Method for reconstructing X-ray image data, method for providing a trained model, processing device, X-ray equipment, computer program and data carrier

Abstract

Computer-implemented method for the reconstruction of three-dimensional or four-dimensional X-ray image data (1), comprising the steps: - Receiving a first group (2) of two-dimensional X-ray images (3), each depicting at least a part of a relevant section (4) of a vascular system (5) of a patient (6), - automatic determination of a three-dimensional relevant region (7) of the patient (6) that includes the relevant section (4) of the vascular system (5) of the patient (6) by processing the X-ray images (3) of the first group (2) by an analysis algorithm (8), - Reconstruction of the three-dimensional or four-dimensional X-ray image data (1) based on the first group (2) of the two-dimensional X-ray images (3) and / or a 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 identified relevant region (7).

Description

[0001] The 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 machine-learning trained model, a processing device, an X-ray device, a computer program and a data carrier.

[0002] In three-dimensional or four-dimensional X-ray imaging, for example in subtraction angiography of a vascular system, the achievable level of detail or resolution of the X-ray image data is in many applications not limited by the imaging of the underlying two-dimensional X-ray images themselves, but by the voxel resolution of the reconstructed X-ray image data, which is fixed in conventional reconstruction approaches.

[0003] Since a high level of detail is essential for the robust detection of relevant features in X-ray image data, for example, for the diagnosis of neurovascular pathologies such as arteriovenous malformations or aneurysms, a multi-stage approach has been standard practice. This involves acquiring all relevant X-ray images and performing a preliminary reconstruction of the entire area reconstructible from the X-ray images. Based on this preliminary reconstruction, medical professionals then select a relevant region, such as one representing the relevant part of the vascular system. The reconstruction is then repeated for this relevant area to achieve optimal resolution.

[0004] This approach leads to an additional workload for medical professionals and can be a potential source of error, as areas relevant for subsequent automated image data processing may lie 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 lies within the relevant region selected for reconstruction at each time step. Otherwise, an X-ray attenuation image caused by vascular segments containing contrast medium that lie outside the reconstructed area cannot be correctly assigned to the appropriate vessel, thus distorting the image values ​​in the reconstructed volume.

[0005] Secondly, reconstructing three-dimensional or even four-dimensional X-ray image data is computationally intensive. The requirement to perform two complete reconstructions, with user review of the first reconstruction's result between them, necessitates high computing power to ensure smooth operation for medical professionals. This results in additional costs for the imaging equipment used or other computing resources, such as a workstation, server, or cloud solution used for the reconstruction.

[0006] As an alternative to the multi-stage approach described above, higher levels of detail could, in principle, be achieved by directly reconstructing higher-resolution X-ray image data. However, adjusting the resolution of the X-ray image data typically requires significant modification of the reconstruction algorithm, for example, by adjusting the sizes of filter kernels or other matrices. Furthermore, increasing the resolution leads to a disproportionate increase in computational effort, resulting in even higher hardware requirements and / or longer processing times compared to the approach described above.

[0007] The invention is therefore based on the objective of providing an improved approach to the reconstruction of X-ray image data relating to a vascular system.

[0008] The object is solved according to the invention by a computer-implemented method for reconstructing three-dimensional or four-dimensional X-ray image data, which comprises the following steps: - Receiving an initial set of two-dimensional X-ray images, each depicting at least part of a relevant section of a patient's vascular system, - Automatic identification of a three-dimensional relevant region of the patient, encompassing the relevant section of the patient's vascular system, by processing the X-ray images of the first group using an analysis algorithm, - Reconstruction of the three-dimensional or four-dimensional X-ray image data based on the first group of two-dimensional X-ray images and / or a received second group of X-ray images of the patient, such that all voxels of the X-ray image data are located within the identified relevant region.

