Motion-compensated reconstruction in an imaging procedure

Abstract

For motion-compensated reconstruction in an imaging process, projection images (7) generated by the imaging process are obtained, representing an object from a multitude of different projection directions. Depending on the projection images (7), three-dimensional first image data (8) are reconstructed. Consistency data (9), which quantifies data consistency for each pair of projection images (7), are generated. Depending on the consistency data (9) and the three-dimensional image data (8), input data (10) are generated. Motion data (12), which characterizes object movement and / or motion artifacts in the first image data (8), are generated by applying a trained machine learning model (MLM) (11) to the input data (10). Depending on the projection images (7) and the motion data (12), three-dimensional, at least partially motion-compensated second image data are reconstructed.

Description

[0001] The present invention relates to a computer-implemented method for motion-compensated reconstruction in an imaging process, wherein projection images generated by the imaging process are obtained that represent an object from a multitude of different projection directions, and three-dimensional image data are reconstructed depending on the projection images. The invention further relates to a computer-implemented training method for training a machine learning model for use in a motion-compensated reconstruction method. The invention also relates to a corresponding data processing system and corresponding computer program products.

[0002] Patient movement poses a major challenge in medical imaging. This is particularly true for imaging techniques that use projection images with different projection directions to reconstruct three-dimensional image data, such as X-ray-based imaging techniques like computed tomography (CT), cone beam computed tomography (CBCT), or tomosynthesis techniques.

[0003] Motion-compensated reconstruction techniques can reduce motion artifacts by estimating or assuming a model of patient movement and incorporating it into the reconstruction. To avoid poor-quality reconstructions and / or the need to repeat the scan, the reliability and accuracy of the model are crucial.

[0004] In the publication A. Preuhs et al.: “Appearance learning for image-based motion estimation in tomography”, IEEE Transactions on Medical Imaging, 39(11), 3667-3678 (2020), a deep autofocus (DAF) method for motion-compensated reconstruction is described. In this method, a motion model is estimated by optimizing a learned image quality metric (IQM). The learned IQM is based on an artificial neural network (ANN) trained to predict a reprojection error (RPE) for tomographic images reconstructed from CBCT data. The RPE is then minimized through iterative reconstruction to compensate for motion.

[0005] An alternative approach is based on data consistency conditions (DCC) regarding the projection data, as described in the publication by R. Frysch & G. Rose: “Rigid motion compensation in interventional C-arm CT using consistency measure on projection data.”, Medical Image Computing and Computer-Assisted Intervention, MICCAI 2015: 18th International Conference, Munich, 2015, Proceedings, Part I 18, 298-306, Springer International Publishing. DCC-based motion compensation requires minimal computational effort but is limited to movements outside the acquisition plane.

[0006] The publication N. Hansen: “The CMA evolution strategy: A tutorial.” (arXiv:1604.00772) describes the CMA-ES optimization method.

[0007] In the publication Z. Liu et al.: “Swin transformer: Hierarchical vision transformer using shifted windows.”, Proceedings of the IEEE / CVF international conference on computer vision, 10012-10022, a neural network called Swin Transformer is described.

[0008] It is an object of the present invention to increase the accuracy of motion compensation in reconstruction using an imaging method.

[0009] This problem is solved by the respective subject matter of the independent claims. Advantageous further developments and preferred embodiments are the subject matter of the dependent claims.

[0010] The invention is based on the idea of ​​using a trained machine learning model (MLM) to predict motion data that characterizes the movement of the object being imaged and / or corresponding motion artifacts. The MLM is applied to input data that depends on reconstructed three-dimensional image data and consistency data that quantifies data consistency for each pair of projection images used in the reconstruction. A subsequent reconstruction is then performed based on the projection images and the predicted motion data.

