X-ray image data reconstruction method, method for providing training model, processing device, X-ray device, computer program product and data carrier

Abstract

A computer-implemented method for reconstructing three-dimensional or four-dimensional X-ray image data (1), comprising the steps of: receiving a first set (2) of two-dimensional X-ray images (3) that respectively image at least a portion of an associated section (4) of a vascular system (5) of a patient (6); automatically determining a three-dimensional region of interest (7) of the patient (6) by processing the X-ray images (3) of the first set (2) by an analysis algorithm (8), the region comprising a section of interest (4) of the vasculature (5) of the patient (6); and reconstructing three-dimensional or four-dimensional X-ray image data (1) on the basis of the first group (2) of two-dimensional X-ray images (3) and / or the received second group (9) of X-ray images (10) of the patient (6) such that all voxels (11) of the X-ray image data (1) are located within the determined region of interest (7). Furthermore, the application also relates to a method for providing a training model, a processing device, an X-ray device, a computer program product and a data carrier.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for reconstructing three-dimensional or four-dimensional X-ray image data, a computer-implemented method for providing a model trained by machine learning, a processing device, an X-ray apparatus, a computer program product, and a data carrier. Background Technology

[0002] In three-dimensional or four-dimensional X-ray imaging, such as in subtraction angiography of the vascular system, in many applications, the achievable level of detail or the achievable resolution of the X-ray image data is not limited by the imaging of the underlying two-dimensional X-ray image itself, but by the voxel resolution of the reconstructed X-ray image data that is fixed in the usual reconstruction method.

[0003] Because high levels of detail are crucial for robustly identifying relevant features in X-ray image data, such as for diagnosing neurovascular diseases like arteriovenous malformations or aneurysms, multi-stage approaches have been commonly used to date. In this approach, after examining all relevant X-ray images, a provisional reconstruction is performed on the entire region that can be reconstructed based on the X-ray images. Based on this provisional reconstruction, medical personnel then select relevant regions, such as areas imaged that are relevant to the vascular system, and repeat the reconstruction on these relevant regions to achieve optimal resolution.

[0004] This process adds an extra workload for medical personnel and can be a potential source of error because, in some cases, the region relevant to subsequent automated image data processing may be located outside the manually segmented area. For example, in four-dimensional digital subtraction angiography, it is necessary that the entire contrast-enhanced portion of the vascular structure be located within the relevant region selected for reconstruction at each time step, because otherwise, X-ray attenuated images of vascular segments containing contrast and located outside the reconstruction area cannot be assigned to the correct vessel, resulting in distorted image values ​​in the reconstructed volume.

[0005] Furthermore, the computational cost of reconstructing three-dimensional or even four-dimensional X-ray image data is extremely high. Two complete reconstructions are required, with the user needing to evaluate the results of the first reconstruction between them. This necessitates high computing power to ensure smooth workflow for medical personnel. This results in additional costs for the imaging equipment used or another computing device for reconstruction (such as workstation computers, servers, or cloud solutions).

[0006] As an alternative to the aforementioned multi-stage method, a higher level of detail can theoretically be achieved by directly reconstructing higher-resolution X-ray image data. However, to adjust the resolution of the X-ray image data, the reconstruction algorithm typically requires significant modifications, such as adjusting the size of filter kernels or other matrices. Furthermore, the increased resolution leads to a disproportionately increased computational cost, resulting in higher hardware costs or longer processing times compared to the methods described above. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to provide an improved method for reconstructing X-ray image data involving the vascular system.

[0008] The present invention solves this problem by a computer-implemented method for reconstructing three-dimensional or four-dimensional X-ray image data, the method comprising the following steps:

[0009] - Receive a first set of two-dimensional X-ray images, which respectively image at least a portion of relevant segments of the patient's vascular system;

[0010] - By processing the first set of X-ray images with analytical algorithms, the patient's three-dimensional relevant regions, including relevant segments of the patient's vascular system, are automatically determined;

[0011] - Based on the first set of two-dimensional X-ray images and / or the second set of X-ray images received from the patient, reconstruct three-dimensional or four-dimensional X-ray image data such that all voxels of the X-ray image data are located within the determined relevant regions.