[0009] The automatic identification of the patient's relevant region eliminates the need for manual selection by medical professionals to achieve optimal resolution within that region. This reduces the workload for medical staff and prevents potential operator errors. Furthermore, as will be explained in more detail later, it has been found that with a suitable implementation of the analysis algorithm, it can require significantly less computational effort than the additional reconstruction required for manually identifying the relevant region, which initially results in lower spatial resolution.Thus, the inventive method not only relieves the burden on medical personnel, but also significantly reduces the computational effort required to provide the high-resolution reconstruction compared to the method described above, so that a result of the same quality can be provided faster with the same hardware effort or with less hardware effort given the same available computing time.

[0010] The voxel resolution of the three-dimensional or four-dimensional X-ray image data can be independent of the dimensions of the identified relevant region, at least in one or two of the spatial dimensions, or in all spatial dimensions. As explained in the introduction, common reconstruction algorithms can specify a fixed resolution for the reconstructed X-ray image data or, for example, for individual planes of the reconstructed X-ray image data. It is possible that, during the automatic determination of the relevant region, different dimensions are determined for all spatial dimensions, such as height, width, and depth. However, it is also possible that the identified relevant region is required to have the same dimensions for two or three of the spatial dimensions, i.e., the shape of a cuboid with a square base or a cube.

[0011] The number of X-ray images in the first group can be at least two or at least four times lower than the number of X-ray images used to reconstruct the three-dimensional or four-dimensional X-ray image data. Alternatively, or in addition, the number of X-ray images in the first group can be a maximum of six, exactly three, or exactly two.

[0012] In particular, the reconstruction of three-dimensional or four-dimensional X-ray image data can be based on the first and second groups of X-ray images, wherein the second group is at least as large, and preferably at least three times as large, as the first group. When reconstructing four-dimensional X-ray image data, the number of X-ray images in the first group can be smaller by a predetermined factor than the number of X-ray images used to reconstruct the four-dimensional X-ray image data, wherein the predetermined factor is at least twice as large or at least four times as large as the number of resolved time steps in the four-dimensional X-ray image data.

[0013] During the development of the invention, it was recognized that by using a suitable analysis algorithm to identify the relevant region of the patient, only a few X-ray images are required in typical applications, and in particular, significantly fewer X-ray images are needed than for a high-quality reconstruction. Thus, the analysis algorithm only needs to process a small amount of input data, resulting in significantly lower storage requirements and, typically, lower computational effort for running the analysis algorithm compared to performing an additional reconstruction after which the relevant region could be selected, for example, manually or through automatic segmentation in three-dimensional space.The X-ray images of the first group are thus preferably processed in the inventive method directly or after purely two-dimensional preprocessing by the analysis algorithm. Preferably, no reconstruction of three-dimensional or four-dimensional image data takes place within the analysis algorithm.

[0014] A machine learning-trained model can be used as an analysis algorithm or as a sub-algorithm of the analysis algorithm. A "machine learning-trained model" can also be referred to as a "trained function".

[0015] In general, a machine learning-trained model mimics cognitive functions that humans associate with the minds of others. Specifically, through training based on training data, the trained model is able to adapt to new circumstances and recognize and extrapolate patterns. Thus, after appropriate training, such a model can, almost intuitively, recognize the boundaries or dimensions that the relevant region of the patient must have to fully depict the relevant section of the vascular system, simply by evaluating just a few projection images, for example, a frontal and a lateral view of the patient or the area being imaged.

[0016] In general, model parameters can be adjusted through training. Specifically, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be employed. Furthermore, representational 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 predefined cost function can be minimized during training. Error feedback can be used for training, especially when training a neural network.

[0017] Preferably, the method according to the invention uses a model trained by supervised learning, wherein the training datasets used comprise, in particular, the X-ray images of the first group as input data for the model to be trained and describe the relevant region of the patient as the desired output data for processing the input data of the training dataset. The desired output data specified in the respective training dataset, and thus the output data for which the model is trained, could, for example, be the coordinates of the center point of the relevant region and its dimensions and / or the coordinates of the boundary surface of the relevant region.

