be permanently stored on non-volatile MEM 53 or another computer readable medium.
[0194] In particular, the CPU 51 and RAM 53 for executing the computer program may include, for example, several CPUs 51 and several RAMs 53 in a computing cluster or cloud system. The HID 54 and MON 55 for controlling the execution of the computer program may be included by different data processing systems, such as terminals communicatively connected to the data processing system 50 (e.g., a cloud system).
[0195] exist Fig.11 , schematically depicts an embodiment of a computer-implemented method for training an AI unit according to the sixth aspect of the invention. The computer-implemented method comprises the following steps: receiving 66 an input training data set; receiving 67 an output training data set; and training 68 an AI unit. Optionally, the computer-implemented method further comprises the following initial steps: determining 61 shape variability and geometric variability; synthesizing 62 at least one synthetic vessel shape model; determining 63 a corresponding training flow distribution; aggregating 64 the input data set; and calculating 65 at least one corresponding hemodynamic parameter.
[0196] In the step of receiving 66, an input training data set is received. The received input training data set includes a training vessel shape distribution and a corresponding training flow distribution of a vessel (here, the aorta of a human subject of a target population). The training vessel shape distribution and the corresponding training flow distribution are the same as the vessel type of the future target vessel for which at least one hemodynamic parameter (e.g., four hemodynamic parameters p1, p2, p3, p4) is predicted. The training vessel shape distribution and the corresponding training flow distribution may be obtained from an image data set acquired by a medical imaging system, such as a 4D MRI flow data set acquired by a 4D MRI system.
[0197] In the step of receiving 67, an output training data set is received. The received output data set includes at least one training hemodynamic parameter corresponding to each of the training vessel shape distribution and the corresponding training flow distribution of the input training data set (here exemplarily, each of the training vessel shape distribution and the corresponding training flow distribution corresponds to four training hemodynamic parameters). The at least one training hemodynamic parameter can be obtained from an image data set acquired by a medical imaging system that is the same as the image data set used to obtain the training vessel shape distribution and the corresponding training flow distribution of the input training data set (for example, from a 4D MRI flow data set acquired by a 4D MRI system).
[0198] In the step of training 68, the AI unit is continuously provided with a training vessel shape distribution and a corresponding training flow distribution of the input training data set from the first pair to the last pair. During training 68, the internal weights of the layers of the AI unit (here exemplarily, a fully connected input layer with 64 nodes of the first input block, a fully connected input layer with 64 nodes of the second input block, two fully connected LSTM deductive layers with 64 nodes of the deductive block, and a fully connected cascade layer with 48 nodes of the output block and a fully connected output layer with 4 nodes) are adapted based on the difference of the current output of the (output layer of) the AI unit compared with the corresponding training hemodynamic parameters of the output training data set. After training 68 has been completed, the trained AI unit is arranged and configured to predict the four hemodynamic parameters p of the aorta as the target vessel. k , k=1,...,4.
[0199] In an optional initial step of determining 61, shape variability and geometric variability of a blood vessel (here exemplarily the aorta) within a given population (here exemplarily 200 subjects) is determined by applying statistical shape analysis to a set of true vessel shape models of the aorta. The shape variability and geometric variability of the aorta within the given population is determined via a data-driven approach, wherein the statistical shape analysis receives the set of true vessel shape models of the aorta. The true vessel shape models are derived from or based specifically on an image dataset (e.g., a 4D MRI flow dataset) of the aorta from the 200 subjects of the given population. Via the statistical shape analysis, the average and variation of the shape of the aorta in the given population is calculated.
[0200] In an optional initial step of synthesis 62, at least one synthetic vessel shape model is synthesized based on the determined shape variability and geometric variability of the aorta within the given population. Exemplarily, here are 800 synthetic vessel shape models synthesized based on the determined shape variability and geometric variability of the aorta within the given population. Therefore, there are synthetic vessel shape models for each of the two different vessel shape model types (mitral valve and tricuspid valve inlet model types of the aorta) that are synthesized. The synthetic vessel shape models for each of the two different vessel shape model types are synthesized based on the corresponding different shape variability and geometric variability of the aorta with a bicuspid valve and the aorta with a tricuspid valve in the given population. The AI unit is later trained with all the synthetic vessel shape models of the two different vessel shape model types so that the trained AI unit can predict the four hemodynamic parameters p of the bicuspid and tricuspid aortas. k , k=1,...,4.