[0011] According to one aspect of the invention, a computer-implemented method for motion-compensated, and in particular at least partially motion-compensated, reconstruction in an imaging process, especially a medical imaging process, is described. Projection images generated by the imaging process are obtained, representing an object, in particular an object to be imaged or imaged, from a multitude of different projection directions. Depending on the projection images, three-dimensional first image data are reconstructed. Consistency data, which quantifies data consistency for each pair of projection images, are generated, in particular based on the projection images.Depending on the consistency data and the initial image data, input data is generated, and motion data is generated or predicted. This motion data characterizes the movement of the object, particularly during the generation of the projection images or of the raw data used to generate the projection images, and / or motion artifacts in the initial image data. A trained machine learning model (MLM) is applied to the input data to generate the motion data. Based on the projection images and the motion data, three-dimensional, at least partially motion-compensated, second image data is reconstructed.

[0012] Unless otherwise specified, all steps of the computer-implemented method can be performed by a data processing system comprising at least one data processing device. In particular, the at least one data processing device is configured or adapted to perform the steps of the computer-implemented method. For this purpose, the at least one data processing device may, for example, store a computer program containing instructions which, when executed by the at least one data processing device, cause the at least one data processing device to perform the computer-implemented method. The computer-implemented method may also be implemented wholly or partly in hardware. The terms "data processing system" and "at least one data processing device" may be used interchangeably here and in the following. This also applies to corresponding derivatives.

[0013] In the event that the at least one data processing device comprises two or more data processing devices, certain steps performed by the at least one data processing device can also be understood as different data processing devices performing different steps or different parts of a step. In particular, it is not necessary for each data processing device to perform the steps. In other words, the execution of the steps can be distributed among the two or more data processing devices.

[0014] In general terms, a trained MLM can replicate cognitive functions that connect people with other human minds. Specifically, through training based on training data, the MLM can adapt to new circumstances and detect and extrapolate patterns. Another term for a trained MLM is "trained function."

[0015] In general, the parameters of a multi-level marketing (MLM) system can be adjusted or updated through training. This can involve supervised training, semi-supervised training, unsupervised training, reinforcement learning, and / or active learning. Furthermore, representational learning, also known as feature learning, can be employed. Specifically, the parameters of MLMs can be adjusted iteratively through multiple training steps. In particular, a specific loss function, also referred to as the cost function, can be minimized during training. When training an artificial neural network (ANN), the backpropagation algorithm can be used.

[0016] An MLM can, in particular, include an ANN, a support vector machine, a decision tree, and / or a Bayesian network, and / or the MLM can be based on k-means clustering, Q-learning, genetic algorithms, and / or association rules. Specifically, an ANN can be or include a deep neural network, a convolutional neural network (CNN), or a convolutional deep neural network. Furthermore, an ANN can be an adversarial network, a deep adverse network, and / or a generative adverse network (GAN).

[0017] In the computer-implemented method according to the invention, the MLM generates output data by applying it to the input data, which includes or consists of the movement data.

[0018] In particular, the projection images contain a corresponding projection image for each of the many different projection directions. For example, one, and in particular exactly one, projection image is obtained for each of the many different projection directions.

[0019] The projection images are generated, in particular, based on raw data that is generated, captured, or acquired using an imaging device. The generation of the raw data and the generation of the projection images are not necessarily part of the computer-implemented method according to the invention; rather, the computer-implemented method according to the invention is, for example, a subsequent step after the generation of the projection images. However, each embodiment of the computer-implemented method results in a corresponding embodiment of a method for motion-compensated reconstruction in an imaging process that is not purely computer-implemented, by including corresponding steps for generating the projection images.

[0020] The fact that the second set of image data is at least partially motion-compensated can be understood, in particular, to mean that the reconstruction used to generate the second set of image data is adapted, compared to the reconstruction used to generate the first set of image data, in such a way that signatures or artifacts resulting from the object's movement are completely or partially removed. Specifically, the motion artifacts are less pronounced in the second set of image data than in the first set.

[0021] The corresponding process steps can also be performed iteratively. In other words, depending on the consistency data and the second set of image data, further input data can be generated, and further motion data can be generated by applying the MLM to this additional input data. Then, depending on the projection images and the additional motion data, three-dimensional third set of image data can be reconstructed, which, in particular, exhibits stronger motion compensation than the second set of image data, and so on. In other words, the steps of generating the input data, generating the motion data, and reconstructing, depending on the projection images and the motion data, can be repeated iteratively, with the most recently reconstructed image data being used to generate the input data for the next iteration. The iterations can also be performed within the framework of an optimization procedure.