[0018] Additional requirements for the identified relevant region or the input data, such as the requirement that two or three of the spatial dimensions of the relevant region are equal, 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 a preliminary relevant region using the trained model, the dimensions of which can then be adjusted to satisfy the corresponding boundary conditions. For example, if the width and depth of the relevant region are to be equal, the larger of these two dimensions of the preliminary relevant region can be chosen for both dimensions.

[0019] The initial data for each training dataset, such as the boundaries or dimensions of the relevant region, can be determined during training preparation, for example, manually by medical professionals or by another algorithm that may be more computationally and / or memory-intensive than the analysis algorithm. For instance, this might involve first reconstructing three-dimensional or four-dimensional X-ray data based on the X-ray images from the first group and preferably further X-ray images taken of the same patient. The initial data for each training dataset can then be determined, for example, by manually or automatically segmenting this three- or four-dimensional X-ray image data.

[0020] A machine learning-trained model can be based on, for example, a neural network, a support vector machine (SVM), a decision tree, a Bayesian network, k-means clustering, a transformer, Q-learning, genetic algorithms, and / or assignment rules. Specifically, a neural network can be a deep neural network, a convolutional neural network (CNN), or a deep CNN. Furthermore, the neural network can be an adversarial network, a deep adversarial network, and / or a generative adversarial network (GAN).

[0021] Preferably, the training of the machine learning-trained model takes place outside of the claimed computer-implemented method. For example, the computer-implemented method can be carried out in the context of the application of medical imaging and / or the evaluation of image data from such imaging, for example, in a clinic or at a physician's office. The training of the model can be carried out independently of this in terms of location and time and / or by other persons, for example, by a manufacturer of the imaging device used, during its manufacture or development. However, it would also be possible, in principle, to include the training as an additional process step in the computer-implemented method according to the invention.

[0022] At least during the acquisition of the respective X-ray image of the second group, a collimator can be arranged between an X-ray source used to acquire the respective X-ray image and the patient, whereby the collimator is adjusted for acquiring the respective X-ray image of the second group depending on the identified relevant region of the patient.

[0023] A collimator used in X-ray imaging is typically a diaphragm that blocks the X-rays from the X-ray source outside a certain transmission range. For example, a one-dimensional collimator can be used, in which the width of an X-ray cone is limited by two essentially parallel rectangular apertures, thus controlling the area of ​​the patient illuminated by X-rays. Alternatively, an iris diaphragm or similar device can be used as a collimator.

[0024] Since the relevant region can already be determined from the X-ray images of the first group, and areas of the X-ray images of the second group that do not affect the reconstruction of the relevant region have no relevance for the reconstructed X-ray image data, the radiation exposure of the patient during X-ray imaging can be reduced by such a collimator setting.

[0025] The collimator can preferably be automatically adjusted to acquire the second group of X-ray images, for example, by an actuator assigned to the collimator, depending on the identified relevant region of the patient. In principle, however, it would also be possible to output the respective collimator setting to a user who would then manually adjust the collimator accordingly. Since, in the described case, X-ray images of the second group are acquired, the reconstruction of the three-dimensional or four-dimensional X-ray image data is then preferably based on the first and second groups of X-ray images or on the second group of X-ray images alone.

[0026] Each X-ray image in the second group can be acquired using a predefined acquisition geometry, with the collimator being adjusted accordingly depending on this geometry. While it would theoretically be possible to use the same collimator setting for all second X-ray images, for example, by adjusting the width of a gap defined by the collimator based on the maximum dimensions of the relevant region, a separate setting for each image in the second group is advantageous. This allows for stronger collimation and thus a further reduction in the patient's radiation exposure for at least some of these images.