[0201] In the step of determining 63, corresponding training flow distributions of 800 synthetic vessel shape models of mitral and tricuspid valve shape model types and additionally or alternatively of a set of real vessel shape models are determined by statistical methods (e.g., interpolation and / or extrapolation or by matching the obtained shape parameters of the real vessel shape models and / or synthetic vessel shape models with the shape parameters of a given vessel shape model in the target population). Alternatively, the flow distribution of at least a portion of the set of real vessel shape models may be directly determined based on an image dataset generated by a medical imaging system (e.g., a 4D MRI flow dataset generated by a 4D MRI system).
[0202] In the step of aggregation 64, the input training dataset is aggregated from the set of 200 real vessel shape models and 800 synthetic vessel shape models of the mitral and tricuspid valve shape model types, so that a total of 1000 (real and synthetic) vessel shape models are included in the input training dataset. For each (real / synthetic) vessel shape model, the determined corresponding flow distribution is also included in the input training dataset.
[0203] In the step of calculating 65, the corresponding four training hemodynamic parameters of each of the 1000 (real and synthetic) blood vessel shape models are calculated for the output training data set. Based on the aggregated input data set, i.e., the 1000 (real and synthetic) blood vessel shape models and the corresponding flow distributions, the corresponding four training hemodynamic parameters of each (real and synthetic) blood vessel shape model and the corresponding flow distribution of the input data set are determined by physical calculation.
[0204] exist Fig.12 In FIG. 1 , an exemplary algorithm of a computer program according to the seventh aspect of the invention is schematically depicted. The computer program implements a computer-implemented method for training an AI unit according to the sixth aspect of the invention.
[0205] Based on optional first instruction I.61, the computer program causes the computer system to perform the following steps: determine the shape variability and geometric variability of a blood vessel (e.g., the aorta) within a given population (e.g., 200 human subjects) by applying statistical shape analysis to a set of real vascular shape models of the blood vessels.
[0206] Based on the optional second instruction I.62, the computer program causes the computer system to perform the following steps: synthesizing at least one synthetic vessel shape model of the blood vessels based on the determined shape variability and geometric variability of the blood vessels within the given population.
[0207] Based on the optional third instruction I.63, the computer program causes the computer system to perform the following steps: determining a corresponding training flow distribution for at least one synthetic vessel shape model and / or a set of real vessel shape models.
[0208] Based on the optional fourth instruction I.64, the computer program causes the computer system to perform the following steps: aggregating the input training data set with a set of real shape models and optionally at least one synthetic shape model and with the determined corresponding training flow distribution.
[0209] Based on the optional fifth instruction I.65, the computer program causes the computer system to perform the following steps: calculate at least one training hemodynamic parameter (for example, four hemodynamic parameters) corresponding to the output training set based on a set of real vascular shape models and / or at least one synthetic vascular shape model and the corresponding flow distribution of the input training data set.
[0210] Based on the sixth instruction I.66, the computer program causes the computer system to perform the following steps: receiving an input training data set, the input training data set comprising a training blood vessel shape model and a corresponding training flow distribution.
[0211] Based on the seventh instruction I.67, the computer program causes the computer system to perform the following steps: receiving an output training data set corresponding to the received training input data set, the output training data set including at least one (here exemplarily four) corresponding training hemodynamic parameter.
[0212] Based on the eighth instruction I.68, the computer program causes the computer system to perform the following steps: training the AI unit to predict at least one hemodynamic parameter p of the target vessel (e.g., the aorta) using the received training vessel shape model and the corresponding training flow distribution and the received corresponding training output set; k (Here are the four hemodynamic parameters p1, p2, p3, p4).
[0213] exist Fig.13 , an embodiment of a computer readable medium 70 according to the eighth aspect of the present invention is schematically depicted.