[0022] The input data can include the consistency data and the initial image data, or consist of the consistency data and the initial image data. Alternatively, further processing or manipulation steps may be performed to process the consistency data and / or the initial image data in order to generate the input data from them.

[0023] In some embodiments, it is also possible to fuse, for example concatenate, the consistency data and the initial image data, or the consistency data and processed image data dependent on the initial image data, to generate the input data.

[0024] Training the MLM is also not necessarily part of the computer-implemented method for motion-compensated reconstruction according to the invention. In other words, the MLM can be trained prior to the execution of the computer-implemented method according to the invention. The MLM can, for example, be designed as an artificial neural network (ANN), such as a convolutional neural network (CNN), or as a transformer network, or it can include one or more of the aforementioned networks. For example, the MLM can also be designed or based on, or be configured as described in, or based on, the publication by Preuhs et al.

[0025] To reconstruct the first image data and the second image data, a known image reconstruction method can be used, for example a method in which the reconstruction is based on a motion model of the object's movement, such as that explained in the publication by Preuhs et al.

[0026] In particular, the motion model can be updated or adjusted based on predicted or generated motion data to enable improved motion-compensated reconstruction. Alternatively or additionally, the motion data, or parts thereof, can be used as a measure of motion artifacts or the severity of motion artifacts in an objective function for iterative optimization of the reconstruction.

[0027] The object is, in particular, an object that can be imaged using the imaging procedure, for example, a patient or a part of the patient's body. Depending on the design of the imaging device or the imaging procedure, the projection images can be of different types. For example, it could be an X-ray-based imaging procedure, such as a CT scan, a CBCT scan, or a tomosynthesis procedure. In this case, the projection images are X-ray projection images.

[0028] Through reconstruction, a three-dimensional reconstruction is generated from the multitude of two-dimensional projection images in the form of the first image data or the second image data. The first image data and the second image data, for example, contain corresponding voxel values ​​for each voxel of a given three-dimensional voxel grid, where the voxel values ​​can correspond, for example, to attenuation values ​​relating to the attenuation of X-rays or similar radiation used.

[0029] Since the projection images represent the same object according to a variety of different projection directions, the pixel values ​​of the projection images are generally not independent. This can be quantified by determining the consistency data for each pair of projection images. In particular, the consistency conditions described in the aforementioned document by Frysch and Rose can be used.

[0030] In contrast to the DAF method mentioned at the outset, the input data for the MLM in the computer-implemented method according to the invention depend not only on the reconstructed data but also on the consistency data. The data consistency of a pair of projection images is disturbed or modified by movement of the object during the generation of the projection images, at least for some pairs or a subset of the pairs. The consistency data thus includes a characteristic signature of the object's movement.By supplying the MLM with consistency data in addition to the initial image data, or with information dependent on these data, in order to predict the motion data, a more accurate prediction of the motion data is achieved, and accordingly, improved accuracy in motion compensation and / or faster convergence of the optimization or iterative reconstruction is achieved, if such a method is used.

[0031] According to at least one embodiment, the first image data are reconstructed based on the projection images and a predefined initial motion model for the object's movement. Depending on the motion data and the initial motion model, an adapted motion model is generated. The second image data are then reconstructed based on the adapted motion model, again depending on the projection images.

[0032] The motion model may, for example, correspond to a description of a trajectory of the object's movement, or an approximate description of the trajectory, or may include such a description, or other parameters or characteristics of the object's movement.

[0033] Generating the first image data can, for example, correspond to an initial iteration of an iterative reconstruction or optimization procedure for motion-compensated reconstruction. Based on this initial image data, an objective function for optimization can then be calculated, and the initial motion model adjusted accordingly. For instance, the objective function can depend on the motion data. Generating the second image data then corresponds to an iteration following the first. Further iterations can be performed analogously, for example, until the value of the objective function lies within a predefined target range.

[0034] According to at least one embodiment, the motion data includes a reprojection error or at least another value that quantifies the strength of the motion artifacts in the first image data.