[0027] The specified imaging geometry of the respective X-ray image of the second group can determine the position of a collimator plane in which the collimator acts as an aperture for the X-ray radiation of the X-ray source, whereby the relevant region is projected into the collimator plane to determine the collimator setting for the imaging of the respective X-ray image.

[0028] The projection preferably follows a known beam geometry of the X-rays emitted by the X-ray source. The collimator can be adjusted such that areas of the collimator plane potentially irradiated by the X-rays, which lie outside the area into which the relevant region is projected, are, if possible, obscured by a collimator component of the collimator, while areas into which the relevant region is projected remain free of collimator components.

[0029] The X-ray images of the first and / or second group can be acquired, in particular, during digital subtraction angiography. The four-dimensional X-ray image data can then depict the temporal course of a local contrast agent concentration in the relevant section of the patient's vascular system.

[0030] A relevant region can be identified as an area that completely encompasses the vascular system within the patient's head or an organ. For example, the brain, liver, heart, or lungs could be imaged as the organ.

[0031] The invention also relates to a computer-implemented method for providing a machine learning-trained model for use as an analysis algorithm or as a sub-algorithm of the analysis algorithm in the inventive method for reconstructing three-dimensional or four-dimensional X-ray image data, comprising the following steps: - Receiving multiple training datasets, each comprising, on the one hand, multiple X-ray images of a respective patient, each depicting at least part of a relevant section of a vascular system of the respective patient, as input data, and, on the other hand, a definition of a three-dimensional relevant region within which the relevant section of the patient's vascular system is located, as the target result, - Training a model based on the training datasets to determine the machine learning-trained model, - Provision of the machine learning-trained model.

[0032] The described design of the training datasets enables supervised learning. Starting with an initial, for example, random parameterization of the model to be trained, the parameterization can be iteratively adjusted to minimize the deviation of a relevant three-dimensional region, determined based on the X-ray images of the respective training dataset, from the definition of this region specified as the target result in the training dataset. For example, a cost function can be or include a sum of measures for differences between the dimensions and / or coordinates of the region determined by the model and the dimensions and / or coordinates of the region defined by the respective target result. This cost function can then be minimized by varying the model parameters within the framework of error feedback, for example, using a gradient descent method.Approaches to providing suitable training datasets have already been explained above.

[0033] The invention also relates to a processing device that is set up to carry out the computer-implemented method according to the invention for the reconstruction of three-dimensional or four-dimensional X-ray image data and / or the computer-implemented method according to the invention for providing a machine learning-trained model.

[0034] The processing device can, for example, be designed as suitably programmed data processing equipment, or the aforementioned functionality can alternatively be implemented, at least partially, via hardwiring. The processing device can be integrated into a medical imaging device, particularly an X-ray device, or it can be separate from it. It can be implemented, for example, as a workstation, server, or cloud solution.

[0035] The invention also relates to an X-ray device comprising an X-ray source and an X-ray detector for obtaining X-ray images of a patient, as well as a processing device according to the invention. Integrating the processing device according to the invention into the X-ray device can further simplify and accelerate the workflow for X-ray imaging.

[0036] The invention also relates to a computer program with instructions configured to carry out, when executed on a data processing device, the computer-implemented method according to the invention for reconstructing three-dimensional or four-dimensional X-ray image data and / or the computer-implemented method according to the invention for providing a machine learning-trained model.

[0037] Furthermore, the invention relates to a data carrier that includes the computer program according to the invention.

[0038] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

[0039] Further advantages and details of the invention will become apparent from the following exemplary embodiments and the accompanying drawings. These schematically illustrate: Fig. 1 an embodiment of the X-ray device according to the invention, comprising an embodiment of the processing device according to the invention, Fig. 2 a flowchart of an embodiment of the inventive method for reconstructing three-dimensional or four-dimensional X-ray image data, Fig. 3 exemplary X-ray images of the first group in the in Fig. 2 methods shown and a preliminary relevant region determined therefrom, Fig. 4 a flowchart of an embodiment of the inventive method for providing a machine learning-trained model, and Fig. 5 the structure of an exemplary procedure according to Fig. 2 and Fig. 4 usable models trainable through machine learning.