[0012] By automatically identifying the patient-related regions, medical personnel are no longer required to manually select these regions to achieve optimal resolution, thus reducing their workload and avoiding potential operational errors. As will be explained in more detail later, it has also been found that, with proper implementation of the analysis algorithm, the computational cost required is far less than the additional reconstruction with the initial lower spatial resolution required to manually find the relevant regions. Therefore, the method according to the invention not only reduces the burden on medical personnel but also significantly reduces the computational cost required to provide high-resolution reconstructions compared to the methods described above, thereby enabling the delivery of the same quality results faster with the same hardware cost, or the same quality results with lower hardware cost within the same available computation time.

[0013] The voxel resolution of three-dimensional or four-dimensional X-ray image data can be independent of the size of the determined correlated region in at least one or two spatial dimensions, or in all spatial dimensions. As explained in the introduction, typical reconstruction algorithms can predetermine a fixed resolution for the reconstructed X-ray image data, or for example, for a single layer of the reconstructed X-ray image data. Within the scope of automatic determination of the correlated region, different dimensions can be determined for all spatial dimensions, such as different dimensions for the height, width, and depth of the correlated region. However, it is also possible to require that the determined correlated region have the same dimensions for two or three spatial dimensions, i.e., a cuboid shape or a cube shape with a square base.

[0014] The number of X-ray images in the first set of X-ray images may be at least two or at least four fewer than the number of X-ray images used to reconstruct the three-dimensional or four-dimensional X-ray image data. Additionally or alternatively, the number of X-ray images in the first set of X-ray images may be at most six, exactly three, or exactly two.

[0015] Specifically, the reconstruction of three-dimensional or four-dimensional X-ray image data can be based on a first set and a second set of X-ray images, wherein the second set is at least as large as the first set, preferably at least three times the size of the first set. When reconstructing four-dimensional X-ray image data, the number of X-ray images in the first set of X-ray images can be less than the number of X-ray images used to reconstruct the four-dimensional X-ray image data by a predetermined coefficient, wherein the predetermined coefficient is at least twice or at least four times the number of resolved time steps in the four-dimensional X-ray image data.

[0016] During the research of this invention, it was discovered that identifying relevant areas of a patient using a suitable analysis algorithm requires only a small number of X-ray images in typical use cases, and particularly far fewer than the number of X-ray images required for high-quality reconstruction. Therefore, the analysis algorithm only needs to process a small amount of input data, resulting in storage requirements and computational costs typically used to execute the analysis algorithm that are far lower than those in cases where additional reconstruction is performed before selecting relevant areas, for example, manually or through automatic segmentation in three-dimensional space. Particularly preferably, the first set of X-ray images is thus processed by the analysis algorithm directly or after purely two-dimensional preprocessing in the method according to the invention. Reconstruction of three-dimensional or four-dimensional image data is preferably not performed within the analysis algorithm.

[0017] A model trained by machine learning can be used as an analytical algorithm or a sub-algorithm of an analytical algorithm. A "model trained by machine learning" can also be called a "trained function".

[0018] Typically, machine learning-trained models mimic human cognitive functions related to other people's minds. In particular, by training on training data, the trained model can adapt to new situations and recognize and extrapolate patterns. Therefore, after proper training, such a trained model can, by analyzing a small number of projected images (e.g., only front and side views of the patient or imaging area), quasi-intuitively identify the necessary boundaries or dimensions of relevant areas of the patient to fully map the relevant segments of the vascular system.

[0019] Typically, model parameters can be tuned through training. Specifically, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used. Additionally, representation learning (also known as feature learning) can be used. In particular, the parameters of the trained model can be iteratively adjusted through multiple training steps. Specifically, a predetermined cost function can be minimized during training. Especially when training neural networks, error feedback can be used for training.