[0035] The reprojection error can be defined, for example, as described in the aforementioned document by Preuhs et al. However, the exact calculation of the reprojection error would require not only the initial image data itself but also reference image data generated without any object movement, or an exact description of the object's movement during the generation of the projection images. Since this is naturally not possible in the present scenario, the reprojection error is predicted in the aforementioned embodiments by the trained MLM, or at least by a value that quantifies the strength of the motion artifacts in the initial image data.

[0036] The reprojection error, or at least one value quantifying the strength of motion artifacts, represents a measure of how strongly the object's movement affects the initial image data, particularly compared to a hypothetical scenario in which no such movement occurred. Accordingly, the reprojection error, or at least one value quantifying the strength of motion artifacts, can also be seen as a measure of how successfully motion compensation was implemented during the reconstruction of the initial image data. In an iterative reconstruction process, the reprojection error, or at least one value quantifying the strength of motion artifacts, can be used, for example, as an error to be minimized within a corresponding optimization objective function.

[0037] The at least one value that quantifies the strength of the motion artifacts in the initial image data can, for example, include a voxel-wise determined metric as an alternative to or in addition to the reprojection error.

[0038] The objective function of the optimization can, for example, depend on or correspond to the reprojection error or to at least one value that quantifies the strength of the motion artifacts.

[0039] According to at least one embodiment, the motion data includes a direction of movement of a translational movement of the object, particularly during the generation of the projection images, and / or a direction of rotation of a rotational movement of the object, particularly during the generation of the projection images.

[0040] The movement of the object therefore includes, in particular, translational movement and / or rotational movement.

[0041] The direction of translational motion can be given, for example, as a direction vector in a three-dimensional coordinate system. Alternatively, the direction of motion can be specified by indicating, for each of the three coordinate axes of a Cartesian coordinate system, whether or not a movement, or a significant movement, has occurred along the corresponding axis. This can be expressed, for example, as a binary vector.

[0042] The direction of rotation can also be given, in particular, by a direction vector that is parallel to an axis of rotation of the rotational motion and whose orientation indicates a direction of rotation about the axis of rotation. Here, too, it is possible that the direction of rotation is specified in such a way that for each of the three coordinate axes, it is indicated whether a rotation about the corresponding coordinate axis is given as the axis of rotation and, if so, in which direction.

[0043] The direction of translational motion and / or rotational motion can be taken into account, particularly during the reconstruction of the second image data for motion compensation. Specifically, the solution space underlying the reconstruction optimization can be restricted such that only solutions corresponding to the direction of translational motion and / or rotational motion are considered or preferred.

[0044] In particular, the direction of movement and / or the direction of rotation can be used to generate the adapted motion model depending on the initial motion model. The adapted motion model can be specifically designed to include translational movement according to the object's direction of movement and / or rotational movement according to the object's direction of rotation, as predicted by the MLM.

[0045] Such embodiments can, in particular, achieve more effective motion compensation and / or faster convergence of the motion compensation, i.e., motion compensation in fewer iterations.

[0046] According to at least one embodiment, the motion data includes a motion trajectory of the object's movement.

[0047] The motion trajectory can, for example, be given as a time series of a position and / or orientation of the object in three-dimensional space, i.e., for example, as a time series of a six-dimensional vector that contains three values ​​for each point in time that define the position, and three values ​​that define the orientation, for example in the form of Euler angles or the like.

[0048] The motion trajectory can be taken into account, particularly when reconstructing the second image data for motion compensation. Specifically, the solution space underlying the reconstruction optimization can be restricted in such a way that only solutions corresponding to the motion trajectory are drawn, or such solutions are preferred.

[0049] In particular, the motion trajectory can be used to generate the adapted motion model depending on the initial motion model. The adapted motion model can, in particular, be modified to include a motion trajectory as predicted by the MLM.

[0050] Such embodiments can, in particular, achieve more effective motion compensation and / or faster convergence of the motion compensation, i.e., motion compensation in fewer iterations.