[0214] Here, illustratively, a computer readable storage disc 70 such as a compact disc (CD), a digital video disc (DVD), a high-definition DVD (HD DVD) or a Blu-ray disc (BD) has stored thereon the data according to the seventh aspect of the present invention and as described in Fig.11 However, the computer readable medium 70 may also be a data storage device, such as a magnetic storage device / memory (e.g., a core memory, a magnetic tape, a magnetic card, a magnetic stripe, a bubble storage device, a drum storage device, a hard disk drive, a floppy disk, or a removable storage device), an optical storage device / memory (e.g., a holographic memory, an optical tape, a Tesa tape, a laser disk (Laserdisc), a Phasewriter (Phasewriter Dual, PD) or an ultra-density optical (UDO)), a magneto-optical storage device / memory (e.g., a MiniDisc or a magneto-optical disk (MO-Disk)), a volatile semiconductor / solid-state memory (e.g., a random access memory (RAM), a dynamic RAM (DRAM) or a static RAM (SRAM)), a non-volatile semiconductor / solid-state memory (e.g., a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrical EPROM (EEPROM), a flash-EEPROM (e.g., a USB-stick), a ferroelectric RAM (FRAM), a magnetoresistive RAM (MRAM) or a phase-change RAM).
[0215] exist Fig.14 , an embodiment of a data processing system 80 according to the ninth aspect of the present invention is schematically depicted.
[0216] The data processing system 80 may be a personal computer (PC), a laptop computer, a tablet computer, a server, a distributed system (e.g., a cloud system), etc. The data processing system 80 includes a central processing unit (CPU) 81, a memory having a random access memory (RAM) 82 and a non-volatile memory (MEM, such as a hard disk) 83, a human-machine interface device (HID, such as a keyboard, a mouse, a touch screen, etc.) 84, and an output device (MON, such as a monitor, a printer, a speaker, etc.) 85. The CPU 81, the RAM 82, the HID 84, and the MON 85 are connected in communication via a data bus. The RAM 82 and the MEM 83 are connected in communication via another data bus. The data processing system 80 may be connected in communication via a data bus according to the seventh aspect of the present invention and Fig.11 The computer program schematically depicted in FIG. 8 is loaded from the MEM 83 or another computer readable medium 70 into the RAM 82. According to the computer program, the CPU 81 executes the sixth aspect of the present invention and Fig.11 8. The execution may be initiated and controlled by a user via HID 84. The status and / or results of the executed computer program may be indicated to the user by MON 85. The results of the executed computer program may be permanently stored on non-volatile MEM 83 or another computer readable medium.
[0217] In particular, the CPU 81 and RAM 83 for executing the computer program may include, for example, several CPUs 81 and several RAMs 83 in a computing cluster or cloud system. The HID 84 and MON 85 for controlling the execution of the computer program may be included by different data processing systems such as terminals communicatively connected to the data processing system 80 (e.g., cloud system).
[0218] Although specific embodiments have been illustrated and described herein, a person of ordinary skill in the art will recognize that there are various alternative and / or equivalent implementations. It should be understood that one or more exemplary embodiments are merely examples and are not intended to limit the scope, applicability, or configuration in any way. On the contrary, the foregoing summary of the invention and detailed description will provide a convenient roadmap for those skilled in the art to implement at least one exemplary embodiment, and it should be understood that various changes may be made to the functions and arrangements of the elements described in the exemplary embodiments without departing from the scope described in the attached claims and their legal equivalents. In general, the present application is intended to cover any adaptation or variation of the specific embodiments discussed herein.
[0219] In the foregoing detailed description, various features are combined together in one or more examples for the purpose of simplifying the present disclosure. It should be understood that the above description is intended to be illustrative rather than restrictive. It is intended to cover all alternatives, modifications and equivalents that may be included within the scope of the present invention. After reviewing the above description, many other examples will be apparent to those skilled in the art.
[0220] The specific nomenclature used in the foregoing description is used to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art from the description provided herein that the present invention can be practiced without specific details. Therefore, for the purpose of illustration and description, the foregoing description of specific embodiments of the present invention is given. They are not intended to be exhaustive or limit the present invention to the precise form disclosed; obviously, in view of the above teachings, many modifications and variations are possible. The embodiments are selected and described in order to best illustrate the principles of the present invention and its practical application, so that those skilled in the art can best utilize the present invention and various embodiments with various modifications suitable for the specific purposes contemplated. Throughout the specification, the terms "including" and "in..." are used as the ordinary English equivalents of the corresponding terms "including" and "wherein", respectively. In addition, the terms "first", "second", and "third", etc. are used only as labels and are not intended to impose numerical requirements on their objects or establish a specific order of importance of their objects. In the context of this specification and the claims, the conjunction "or" should be understood as including ("and / or") rather than exclusive ("either...or").