[0020] Preferably, in the method according to the invention, a model trained through supervised learning is used, wherein the training dataset specifically includes a first set of X-ray images as input data for the model to be trained, and the relevant regions of the patient are defined as output data for processing the input data of the training dataset. The predetermined desired output data in the corresponding training dataset, and thus the output data produced by the model training, may, for example, be the coordinates and dimensions of the center point of the relevant region and / or the coordinates of the boundary surface of the relevant region.

[0021] Additional requirements for defined relevant regions or output data, such as requiring two or three spatial dimensions of the relevant region to be the same, can be considered as additional boundary conditions during training, for example, as an additional term in the cost function. However, it is often simpler to first determine temporary relevant regions using the trained model, and then adjust their size to satisfy the relevant boundary conditions. For example, if the width and depth of the relevant region should be the same, then for two sizes, the larger of the two sizes of the temporary relevant region can be chosen.

[0022] The output data of the corresponding training dataset, such as the boundaries or dimensions of the relevant regions, can be determined within the training preparation scope, for example, manually by medical personnel or by other algorithms, which may be more computationally and / or storage-intensive than analytical algorithms. For example, here, three-dimensional or four-dimensional X-ray data reconstruction can first be performed based on a first set of X-ray images and, preferably, other X-ray images taken on the same patient as these images. The output data for the corresponding training dataset can then be determined, for example, through manual or automatic segmentation of the three-dimensional or four-dimensional X-ray image data.

[0023] Models trained using machine learning can be based on, for example, neural networks, support vector machines (SVMs), decision trees and / or Bayesian networks, and / or k-means clustering and / or transformers, Q-learning, genetic algorithms and / or association rules. In particular, neural networks can be deep neural networks, convolutional neural networks (CNNs), or deep CNNs. Furthermore, neural networks can be adversarial networks, deep adversarial networks, and / or generative adversarial networks (GANs).

[0024] Preferably, the training of the machine learning-trained model is performed outside the claimed computer-implemented method. For example, the computer-implemented method may be performed within the scope of medical imaging use and / or during the analysis of image data from such imaging, e.g., at a medical center or general practitioner's office. The model training may be performed independently of this in space and time, and / or by others, e.g., by the manufacturer of the imaging device used within its manufacturing or development scope. However, in principle, training may also be included as an additional method step in the computer-implemented method according to the invention.

[0025] At least during the detection of the corresponding X-ray images of the second group, a collimator can be positioned between the X-ray source used to detect the corresponding X-ray images and the patient, wherein the collimator is set according to the patient's determined relevant area to detect the corresponding X-ray images of the second group.

[0026] Collimators used in X-ray imaging are typically apertures that shield the X-ray source from X-ray radiation outside the area through which it passes. For example, a one-dimensional collimator can be used, thereby limiting the width of the X-ray cone by two substantially parallel rectangular apertures, thus allowing the area of ​​the patient exposed to X-ray radiation to be positioned. Alternatively, variable apertures (German: Irisblende) or similar devices can be used as collimators, for example.

[0027] Since the relevant area can be determined based on the first set of X-ray images, and the areas in the second set of X-ray images that do not affect the reconstruction of the relevant area are not related to the reconstructed X-ray image data, this collimator setting can reduce the patient's radiation exposure in the X-ray imaging range.

[0028] To determine the second set of X-ray images, the collimator can preferably be set automatically, for example by an actuator assigned to the collimator, based on the determined patient-related region. However, in principle, the corresponding collimator settings can also be output to the user, who then manually sets the collimator according to these preset values. Since the second set of X-ray images is detected in the described case, the reconstruction of the three-dimensional or four-dimensional X-ray image data is then preferably based on either the first or second set of X-ray images.

[0029] The corresponding X-ray images of the second group can be determined by separately predetermined imaging geometry, wherein the collimator is configured to additionally detect the corresponding X-ray images of the second group according to the corresponding imaging geometry. In principle, the same collimator setting can be used for all second X-ray images, for example by setting the width of the slit defined by the collimator according to the maximum size of the determined relevant region, but individual settings for the corresponding X-ray images of the second group are advantageous because this usually allows for stronger collimation of at least a portion of these X-ray images, thereby further reducing the patient's radiation exposure.