[0051] According to at least one embodiment, the adapted motion model is generated within the framework of an optimization procedure in which the reprojection error or the at least one value that quantifies the strength of the motion artifacts in the first image data is optimized, in particular minimized.

[0052] According to at least one embodiment, the adapted motion model is generated depending on the direction of movement and / or the direction of rotation.

[0053] According to at least one embodiment, the adapted motion model is generated depending on the motion trajectory.

[0054] According to at least one embodiment, the consistency data includes an error matrix, wherein each entry of the error matrix is ​​assigned to a pair of projection images and quantifies an inconsistency between the projection images of the respective pair corresponding to the motion artifacts.

[0055] In particular, each row of the error matrix corresponds to one of the projection images, and each column of the error matrix corresponds to one of the projection images. Each entry defined by a row and a column of the error matrix then corresponds to the corresponding pair of projection images.

[0056] For example, if there is no movement of the object during the generation of the projection images, then the inconsistency between different projection images, or projection images from different projection directions, is always zero. The inconsistency value thus reflects the influence of movement and the data consistency between the projection images of the respective pair.

[0057] One advantage of using the error matrix as part of the input data or as the basis of the input data is that it can be processed essentially analogously to image data or images derived from it.

[0058] The error matrix can be determined, for example, as described in the publication by Frysch et al. mentioned at the beginning, i.e., given by εij=1N∑n=1N|Si(ϑin,sin)−Sj(ϑjn,sjn)|p, as defined in section 2.1 of the publication by Frysch et al.

[0059] According to at least one embodiment, a multitude of layer images are generated based on the three-dimensional image data, and the input data includes the multitude of layer images.

[0060] A layered image is, in particular, a two-dimensional image extracted from the three-dimensional reconstruction of the initial image data. Specifically, a layered image corresponds to a two-dimensional section through the three-dimensional reconstruction of the initial image data.

[0061] According to at least one embodiment, the imaging method is a computed tomography (CT) method, a cone-beam computed tomography (CBCT) method, or a tomosynthesis method.

[0062] In the case of a CT or CBCT scan, the initial image data can be generated, for example, through full-angle reconstruction. In other words, the angular range covered by the various projection directions is large enough to allow for a complete three-dimensional reconstruction. For example, the angular range is at least 180° plus the so-called fan angle of the X-ray beam used. This can result in an angular range of at least 220°.

[0063] In alternative embodiments, the three-dimensional first image data include at least two, in particular at least two different, partial angle reconstructions, and the input data include the at least two partial angle reconstructions.

[0064] In particular, not all necessary projection directions are used to generate a partial angle reconstruction, which would be required to create a complete three-dimensional reconstruction. In the example above, the angular range per partial angle reconstruction would be less than 220° or less than 180°, for example, in the range of 10° to 120°.

[0065] In particular, the two or more partial angle reconstructions together cover an angle range that would be required for a full angle reconstruction, for example at least 180° plus the fan angle or at least 220°.

[0066] By providing two or more partial angle reconstructions, the MLM is supplied with different information sources in which the object's movements may have varying degrees or effects. A suitably trained MLM can therefore predict motion data with increased accuracy.

[0067] According to a further aspect of the invention, a computer-implemented training method for training an MLM for use in a computer-implemented method, in particular in a computer-implemented method according to the invention, for motion-compensated reconstruction in an imaging process is described. In this method, projection images are simulated by simulating the imaging process under the assumption of a predetermined movement of the object to be imaged. These projection images represent the object from a multitude of different projection directions. Depending on the simulated projection images, three-dimensional training image data are reconstructed. Depending on the predetermined movement, annotation data for motion data, which characterizes the movement of the object and / or motion artifacts in the training image data, are generated.Consistency training data, which quantifies data consistency for each pair of simulated projection images, is generated. Based on the consistency training data and the training image data, training input data is created. Motion data is predicted by applying the MLM, particularly in an untrained or partially trained state, to the training input data. A value for a predefined loss function is determined based on the predicted data and the annotation data. The MLM is then updated based on the value of the loss function.