Claims
1. A method for predicting a hemodynamic parameter (p k ), comprising the following steps: - receiving (1) a blood vessel shape model (21) of the target blood vessel (20), wherein the blood vessel shape model (21) of the target blood vessel (20) is a data set representing a polygonal mesh, and / or wherein the blood vessel shape model (21) of the target blood vessel (20) comprises a volume mesh formed by 3D elements; - receiving (2) a corresponding flow profile (22) of the target blood vessel (20), wherein the corresponding flow profile (22) of the target blood vessel (20) comprises characteristics of blood flow at an entrance of the target blood vessel (20); and - predicting (3) at least one hemodynamic parameter (p) by the artificial intelligence AI unit (10) based on the received blood vessel shape model (21) and the received flow distribution (22); k ).
2. The method according to claim 1, wherein: In the step of predicting (3), the at least one hemodynamic parameter is predicted for at least one blood vessel shape point (21.i) of the received blood vessel shape model (21).
3. The method according to claim 1 or 2, wherein: The step of predicting (3) is iteratively performed for each of at least two consecutive blood vessel shape points (21.1, ..., 21.N) of the received blood vessel shape model (21).
4. The method according to claim 1 or 2, wherein: The prediction step (3) includes the following iterative steps: - providing (4) a blood vessel shape point (21.i) of the received blood vessel shape model (21) to a first input block (11) of the AI unit (10) at each iteration (i); - providing (5) the received flow distribution (22) to a second input block (12) of the AI unit (10) at a first iteration (i=1); - based on a provided vessel shape point (21.i) and based on at least one previous hidden state (h) of the deductive block (13) of the AI unit (10) at the provided flow distribution (22) at the first iteration (i=1) or at any iteration (i>1) after the first iteration j,i-1 ), generating (6) at least one current hidden state (h j,i );as well as - based on at least one generated current hidden state (h j,i ) is the last generated current hidden state (h M,i ), determining (7) the at least one hemodynamic parameter (p) in an output block (15) of the AI unit (10) k ).
5. The method according to claim 4, wherein: The steps of generating (6) include: - based on a provided blood vessel shape point (21.i) and based on a provided flow distribution (22) at the first iteration (i=1) or at any iteration after the first iteration (i>1) a provided first previous hidden state (h) of the first deductive layer (13.1) of the AI unit (10) 1,i-1 ), generating (6.1) a first current hidden state (h 1,i );as well as - The current hidden state (h) based on the previous deductive layer (13.M-1) M-1,i ) and the initial state (s) of the final deductive layer (13.M) based on the AI unit (10) M,1 ) or the last previous hidden state (h M,i-1 ), generating (6.M) the final current hidden state (h M,i ).
6. The method according to claim 5, wherein: The step of generating (6) further includes: - The current hidden state (h) based on the previous deductive layer (13.j-1) j-1,i ) and an initial state (s ) of at least one further deductive layer (13.j) of the AI unit (10) at the first iteration (i=1) j,1 ) or any iteration (i>1) after the first iteration, the provided additional previous hidden state (h j,i-1 ), generating (6.j) at least one further current hidden state (h j,i ).
7. The method according to claim 6, wherein: The initial state (s) of the at least one further deductive layer (13.j) j,1 ) is the initial state (s) of the memory unit (14.j) of the at least one further deductive layer (13.j) j,1 ).
8. The method according to any one of claims 1, 2 and 5, wherein: The blood vessel shape model (21) comprises a center line (CL) along a main direction of the target blood vessel (20), and wherein each continuous blood vessel shape point (21.i) comprises a continuous center line point (CLP.i) on the center line (CL) and a radius (ri) of the target blood vessel (20) at the corresponding center line point (CLP.i).