[0030] The predetermined imaging geometry of the corresponding X-ray images in the second group can be used to determine the position of the collimator plane, in which the collimator acts as an aperture for X-ray radiation from the X-ray source, wherein, in order to determine the collimator setting for imaging the corresponding X-ray images, the relevant region is projected onto the collimator plane.

[0031] The projection is preferably based on the known beam geometry of the X-ray radiation emitted from the X-ray source. The collimator can be configured such that the region of the collimator plane that may be traversed by the X-ray radiation, located outside the region in which the projection is related, is covered as much as possible by the collimator assembly of the collimator, while the region in which the projection is related remains without the collimator assembly.

[0032] The first and / or second set of X-ray images can be detected specifically within the scope of digital subtraction angiography. Four-dimensional X-ray image data can be used to image the temporal changes in local contrast agent concentrations within relevant segments of the patient's vascular system.

[0033] The relevant area can be defined as the region that completely encompasses the vascular system within the patient's head or organs. Organs, for example, can be imaged, such as the patient's brain, liver, heart, or lungs.

[0034] The present invention also relates to a computer-implemented method for providing a machine learning-trained model for use as an analysis algorithm or a sub-algorithm of an analysis algorithm in a method for reconstructing three-dimensional or four-dimensional X-ray image data according to the present invention, the method comprising the following steps:

[0035] - Receive multiple training datasets, each comprising, on the one hand, multiple X-ray images of at least a portion of the relevant segment of the vascular system of the corresponding patient as input data, and on the other hand, a definition of a three-dimensional relevant region as the target result, in which the relevant segment of the patient's vascular system is located;

[0036] - Train the model based on the training dataset to determine the machine learning-trained model;

[0037] - Provides models trained using machine learning.

[0038] Supervised learning is achieved through the design of the training dataset. This can be done by starting with initial, for example, randomized parameterization of the model to be trained, iteratively adjusting the parameterization to minimize the deviation between the 3D relevant region determined based on X-ray images from the corresponding training dataset and the definition of that region predetermined as a target result in the training dataset. For example, the cost function can be, or include, the sum of measures of the difference between the size and / or coordinates of the region determined by the model and the size and / or coordinates of the region defined by the corresponding target result. This cost function can then be minimized within the error feedback range, for example, using gradient descent. Methods for providing a suitable training dataset have been described above.

[0039] The present invention also relates to a processing apparatus configured to perform a computer-implemented method according to the invention for reconstructing three-dimensional or four-dimensional X-ray image data and / or a computer-implemented method according to the invention for providing a machine learning-trained model.

[0040] The processing device can be configured, for example, as a suitably programmed data processing unit, or the aforementioned functions can be alternatively implemented at least partially via fixed wiring. The processing device can be integrated into a medical imaging apparatus, particularly an X-ray apparatus, or constructed separately from it. For example, it can be implemented as a workstation computer, server, or cloud solution.

[0041] The present invention also relates to an X-ray apparatus comprising an X-ray source and an X-ray detector for determining an X-ray image of a patient, and a processing device according to the invention. Integrating the processing device according to the invention into the X-ray apparatus can further simplify and accelerate the workflow of X-ray imaging.

[0042] The present invention also relates to a computer program comprising instructions configured to, when executed on a data processing apparatus, perform a computer-implemented method according to the invention for reconstructing three-dimensional or four-dimensional X-ray image data and / or a computer-implemented method according to the invention for providing a machine learning-trained model.

[0043] The present invention also relates to a data carrier comprising a computer program according to the present invention. Attached Figure Description

[0044] Other advantages and details of the invention will become apparent from the following embodiments and drawings, wherein:

[0045] Figure 1 An embodiment of an X-ray apparatus according to the invention is shown, which includes an embodiment of a processing device according to the invention;

[0046] Figure 2 A flowchart illustrating an embodiment of the method for reconstructing three-dimensional or four-dimensional X-ray image data according to the present invention;

[0047] Figure 3 Show Figure 2 Exemplary X-ray images of the first group in the method shown, and the temporary related regions determined therefrom;