[0068] The steps mentioned can be considered an iteration of the training process. These steps can be repeated with different objects and / or different movements of the object in a multitude of iterations. In particular, the iterations can be performed until a predefined termination or convergence criterion is met, specifically until the value of the loss function lies within a corresponding predefined target range.

[0069] The predefined movement can, for example, be specified by a corresponding motion trajectory. The annotation data structurally corresponds, for example, to the predicted motion data. For instance, the motion data can include a reprojection error of the training image data and / or a direction of movement of a translational movement of the object and / or a direction of rotation of a rotational movement of the object and / or the motion trajectory of the object, as described above with regard to the computer-implemented method for motion-compensated reconstruction according to the invention. Accordingly, the annotation data includes corresponding ground truth data for the reprojection error and / or the direction of movement of the translational movement and / or the direction of rotation of the rotational movement and / or the motion trajectory.Since the projection images are generated through simulation, the associated annotation data is obtained with particularly little effort.

[0070] Further embodiments of the computer-implemented training method according to the invention follow directly from the various configurations of the computer-implemented method according to the invention for motion-compensated reconstruction and vice versa. In particular, individual features and corresponding explanations as well as advantages relating to the various configurations of the computer-implemented method according to the invention for motion-compensated reconstruction can be transferred analogously to corresponding configurations of the computer-implemented training method according to the invention.

[0071] According to at least one embodiment of the computer-implemented method for motion-compensated reconstruction according to the invention, the MLM is or was trained using a computer-implemented training method according to the invention.

[0072] According to a further aspect of the invention, a data processing system is provided which is configured to carry out a computer-implemented method for motion-compensated reconstruction and / or a computer-implemented training method according to the invention.

[0073] In the present disclosure, the terms "data processing system" and "at least one data processing device" can be used interchangeably. A data processing device can be understood to be, in particular, a data processing device that contains a processing circuit. The data processing device can thus, in particular, process data to perform arithmetic operations. This may also include operations to perform indexed accesses to a data structure, for example, a lookup table (LUT), as well as a data processing process implemented in hardware.

[0074] The data processing device may, in particular, contain one or more computers, one or more microcontrollers, and / or one or more integrated circuits, for example, one or more application-specific integrated circuits (ASICs), one or more field-programmable gate arrays (FPGAs), and / or one or more systems on a chip (SoCs). The data processing device may also contain one or more processors, for example, one or more microprocessors, one or more central processing units (CPUs), one or more graphics processing units (GPUs), and / or one or more signal processors, in particular one or more digital signal processors (DSPs).The data processing device may also include a physical or virtual network of computers or other units of the aforementioned type.

[0075] In various embodiments, the data processing device includes one or more hardware and / or software interfaces and / or one or more storage units.

[0076] A storage unit can be volatile data storage, for example as dynamic random access memory (DRAM) or static random access memory (SRAM), or as non-volatile data storage, for example as read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or flash EEPROM, ferroelectric random access memory (FRAM), or magnetoresistive random access memory.It can be designed as MRAM (magnetoresistive random access memory) or as phase-change random access memory, PCRAM (phase-change random access memory).

[0077] According to a further aspect of the invention, an imaging arrangement is specified which includes a data processing system configured to perform a computer-implemented method for motion-compensated reconstruction according to the invention, and an imaging device configured to generate the projection images.

[0078] The imaging device can be, for example, an X-ray-based imaging device, such as a CT scanner, a CBCT scanner, a tomosynthesis device, or a C-arm X-ray angiography system. The CBCT scanner or tomosynthesis device can also be configured as a C-arm device.

[0079] According to a further aspect of the invention, a first computer program with first instructions is specified. When the first instructions are executed by a data processing system, the first instructions cause the data processing system to perform a computer-implemented method according to the invention for motion-compensated reconstruction in an imaging procedure.

[0080] The initial instructions can be provided as program code. This program code can be, for example, binary code or assembly language, and / or source code in a programming language such as C, and / or a program script such as Python.

[0081] According to a further aspect of the invention, a second computer program with second instructions is provided. When the second instructions are executed by a data processing system, they cause the data processing system to perform a computer-implemented training method according to the invention.