[0040] Fig. Figure 1 shows an X-ray device 44 for acquiring X-ray images of a patient 6, wherein in the example shown, the area of ​​the head 19 of the patient 6 is captured. X-ray images can be taken from different perspectives in order to reconstruct three-dimensional image data, for example for digital subtraction angiography, in particular for four-dimensional digital subtraction angiography.

[0041] As already explained in the general section of the description, the reconstruction of the three-dimensional or four-dimensional X-ray image data 1 is typically performed with a fixed voxel resolution, so that the achievable spatial resolution decreases with increasing size of the imaged area. To ensure that the entire relevant section 4 of the vascular system, for example, the entire area flooded with contrast medium during digital subtraction angiography, can be imaged, the imaging is typically performed such that a conventional reconstruction of the X-ray image data 1 would image a region of the patient that is larger than a relevant region 7 of the patient 6, which encompasses the relevant section 4 of the vascular system 5 of the patient 6, thus achieving a lower spatial resolution.

[0042] To achieve optimal spatial resolution, the X-ray device 44 therefore includes a processing device 39, which is used to perform a procedure described below with reference to Fig. 2. A computer-implemented method for reconstructing three-dimensional or four-dimensional X-ray image data, which is explained in more detail below, is set up in which several X-ray images 3, in this example two, are processed by an analysis algorithm 8 to automatically determine the three-dimensional relevant region 7 of the patient 6, which comprises the relevant section 4 of the vascular system 5 of the patient 6. The reconstruction of the three-dimensional or four-dimensional X-ray image data 1 is then carried out on the basis of these two-dimensional X-ray images 3 or, preferably, additionally or alternatively, on the basis of a 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 relevant region 7.

[0043] The advantages of this approach are illustrated using the example of imaging the vascular system in the head of patient 6. With a typical imaging geometry used for this purpose, a standard reconstruction might result in a dimension of 248 mm in each spatial direction for the reconstructed area. At a resolution of 1024 voxels in each direction, a spatial resolution of 0.24 mm is achieved. However, if the dimension of the relevant region in the width and depth of the patient is, for example, 16.4 cm and 12 cm respectively, even if square voxels are to be used in this plane, the reconstructed area can be restricted to dimensions of 16.4 cm, thus achieving a resolution of 0.16 mm, and therefore a 33% higher resolution.

[0044] The processing device 39 is located in the Fig. The example shown in Figure 1 is implemented by a programmable data processing device 24, whose processor 41 executes the instructions of a computer program 25 implementing the procedure. The computer program 25 is stored on a data carrier 40.

[0045] An exemplary embodiment of such a method for reconstructing X-ray image data 1 is described below with additional reference to the one in Fig. The flowchart shown in section 2 is explained.

[0046] In step S1, two-dimensional X-ray images 3 of a first group 2 are initially acquired. Alternatively, these X-ray images could, for example, be received from another device or read from a memory.

[0047] As in Fig. Figure 3 is shown schematically; in this example, only two X-ray images are considered, which, however, have essentially perpendicular imaging geometries. The upper of the two in Figure 3 is shown schematically. Fig. In the three X-ray images shown, image 3 was taken from a lateral view, while the C-arm 42 was positioned opposite the one shown in Fig. The position shown in 1 was rotated 90° out of the image plane. The lower of the two in Fig. In contrast, image 3 of the three X-ray images shown is taken in an anterior-posterior perspective, i.e., with the view shown in Fig. 1. Imaging geometry shown.