[0048] Figure 4 A flowchart illustrating an embodiment of a method according to the present invention for providing a model trained by machine learning;

[0049] Figure 5 An example is shown that can be used according to Figure 2 and 4 The structure of the machine learning-trained model in the method. Detailed Implementation

[0050] Figure 1 An X-ray device 44 is shown for detecting X-ray images of a patient 6, wherein, in the illustrated example, the area of ​​the patient 6's head 19 is detected. X-ray images can be taken from different perspectives to detect the reconstruction of three-dimensional image data, such as for digital subtraction angiography, particularly four-dimensional digital subtraction angiography.

[0051] As explained in the general section of the instruction manual, the reconstruction of the three-dimensional or four-dimensional X-ray image data 1 is typically performed at a fixed voxel resolution, resulting in a decrease in achievable spatial resolution as the size of the imaging area increases. To ensure that the entire relevant segment 4 of the vascular system can be imaged, for example, the entire area filled with contrast agent in a digital subtraction angiography range, imaging is typically performed such that the patient's region 7, which contains the relevant segment 4 of the patient's vascular system 5, is larger than the patient's relevant region 7 in the typical reconstruction of the X-ray image data 1, thus resulting in lower spatial resolution.

[0052] To achieve optimal spatial resolution, the X-ray apparatus 44 therefore includes a processing device 39, which is configured to perform the following reference. Figure 2 A more detailed explanation of the computer-implemented method for reconstructing three-dimensional or four-dimensional X-ray image data is provided, wherein multiple (two in this example) X-ray images 3 are processed by an analysis algorithm 8 to automatically determine a three-dimensional relevant region 7 of the patient 6, which contains a relevant segment 4 of the patient 6's vascular system 5. The three-dimensional or four-dimensional X-ray image data 1 is then reconstructed based on these two-dimensional X-ray images 3, or particularly preferably, alternatively or additionally, based on a second set of 9 X-ray images 10 of the patient 6, such that all voxels 11 of the X-ray image data 1 lie within the determined relevant region 7.

[0053] The advantages of this method can be explained, for example, by imaging the vascular system in the head of patient 619. In the typical imaging geometry used for this purpose, a typical reconstruction might result in a reconstruction area size of 248 mm in each spatial direction. This achieves a spatial resolution of 0.24 mm at a resolution of 1024 voxels in each direction. If the dimensions of the relevant areas in the patient's width and depth directions are now, for example, 16.4 cm and 12 cm, then even using square voxels in this plane, the reconstruction area can be limited to a size of 16.4 cm, thus achieving a resolution of 0.16 mm, or a 33% higher resolution.

[0054] exist Figure 1 In the example shown, the processing device 39 is implemented by a programmable data processing device 24, whose processor 41 executes instructions of a computer program 25 that implements the method. The computer program 25 is stored on a data carrier 40.

[0055] An exemplary design of this method for reconstructing X-ray image data 1 is described below. Figure 2 The flowchart shown will be explained.

[0056] In step S1, the two-dimensional X-ray images 3 of the first group 2 are first detected. Alternatively, these X-ray images may be received from other devices or read from memory, for example.

[0057] like Figure 3 As illustrated, only two X-ray images 3 are considered in the example, however, they have a basically mutually perpendicular shooting geometry. Figure 3 The upper part of the X-ray image 3 shown here was produced by side-mounting, with C-arm 42 relative to... Figure 1 The position shown is rotated 90° out of the drawing plane. On the other hand, Figure 3 The lower part of the two X-ray images 3 shown was detected from a front-to-back perspective, i.e., through... Figure 1 The imaging geometry is shown in the figure.

[0058] Figure 3 X-ray image 3 is shown to better identify the vascular system 5, wherein the vascular system 5 of the head or right hemisphere of the brain is filled with contrast agent and is therefore clearly identifiable. If the method is used for reconstruction in four-dimensional digital subtraction angiography, it is advantageous in practice to use the X-ray film examined at the beginning of the X-ray sequence as the first set of X-ray images 3, since the vascular system 5 typically contains little or no contrast agent. This is appropriate because the relevant regions 7 determined based on these X-ray images 3 can then be used for collimation of subsequent X-ray films 10 of the sequence, as described below.