[0082] The second set of instructions can be provided as program code. This program code can be, for example, binary code or assembly language, and / or source code in a programming language such as C, and / or a program script such as Python.

[0083] According to a further aspect of the invention, a computer-readable storage medium is specified, in particular a physical and / or non-volatile computer-readable storage medium that stores a first computer program and / or a second computer program according to the invention.

[0084] The first computer program, the second computer program, and the computer-readable storage medium are each computer program products containing the first instructions and the second instructions, respectively.

[0085] Above and below, the solution according to the invention is described with respect to both the claimed systems and the claimed methods. Features, advantages, or alternative embodiments can be assigned to the other claimed subject matter and vice versa. In other words, the claims and embodiments for the systems can be improved by features described or claimed in connection with the respective methods. In this case, the functional features of the method are implemented by physical units of the system.

[0086] Furthermore, the solution according to the invention is described above and below with respect to methods and systems for motion-compensated reconstruction in an imaging procedure, as well as with respect to methods and systems for providing a trained MLM. Features, advantages, or alternative embodiments can be assigned to the other claimed subject matter and vice versa. In other words, claims and embodiments for providing a trained MLM can be improved with features that are described or claimed in connection with motion-compensated reconstruction in an imaging procedure.In particular, the datasets used in the procedures and systems may have the same properties and characteristics as the corresponding datasets used in the procedures and systems to provide a trained MLM, and the trained MLMs provided by the respective procedures and systems may be used in the procedures and systems for motion-compensated reconstruction in an imaging procedure.

[0087] Further features and combinations of features of the invention will become apparent from the figures and their description, as well as from the claims. In particular, further embodiments of the invention need not necessarily include all features of any one of the claims. Further embodiments of the invention may have features or combinations of features not mentioned in the claims.

[0088] The invention is explained in more detail below with reference to specific embodiments and associated schematic drawings. In the figures, identical or functionally equivalent elements may be designated with the same reference numerals. The description of identical or functionally equivalent elements is not necessarily repeated with respect to different figures.

[0089] The figures show Fig. 1 a schematic representation of an exemplary embodiment of an imaging arrangement according to the invention; Fig. 2 a schematic block diagram of an exemplary embodiment of a computer-implemented method according to the invention for motion-compensated reconstruction in an imaging procedure; Fig. 3 a schematic representation of an artificial neural network; and Fig. 4 a schematic representation of a convolutional neural network.

[0090] In Fig. Figure 1 schematically shows an exemplary embodiment of an imaging arrangement 1 according to the invention. The imaging arrangement 1 comprises an imaging device 3 which is configured to generate a plurality of projection images 7 that depict an object from a plurality of different projection directions.

[0091] The imaging device 3 is shown as an example of a C-arm X-ray unit with an X-ray source 4 and an X-ray detector 5. The imaging device 3 can therefore be configured, for example, as a CBCT device. However, the following explanations can be applied analogously to other imaging devices 3 that can generate projection images 7 depicting an object from a multitude of different projection directions.

[0092] The object can be, for example, a patient or a part of the patient's body. The patient can be placed on a patient table 2 of the imaging setup 1.

[0093] The imaging arrangement 1 comprises a data processing system 6 according to the invention, which is configured to perform a computer-implemented method according to the invention for motion-compensated reconstruction in an imaging process based on the projection images 7.

[0094] Fig. Figure 2 shows a schematic block diagram of an exemplary embodiment of such a computer-implemented method according to the invention.

[0095] The projection images 7 are obtained, and three-dimensional first image data 8 are reconstructed based on these projection images 7. Consistency data 9, which quantify data consistency for each pair of projection images 7, are generated, and input data 10 are generated based on the consistency data 9 and the first image data 8.

[0096] For example, the input data 10 can include the consistency data 9 as well as a multitude of layer images of the object extracted from the initial image data 8. The consistency data 9 correspond, for example, to an error matrix, where each entry of the error matrix is ​​assigned to a pair of the projection images 7 and quantifies an inconsistency between the projection images 7 of the respective pair, corresponding to the motion artifacts. In particular, the error matrix of the error matrix ε ij from the above-mentioned publication by Frysch et al.