[0048] In Fig. Figure 3 shows X-ray images 5 for better visualization of the vascular system, in which the vascular system 5 of the right head or brain hemisphere is flooded with contrast medium and thus clearly visible. If the described procedure is used to reconstruct a four-dimensional digital subtraction angiography, it can be advantageous in actual application to use X-ray images 3 from the first group, which are taken at the beginning of the X-ray sequence, when typically little or no contrast medium is present in the vascular system 5. This is useful because the relevant area 7 determined on the basis of these X-ray images 3, as will be explained later, can then be used in the collimation of the subsequent X-ray images 10 of the sequence.

[0049] Steps S2 and S3 implement the analysis algorithm 8. In this process, the X-ray images 3 of the first group 2 in the example are first processed in step S2 by a sub-algorithm 12 of the analysis algorithm 8 to determine a preliminary relevant region 26. The preliminary relevant region is described, for example, by the positions 27-32 of interfaces of the preliminary relevant region 26 in the different spatial directions, which are shown in Fig. The three shown are...

[0050] In principle, it would be possible to identify the preliminary relevant region 7, 26 based on a segmentation of the vascular system 4 in the X-ray images of the first group 2.

[0051] However, automatic segmentation of the vascular system 5 in the X-ray images 3 of the first group 2 is often not robustly possible with classical segmentation approaches, especially when the vascular system in the X-ray images 3 of the first group 2 is still largely free of contrast agent. Therefore, in the example shown, a machine learning-trained model 13 is used as the sub-algorithm 12 to achieve robust segmentation even in a contrast-free vascular system. A method for training such a model 13 will be described later with reference to Fig. 4 will be explained.

[0052] As in Fig. As can be seen in Figure 3, different dimensions result for the various spatial directions in the example, and dimensions 33-35 of the preliminary relevant region 26 are determined. Since the example aims for a square shape for each voxel of the reconstructed X-ray image data in the plane spanned by dimensions 33 and 34, a boundary condition 36 is applied in step S3. This condition sets the smaller of the two dimensions (33 and 34) to the value of the larger dimension (33 and 34) to determine the relevant region 26 used for reconstruction. Depending on the specific implementation of the reconstruction and the visualization of the reconstructed data, this step can potentially be omitted to achieve a higher level of detail in the reconstructed X-ray image data.

[0053] In steps S4-S7, further X-ray images are acquired, which are assigned to a second group 9. These additional X-ray images 10 serve, on the one hand, to provide additional imaging geometries for high-quality reconstruction, and on the other hand, to acquire image data at different times during the influx of contrast medium into the vascular system 5, thus enabling four-dimensional digital subtraction angiography.

[0054] In step S4, the desired imaging geometry 16 for the respective X-ray image 10 is first set, in particular by swiveling and / or rotating the C-arm 42.

[0055] In step S5, or alternatively during the setting of the imaging geometry 16, a suitable collimator setting 18 of a collimator 14 arranged 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 transferred to the known beam geometry of the radiation source 15 and the imaging geometry 16. Fig. The collimator plane 17 shown is projected. The rectangular apertures, which in this example form the collimator 14, are then adjusted so that the area on which the relevant region 7 is imaged during this projection remains clear, while areas outside this area are covered by the apertures as far as possible in order to minimize the radiation dose emitted onto the patient 6 during the imaging process.

[0056] In step S6, the X-ray source 15 is activated for imaging in order to acquire each X-ray image 10 of the second group 9 using the X-ray detector 45. Subsequently, in step S7, it is checked whether all X-ray images 10 of the second group 9 have already been acquired. If this is not the case, the procedure is repeated from step S4 for the next X-ray image 10 of the second group 9 to be acquired.

[0057] After acquiring all X-ray images 10 of the second group 9, the X-ray image data 1 are reconstructed in step S8 such that all voxels 11 of the X-ray image data 1 are located within the identified relevant region 7. For reconstructing the three-dimensional image data of a specific volume based on projection images that at least partially depict this volume, a variety of approaches are known that can be used in step S8. For example, a filtered backprojection, an iterative reconstruction method, or a reconstruction using another machine learning-trained model can be performed.