[0059] Steps S2 and S3 implement analysis algorithm 8. In this example, the X-ray image 3 in the first group 2 is first processed by sub-algorithm 12 of analysis algorithm 8 in step S2 to determine the temporary correlation region 26. The temporary correlation region is defined, for example, by the positions 27 to 32 of the boundary surfaces of the temporary correlation region 26 in different spatial directions, which... Figure 3 As shown in the image.

[0060] In principle, the vascular system 4 in the X-ray images of the first group 2 can be segmented to determine its or its temporary related regions 7 and 26. However, classical segmentation methods often cannot robustly and automatically segment the vascular system 5 in the X-ray images 3 of the first group 2, especially when the vascular system in the X-ray images 3 of the first group 2 still contains virtually no contrast agent. Therefore, in the example shown, a machine learning-trained model 13 is used as a sub-algorithm 12 to achieve robust segmentation even when the vascular system does not contain a contrast agent. A feasible method for training such a model 13 will be referred to below. Figure 4 explain.

[0061] like Figure 3 As shown, in the example, the temporary relevant region 26 results in different sizes 33 to 35 for different spatial orientations. Since, in the example, the square shape of the corresponding voxels of the reconstructed X-ray image data should be achieved in the plane opened by sizes 33 and 34, a boundary condition 36 is applied in step S3, according to which the smaller of sizes 33 and 34 is set to the larger of these sizes to determine the relevant region 26 for reconstruction. Depending on the specific implementation of the reconstruction and visualization of the reconstructed data, this step can potentially be omitted to potentially achieve a higher level of detail in the reconstructed X-ray image data.

[0062] In steps S4 to S7, additional X-ray images are then detected and assigned to the second group 9. These additional X-ray images 10 serve, on the one hand, to provide additional imaging geometry for high-quality reconstruction, and on the other hand, to detect image data at different times during the filling of the vascular system 5 with contrast agent, thereby performing four-dimensional digital subtraction angiography.

[0063] In step S4, the required recording geometry 16 is first set for the corresponding X-ray film 10, in particular by pivoting and / or rotating the C-arm 42.

[0064] Then, in step S5, or alternatively, even during the setting of the recording geometry 16, a suitable collimator setting 18 for the collimator 14 positioned between the X-ray source 15 and the patient 16 is determined, and the collimator 14 is set accordingly. For this purpose, the relevant region 7 determined in step S3 is projected onto the known beam geometry and imaging geometry 16 of the radiation source 15. Figure 1On the collimator plane 17 shown. Then, a rectangular aperture forming the collimator 14 in the example is set such that the surface of the relevant region 7 being imaged under this projection is not blocked, while the area outside the surface is blocked by the aperture as much as possible to minimize the radiation dose to the patient 6 in the imaging range.

[0065] In step S6, the X-ray source 15 is then activated for imaging so that the corresponding X-ray image 10 of the second group 9 is recorded by the X-ray detector 45. Then in step S7, it is checked whether all X-ray images 10 of the second group 9 have been detected. If not, the method is repeated from step S4 for the next X-ray image 10 to be detected in the second group 9.

[0066] After detecting all X-ray images 10 in the second group 9, X-ray image data 1 is reconstructed in step S8 such that all voxels 11 of X-ray image data 1 are located within the determined relevant region 7. Various methods are known for reconstructing three-dimensional image data of a volume based on a projected image that determines the volume from at least partial imaging, and these methods can be used in step S8. For example, filtered backprojection, iterative reconstruction methods, or reconstruction via a further machine learning-trained model can be performed.

[0067] The method described achieves at least near-optimal resolution for the reconstructed X-ray image data 1 without requiring complex user interaction and with less computational cost than the multiple reconstructions explained at the beginning of the general section of the specification. Furthermore, the automatic collimator setting reduces the X-ray dose delivered to the patient 6 for imaging compared to conventional imaging methods.