9. The method according to claim 4, wherein: The blood vessel shape model (21) comprises a center line (CL) along a main direction of the target blood vessel (20), and wherein each continuous blood vessel shape point (21.i) comprises a continuous center line point (CLP.i) on the center line (CL) and a predetermined number k of surrounding points (SP.i) surrounding each center line point (CLP.i) in the form of a sphere (Si). k ), where the corresponding centerline point (CLP.i) is located at the center of the corresponding sphere (Si), The step of generating (6) includes: - Based on the provided information including the corresponding center line point (CLP.i) and k surrounding points (SP.i) around the center line point (CLP.i) k ) and a provided first previous hidden state (h) of the first deductive layer (13.1) of the AI unit (10) based on the provided flow distribution (22) at the first iteration (i=1) or at any iteration (i>1) after the first iteration. 1,i-1 ,h 1,i-1 k ), generating (6.1) a first current hidden state (h 1,i ,h 1,i k );as well as - The current hidden state (h) based on the previous deductive layer (13.M-1) M-1,i ,h M-1,i k ) and the initial state (s) of the final deductive layer (13.M) based on the AI unit (10) M,1 ,s M,1 k ) or the last previous hidden state (h M,i-1 ,h M,i-1 k ), generating (6.M) the final current hidden state (h M,i ,h M,i k ),as well as Among them, in the step of determining (7), based on the last generated current hidden state (h M,i ,h M,i k ) determines in an output block (15) of the AI unit (10) the at least one hemodynamic parameter (p k ).
10. The method according to claim 9, wherein: The step of generating (6) further includes: - The current hidden state (h) based on the previous deductive layer (13.j-1) j-1,i ,h j-1,i k ) and an initial state (s ) of at least one further deductive layer (13.j) of the AI unit (10) at the first iteration (i=1) j,1 ,s j,1 k ) or any iteration (i>1) after the first iteration, the provided additional previous hidden state (h j,i-1 ,h j,i-1 k ), generating (6.j) in said at least one further deductive layer (13.j) at least one further set of current hidden states (h j,i ,h j,i k ).
11. The method according to claim 10, wherein: The initial state (s) of the at least one further deductive layer (13.j) j,1 ,s j,1 k ) is a memory unit (14.j, 14.j) of the at least one further deductive layer (13.j) k )'s initial state (s j,1 ,s j,1 k ).
12. The method according to any one of claims 1, 2 and 5, wherein: The flow distribution (22) comprises: - inlet blood flow velocity u in direction x of the corresponding image dataset; - an inlet blood flow velocity v of the image data set in a direction y orthogonal to the direction x; and - an inlet blood flow velocity w in a direction z orthogonal to the directions x and y of the image dataset.
13. The method according to claim 12, wherein: The flow distribution (22) further comprises: - Additional information derived from at least one of said inlet blood flow velocities u, v, w.
14. The method according to claim 13, wherein: The additional information is the secondary fluidity.
15. The method according to claim 12, wherein: The flow distribution (22) further comprises: -Boundary conditions.
16. The method according to any one of claims 1, 2 and 5, wherein: The blood vessel shape model (21) is obtained from an image dataset.
17. The method according to claim 16, wherein: The blood vessel shape model (21) is derived from an image dataset generated by a medical imaging system.
18. The method according to claim 17, wherein: The blood vessel shape model (21) is obtained from a 4D MRI flow data set generated by a 4D MRI system.
19. The method according to any one of claims 1, 2 and 5, wherein: The flow distribution (22) is obtained from an image dataset.
20. The method according to claim 19, wherein: The flow distribution (22) is derived from an image dataset generated by a medical imaging system.
21. The method according to claim 20, wherein: The flow distribution (22) is obtained from a 4D MRI flow dataset generated by a 4D MRI system.
22. A method for predicting a hemodynamic parameter (p k ), the artificial intelligence AI system (30) being arranged and configured to perform the method according to any of the preceding claims, the artificial intelligence AI system (30) comprising: - a first interface (31) arranged and configured to receive a vessel shape model (21) of the target vessel (20), wherein the vessel shape model (21) of the target vessel (20) is a data set representing a polygonal mesh and / or wherein the vessel shape model (21) of the target vessel (20) comprises a volume mesh formed by 3D elements; - a second interface (32) arranged and configured to receive a corresponding flow profile (22) of the target vessel (20), wherein the corresponding flow profile (22) of the target vessel (20) comprises characteristics of blood flow at an inlet of the target vessel (20); and - an AI unit (10) which is communicatively connected to the first interface (31) and the second interface (32) and is arranged and configured to predict at least one hemodynamic parameter (p) based on the received vessel shape model (21) and the received flow distribution (22); k ).