[0058] The described procedure achieves, on the one hand, an at least approximately optimal resolution of the reconstructed X-ray image data 1 without requiring complex user interactions and with less computational effort than would be necessary for the multiple reconstruction described at the beginning of the general section of the description. Furthermore, the automatic collimator setting described reduces the X-ray dose delivered to the patient 6 for imaging compared to conventional imaging approaches.

[0059] Fig. Figure 4 shows a flowchart of a procedure for providing a machine learning-trained model 13, which is used in step S2 of the procedure described in Figure 4. Fig. The procedure shown in section 2 can be used to determine the preliminary relevant region 26. In principle, the procedure can also be used with the method described in Fig. The processing device shown in Figure 1 is used. However, it is typically advantageous to use a separate processing device (not shown), such as a server or cloud solution, for training the model 13, among other reasons because this training is typically performed by different people than the application of the device shown in Figure 1. Fig. The algorithm shown in section 2. For example, the training by a manufacturer of the [product / service] can be [implemented / implemented / etc.] Fig. X-ray equipment 44 shown in 1.

[0060] In step S9, several training datasets 20 are received, each comprising, on the one hand, several X-ray images 3 of a respective patient 6, each depicting at least a part of a relevant section 4 of a 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 within which the relevant section 4 of the vascular system 5 of the patient 6 is located, as the target result 23. In this example, the definition 22 is provided by specifying the positions 27-32 of interfaces of the relevant region. These interfaces 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 originate, in which the relevant region is then segmented automatically or manually by medical personnel.

[0061] In step S10, the model 13, with its initial or current parameterization 37, is applied to the X-ray images 3 of the input data 21 to determine an actual result 43 for the relevant region. In the example, the actual result comprises actual values ​​for positions 27-32 of interfaces of the relevant region.

[0062] In step S11, a cost function 38 can be evaluated, which is a sum of measures for the deviation of the respective position 27-32 in the actual result 43 from the corresponding position 27-31 in the target result 23. The parameterization 37 of the trained function 13 can then be adjusted using a known error feedback method. For example, a gradient descent method can be used to minimize the cost function 38.

[0063] For example, the trained model 13 can be the one in Fig. The convolutional neural network shown in Figure 5 is used. Layers L.1 to L.10 are duplicated, once in the depicted subnetwork 40 and again in subnetwork 41. Layer L.11 processes the output data of the respective layer L.10 of both subnetwork 40 and subnetwork 41 as input data. Subnetworks 40 and 41, or rather their respective layer L.1, each process one of the X-ray images 3 from the first group 2. Thus, in the example shown, the X-ray images 3 are first preprocessed separately in layers L.1-L.10, after which layers L.11-L.13 perform joint processing of the resulting intermediate results.

[0064] In this example, the neural network consists of convolutional layers, pooling layers, and fully connected layers. In the input layer L.1, there is one node for each pixel of the respective input data, with each pixel having one channel (representing its intensity value). Following the input layer are four convolutional layers: L.2, L.4, L.6, and L.8. Each of these four convolutional layers is followed by a pooling layer: L.3, L.5, L.7, and L.9. Each convolutional layer uses a 5x5 kernel (indicated by "5x5 Kernel") with a padding of 2 (indicated by "P: 2") and an increasing number of filters / convolutional kernels (indicated by "F: 2", "F: 4", or "F: 8"). Furthermore, there are four pooling layers L.3, L.5, L.7, L.9, where the first three layers L.3, L.5, L.7 perform averaging over 4x4 fields and the last pooling layer L.9 performs maximum selection over 2x2 fields. Fig.Figure 5 shows an additional layer L.10 that flattens the input images (i.e., the 8 images of size 4x4 are combined into a vector with 128 entries). However, this layer is not relevant for the actual calculation.