[0068] Figure 4 A flowchart is shown for a method of providing a machine learning-trained model 13, which can be used... Figure 2 In step S2 of the method shown, a temporary relevant region 26 is determined. In principle, this method can also be used... Figure 1 The processing device 39 shown performs the operation. However, it is generally preferable to train model 13 using a separate processing device (not shown), such as a server or cloud solution, especially since such training is typically performed using... Figure 2 The algorithm shown can be executed by different people. For example, training can be performed by... Figure 1 The manufacturer of the X-ray device 44 shown is responsible for its execution.

[0069] In step S9, multiple training datasets 20 are received, each comprising, on the one hand, multiple X-ray images 3 of at least a portion of a relevant segment 4 of the vascular system 5 of the corresponding patient 6, respectively, as input data 21; and on the other hand, a definition 22 of a three-dimensional relevant region, in which the relevant segment 4 of the vascular system 5 of the patient 6 is located, as the target result 23. Definition 22 is, in this example, defined by specifying the positions 27 to 32 of the boundary surfaces of the relevant region. These boundary surfaces can be determined, for example, by reconstructing three-dimensional X-ray image data based on a dataset of X-ray images from which the X-ray images 3 are derived, wherein the segmentation of the relevant region is then performed automatically or manually by medical personnel.

[0070] In step S10, model 13 is then applied to the X-ray image 3 of input data 21 using its initial or current parameterization 37 to determine the actual results 43 for the relevant regions. In the example, the actual results include the actual values ​​of the positions 27 to 32 of the boundary surfaces of the relevant regions, respectively.

[0071] In step S11, the cost function 38 can therefore be evaluated. This cost function is the sum of measures of the deviations between the corresponding positions 27 to 32 in the actual results 43 and the corresponding positions 27 to 31 in the target results 23. The parameterization 37 of the training function 13 can then be adjusted using the known error feedback. For example, gradient descent can be used to minimize the cost function 38.

[0072] The trained model 13 could be, for example, Figure 5 The convolutional neural network shown has layers L.1 through L.10, which exist twice, once in sub-network 40 and once in sub-network 41. Layer L.11 processes the output data from the corresponding layers L.10 in sub-networks 40 and 41, which are input data. Sub-networks 40 and 41, or their respective layers L.1, process one of the X-ray images 3 of the first group 2. Therefore, in the example shown, the X-ray image 3 is first preprocessed separately in layers L.1 through L.10, and then layers L.11 through L.13 jointly process the resulting intermediate results.

[0073] The neural network in this example consists of convolutional layers, pooling layers, and fully connected layers. In the input layer L.1, there is a node for each pixel of the corresponding input data, where each pixel has a channel (the corresponding intensity value). Following the input layer are four convolutional layers L.2, L.4, L.6, and L.8, each followed by a pooling layer L.3, L.5, L.7, and L.9. For each convolutional layer, a 5x5 kernel (indicated by "5x5 kernel") is used, padded to 2 (indicated by "P: 2"), and the number of filters / convolutional kernels increases (indicated by "F: 2", "F: 4", or "F: 8"). Additionally, there are four pooling layers L.3, L.5, L.7, and L.9, where the first three layers L.3, L.5, and L.7 determine the average value over a 4x4 domain, and the last pooling layer L.9 performs maximum selection over a 2x2 domain. Figure 5 The diagram shows an additional layer L.10, which flattens the input image (i.e., combines eight 4x4 images into a vector with 128 entries). However, this layer is irrelevant to the actual computation.

[0074] The final layers of the network are three fully connected layers L.11, L.12, and L.13. The first fully connected layer has 256 input nodes and 64 output nodes, the second fully connected layer L.12 has 64 input nodes and 24 output nodes, and the third fully connected layer L.13 has 24 input nodes and 6 output nodes. Each output node provides one of the positions 27 to 32 of the temporary related region 26 defined in the example.