23. The AI system (30) of claim 22, wherein: The AI unit (10) includes a recurrent neural network (RNN).
24. The AI system (30) of claim 23, wherein: The AI unit (10) includes a long short-term memory (LSTM) network.
25. The AI system according to any one of claims 22 to 24, wherein: The AI unit (10) comprises: - a first input block (11) which is communicatively connected to said first interface (31) and is arranged and configured to obtain a provision of one vessel shape point (21.i) of the received vessel shape model (21) at each iteration (i); - a second input block (12) communicatively connected to said second interface (32) and arranged and configured to obtain the provided received flow profile (22) at a first iteration (i=1); a deductive block (13) communicatively connected to the first input block (11) and the second input block (12) and arranged and configured to generate a first hidden state (h) of the deductive block (13) based on a provided one vessel shape point (21.i) and based on a provided flow distribution (22) at the first iteration (i=1) or at any iteration (i>1) after the first iteration; j,i-1 ) to generate at least one current hidden state (h j,i );as well as - an output block (15) communicatively connected to the deduction block (13) and arranged and configured to generate a current hidden state (h j,i ) in the last generated current hidden state (h M,i ) to determine the at least one hemodynamic parameter (p k ).
26. The AI system (30) of claim 25, wherein: The AI unit (10) comprises: a first deductive layer (13.1) communicatively connected to the first input block (11) and the second input block (12) and arranged and configured to generate a first previous hidden state (h) of the first deductive layer (13.1) based on a provided one vessel shape point (21.i) and based on a provided flow distribution (22) at the first iteration (i=1) or at any iteration (i>1) after the first iteration; 1,i-1 ), to generate the first current hidden state (h 1,i );as well as a final interpretation layer (13.M) which is communicatively connected to the first interpretation layer (13.1) or the final additional interpretation layer of the at least one additional interpretation layer (13.j) and is arranged and configured to determine the current hidden state (h) of the previous interpretation layer (13.M-1) based on the current hidden state (h M-1,i ) and the initial state (s ) of the last deductive layer (13.M) based on the first iteration (i=1) M,1 ) or any iteration (i>1) after the first iteration, the last previous hidden state (h M,i-1 ), to generate the final current hidden state (h M,i ).
27. The AI system (30) of claim 26, wherein: The AI unit (10) comprises: - at least one further deductive layer (13.j) communicatively connected to said first deductive layer (13.1) and arranged and configured to determine the current hidden state (h) of a previous deductive layer (13.j-1) based on the current hidden state (h j-1,i ) and based on the initial state (s ) of the at least one further deductive layer (13.j) at the first iteration (i=1) j,1 ) or any iteration (i>1) after the first iteration, the provided additional previous hidden state (h j,i-1 ), to generate at least one additional current hidden state (h j,i ).
28. The AI system (30) of claim 27, wherein: The initial state (s) of the at least one further deductive layer (13.j) j,1 ) is the initial state (s) of the memory unit (14.j) of the at least one further deductive layer (13.j) j,1 ).
29. The AI system (30) of claim 26, wherein: The initial state (s) of the final deductive layer (13.M) M,1 ) is the initial state (s) of the memory cell (14.M) of the last deductive layer (13.M) M,1 ).
30. The AI system (30) of claim 27, wherein: The first input block (11) is arranged and configured to obtain a provided blood vessel shape point (21.i) at each iteration (i), the blood vessel shape point (21.i) comprising a continuous centerline point (CLP.i) on a centerline (CL) along a main direction of the target blood vessel (20) and a predetermined number k of surrounding points (SP.i) surrounding each centerline point (CLP.i) in the form of a sphere (Si). k ), where the corresponding centerline point (CLP.i) is located at the center of the corresponding sphere (Si), The first deductive layer (13.1) is arranged and configured to provide a corresponding center line point (CLP.i) and k surrounding points (SP.i) around the center line point (CLP.i). k ) and a provided first previous hidden state (h) of the first deductive layer (13.1) based on the provided flow distribution (22) at the first iteration (i=1) or at any iteration (i>1) after the first iteration. 1,i-1 ,h 1,i-1 k ), to generate the first current hidden state (h 1,i ,h 1,i k ), wherein the at least one further deductive layer (13.j) is arranged and configured to be based on the current hidden state (h) of the previous deductive layer (13.j-1) j-1,i ,h j-1,i k ) and based on the initial state (s ) of the at least one further deductive layer (13.j) at the first iteration (i=1) j,1 ,s j,1 k ) or any iteration (i>1) after the first iteration, the provided additional previous hidden state (h j,i-1 ,h j,i-1 k ), to generate at least one additional set of current hidden states (h j,i ,h j,i k ), Wherein, the last deductive layer (13.M) is arranged and configured to be based on the current hidden state (h) of the previous deductive layer (13.M-1). M-1,i ,h M-1,i k ) and the initial state (s ) of the last deductive layer (13.M) based on the first iteration (i=1) M,1 ,s M,1 k ) or any iteration (i>1) after the first iteration, the last previous hidden state (h M,i-1 ,h M,i-1 k ), to generate the final current hidden state (h M,i ,h M,i k ),as well as The output block (15) is arranged and configured to generate a current hidden state (h M,i ,h M,i k ) to determine the at least one hemodynamic parameter (p k ).