[0065] The last layers of the network are three fully connected layers L.11, L.12, L.13, wherein the first fully connected layer has 256 input and 64 output nodes, the second fully connected layer L.12 has 64 input and 24 output nodes, and the third fully connected layer L.13 has 24 input and 6 output nodes, each of the output nodes providing one of the positions 27-32, which in the example define the preliminary relevant region 26.

Claims

[1] Computer-implemented method for the reconstruction of three-dimensional or four-dimensional X-ray image data (1), comprising the steps: - Receiving a first group (2) of two-dimensional X-ray images (3), each depicting at least a part of a relevant section (4) of a vascular system (5) of a patient (6), - automatic determination of a three-dimensional relevant region (7) of the patient (6) that includes the relevant section (4) of the vascular system (5) of the patient (6) by processing the X-ray images (3) of the first group (2) by an analysis algorithm (8), - Reconstruction of the three-dimensional or four-dimensional X-ray image data (1) based on the first group (2) of the two-dimensional X-ray images (3) and / or a 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 identified relevant region (7). [2] Computer-implemented method according to claim 1, characterized by , that the number of X-ray images (3) in the first group (2) of the X-ray images (3) is at least two times less or at least four times less than the number of X-ray images (3, 10) used to reconstruct the three-dimensional or four-dimensional X-ray image data (1), and / or that the number of X-ray images (3) in the first group (2) of the X-ray images (3) is at most six or exactly three or exactly two. [3] Computer-implemented method according to claim 1, characterized by, that a machine learning trained model (13) is used as the analysis algorithm (8) or as a sub-algorithm (12) of the analysis algorithm (8). [4] Method according to any of the preceding claims, characterized by , that at least during the acquisition of the respective X-ray image (10) of the second group (9) a collimator (14) is arranged between an X-ray source (15) used for acquiring the respective X-ray image (10) and the patient (6), wherein the collimator (14) is adjusted for acquiring the respective X-ray image (10) of the second group (9) depending on the identified relevant region (7) of the patient (6). [5] Method according to claim 4, characterized by, that the respective X-ray image (10) of the second group (9) is determined with a respective specified recording geometry (16), wherein the collimator (14) is additionally adjusted depending on the respective recording geometry (16) to capture the respective X-ray image (10) of the second group (9). [6] Method according to claim 5, characterized by , that the specified recording geometry (16) of the respective X-ray image (10) of the second group (9) specifies the 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, in order to determine a collimator setting (18) of the collimator (14) for the recording of the respective X-ray image (10), the relevant region (7) is projected into the collimator plane (17). [7] Method according to any of the preceding claims, characterized by, that the X-ray images (3, 10) of the first and / or the second group (2, 9) are acquired as part of a digital subtraction angiography. [8] Method according to claim 7, characterized by , that the four-dimensional X-ray image data (1) depict a temporal course of a local contrast agent concentration in the relevant section (4) of the vascular system (5) of the patient (6). [9] Method of one of the preceding claims, characterized by , that the relevant region (7) is identified as an area that completely encompasses the vascular system (5) within the head (19) or an organ of the patient (6). [10] Computer-implemented method for providing a machine learning-trained model (13) for use as an analysis algorithm (8) or as a sub-algorithm (12) of the analysis algorithm (8) in the method according to any of the preceding claims, comprising the steps: - Receiving multiple training datasets (20), each comprising, on the one hand, multiple X-ray images (3) of a respective patient (6), each depicting at least a part of a relevant section (4) of a 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 a target result (23), - Training a model (13) based on the training data sets (20) to determine the machine learning-trained model (13), - Provision of the machine learning trained model (13). [11] Processing device, characterized by that it is equipped to carry out the computer-implemented method according to one of the preceding claims. [12] X-ray apparatus comprising an X-ray source (15) and an X-ray detector (45) for obtaining X-ray images of a patient (6), characterized by , that the X-ray device (44) comprises a processing device (39) according to claim 11. [13] Computer program with instructions configured to perform the computer-implemented method according to any 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