Claims

1. A computer-implemented method for reconstructing three- or four-dimensional X-ray image data (1), comprising the steps of: - receiving a first set (2) of two-dimensional X-ray images (3) which each image at least a portion of a relevant section (4) of a vascular system (5) of a patient (6); - automatically determining a three-dimensional relevant region (7) of the patient (6) by processing the X-ray images (3) of the first set (2) by an analysis algorithm (8), the three-dimensional relevant region comprising the relevant section (4) of the vascular system (5) of the patient (6); - reconstructing the three- or four-dimensional X-ray image data (1) based on the first set (2) of two-dimensional X-ray images (3) and / or a second set (9) of X-ray images (10) of the patient (6) received, such that all voxels (11) of the X-ray image data (1) are located within the determined three-dimensional relevant region (7).

2. The computer-implemented method of claim 1, wherein, The number of X-ray images (3) in the first set (2) of X-ray images (3) is at least a factor of two or at least a factor of four less than the number of X-ray images (3, 10) used for reconstructing the three- or four-dimensional X-ray image data (1), and / or the number of X-ray images (3) in the first set (2) of X-ray images (3) is at most six or exactly three or exactly two.

3. The computer-implemented method of claim 1, wherein, A model (13) trained by machine learning is used as the analysis algorithm (8) or as a sub-algorithm (12) of the analysis algorithm (8).

4. The method according to one of the preceding claims, characterized in that At least during the detection of the respective X-ray images (10) of the second set (9), a collimator (14) is arranged between an X-ray source (15) for detecting the respective X-ray images (10) and the patient (6), wherein the collimator (14) is set in accordance with the determined relevant region (7) of the patient (6) for detecting the respective X-ray images (10) of the second set (9).

5. The method of claim 4, wherein, The respective X-ray images (10) of the second set (9) are determined by a respectively predetermined acquisition geometry (16), wherein the collimator (14) is set for detecting the respective X-ray images (10) of the second set (9) additionally in accordance with the respective acquisition geometry (16).

6. The method of claim 5, wherein, The predetermined acquisition geometry (16) of the respective X-ray images (10) of the second set (9) prescribes a position of a collimator plane (17) in which the collimator (14) acts as an aperture for the X-ray radiation of the X-ray source (15), wherein, for determining a collimator setting (18) of the collimator (14) for acquiring the respective X-ray images (10), the relevant region (7) is projected into the collimator plane (17).

7. The method according to one of the preceding claims, characterized in that The X-ray images (3, 10) of the first and / or second set (2, 9) are detected in the context of digital subtraction angiography.

8. The method of claim 7, wherein, The four-dimensional X-ray image data (1) images a temporal change of a local contrast agent concentration in the relevant section (4) of the vascular system (5) of the patient (6).

9. The method according to one of the preceding claims, characterized in that The relevant region (7) is determined as a region which completely contains the head (19) of the patient (6) or a vascular system (5) within an organ. The relevant region (7) is determined as a region which completely contains the head (19) of the patient (6) or a vascular system (5) within an organ.

10. Computer-implemented method for providing a model (13) trained by machine learning for use as an analysis algorithm (8) or a sub-algorithm (12) of an analysis algorithm (8) in a method according to one of the preceding claims, the method comprising the steps of: - receiving a plurality of training data sets (20) each comprising on the one hand a plurality of X-ray images (3) of a respective patient (6) imaging at least a portion of a relevant section (4) of the vascular system (5) of the respective patient (6) as input data (21) and on the other hand a definition (22) of a three-dimensional relevant region (7) within which the relevant section (4) of the vascular system (5) of the patient (6) is located as target result (23), - training the model (13) based on the training data sets (20) to determine a model (13) trained by machine learning, - providing the model (13) trained by machine learning.

11. A processing device, characterized by which is configured to perform a computer-implemented method according to one of the preceding claims.

12. An X-ray apparatus comprising an X-ray source (15) and an X-ray detector (45) for determining an X-ray image of a patient (6), characterized in that, The X-ray apparatus (44) comprises a processing device (39) according to claim 11.

13. Computer program product comprising instructions configured to perform a computer-implemented method according to one of claims 1 to 10 when executed on a data processing device (24).

14. Data carrier comprising a computer program (25) according to claim 13.

Images