31. The AI system (30) of claim 30, wherein: The initial state (s) of the at least one further deductive layer (13.j) j,1 ,s j,1 k ) is a memory unit (14.j, 14.j) of the at least one further deductive layer (13.j) k )'s initial state (s j,1 ,s j,1 k ).
32. The AI system (30) of claim 30, wherein: The initial state (s) of the final deductive layer (13.M) M,1 ,s M,1 k ) is the memory unit (14.M, 14.M) of the last deductive layer (13.M) k )'s initial state (s M,1 ,s M,1 k ).
33. The AI system (30) of claim 25, wherein: The output block (15) comprises: - at least one cascading layer (16) communicatively connected to said deduction block (13); and - an output layer (17) communicatively connected to the last cascade layer of the at least one cascade layer (16) and the deduction block (13), wherein the at least one cascade layer (16) and the output layer (17) are jointly arranged and configured to generate a current hidden state (h j,i ) to determine the at least one hemodynamic parameter (p k ).
34. A computer program comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 21.
35. A computer readable medium (40) having stored thereon a computer program according to claim 34.
36. A data processing system (50) comprising means (51, 52, 53, 54, 55) for performing the steps of the method according to any one of claims 1 to 21.
37. A computer-implemented method for training an artificial intelligence (AI) unit (10), the AI unit (10) being comprised by an AI system (30) according to any one of claims 22 to 33, the method comprising the following steps: - receiving (66) an input training data set comprising a training vessel shape model and a corresponding training flow distribution, wherein the training vessel shape model is a data set representing a polygonal mesh and / or wherein the training vessel shape model comprises a volumetric mesh formed by 3D elements and wherein the corresponding training flow distribution comprises characteristics of blood flow at an inlet of a vessel; - receiving (67) an output training data set corresponding to the received training input data set, the output training data set comprising a corresponding at least one training hemodynamic parameter; - training (68) the AI unit (10) to predict at least one hemodynamic parameter (p) of the target vessel (20) using the received training vessel shape model and the corresponding training flow distribution and the received set of corresponding training outputs; k ).
38. The computer-implemented method of claim 37, wherein: The AI unit (10) is a recursive neural network (RNN).
39. The computer-implemented method of claim 38, wherein: The AI unit (10) is a long short-term memory (LSTM) network.
40. The computer-implemented method of claim 37, comprising the following initial steps: - determining (61) the shape variability and geometric variability of blood vessels within a given population by applying statistical shape analysis to a set of realistic vessel shape models of the blood vessels; - synthesizing (62) at least one synthetic vessel shape model of the vessels based on the determined shape variability and geometric variability of the vessels within a given population; - determining (63) a corresponding training flow distribution for the set of real vessel shape models and / or the at least one synthetic vessel shape model; - aggregating (64) the input training data set with the set of true shape models and with the determined corresponding training flow distribution; - calculating (65) a corresponding at least one training hemodynamic parameter of an output training set based on the set of real shape models and / or the at least one synthetic vessel shape model and a corresponding flow distribution of the input training data set.
41. The computer-implemented method of claim 40, wherein: The input training data set is aggregated (64) with the set of true shape models and the at least one synthetic shape model and with the determined corresponding training flow distribution.
42. A computer-implemented method according to any one of claims 37 to 41, wherein: Computational fluid dynamics (CFD) simulations are used to calculate the training hemodynamic parameters of the output training set.
43. A computer program comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 37 to 42.
44. A computer readable medium (70) having stored thereon a computer program according to claim 43.
45. A data processing system (80) comprising means (81, 82, 83, 84, 85) for performing the steps of the method according to any one of claims 37 to 42.