[0097] Motion data 12, which characterize object movement and / or motion artifacts in the first image data 8, are generated by applying a trained MLM 11 to the input data 10. Depending on the projection images 7 and the motion data 12, three-dimensional, at least partially motion-compensated second image data 8 will be reconstructed.

[0098] For example, the first image data and the second image data can be generated using an iterative reconstruction process. In particular, a measure of the motion-induced error in the reconstructed image data is minimized. For example, the reprojection error can be used as such a measure, as described in the publication by Preuhs et al. mentioned at the beginning. Suitable optimization methods include simplex methods or methods based on the Monte Carlo method, such as CMA-ES (covariance matrix adaptation evolution strategy).

[0099] For example, the initial image data 8 can be reconstructed for the object's movement based on the projection images 7 and a predefined initial motion model, such as a motion trajectory. The initial motion model can also correspond to a scenario in which the object is not subject to any movement.

[0100] Depending on the motion data 12 and the initial motion model, a modified motion model is then generated, and the second set of image data is reconstructed based on the projection images 7 and the modified motion model. In the next iteration, the input data 10 is generated based on the consistency data 9 and the second set of image data, and the MLM 11 is applied to it to generate further motion data 12, and so on.

[0101] In particular, the motion data 12 includes the reprojection error 12a. Based on the reprojection error 12a predicted or estimated using the MLM 11, the objective function of the optimization can then be evaluated and the motion model adjusted accordingly so that the reprojection error 12a is minimized.

[0102] For example, the motion data may include a motion classification 12b and / or a motion trajectory 12c for the object's movement. The motion model can then be directly adapted according to the motion classification 12b and / or the motion trajectory 12c.

[0103] The MLM 11 can, for example, be designed as an ANN, specifically as an ANN with a feature extraction stage and a regression stage as described in the publication by Preuhs et al. However, the ANN can also have a different architecture, particularly a CNN-based or transformer-based one, such as the Swin Transformer mentioned above.

[0104] According to the invention, the MLM 11 is applied to the input data 10, which is based not only on the first image data 8, for example the layer images, but also on the consistency data 9, for example the error matrix.

[0105] Different motion patterns are represented differently in the error matrix. Therefore, the shape and pattern of inconsistencies in the error matrix provide a wealth of information that the MLM 11 can exploit in corresponding embodiments to estimate the motion data 12. For example, a translation peak of 1 mm along the cranial / caudal axis of the patient results in a different pattern than the same motion peak as a rotation around this axis.

[0106] In combination with the first image data 8, therefore, enough features are available to predict the motion trajectory in some embodiments.

[0107] In some embodiments, the output of the MLM 11 is designed according to a multi-class learning approach. For example, a first output of the MLM 11 is a scalar that encodes the total motion error, such as the reprojection error 12a.

[0108] In some embodiments, a second output of MLM 11 is a six-parameter encoding in which the respective axis of motion is specified as a vector (tx, ty, tz, rx, ry, rz) of motion classification 12b, in which the motion occurs. tx, ty, tz correspond to translational movements in the x, y, and z directions, respectively, whereas rx, ry, rz correspond to rotational movements about the x, y, and z directions, respectively. The vector entries can be, for example, 0 or 1, depending on whether a corresponding movement is predicted or not. Optionally, a sign can indicate the direction of the respective motion component.

[0109] In some embodiments, a third output of MLM 11 is a direct prediction of the motion trajectory 12c. The motion trajectory 12c can, for example, be encoded as a fixed spline. However, it is also possible for MLM 11 to define the number and distribution of the simplex nodes, or more generally, the structure of the motion, such as high-frequency or low-frequency motion in the middle part of the motion trajectory 12c, and so on.

[0110] The MLM 11 can be trained, for example, by using motion-free scans or simulated scans and synthetically adding the motion. This allows for the provision of a reference standard for the respective outputs and the subsequent training of the MLM 11.

[0111] After training, MLM 11 can be used particularly for optimization. Alternatively or additionally, the movement classification 12b and / or the movement trajectory 12c can be used to restrict the solution space for optimization.

[0112] Fig. Figure 3 shows an embodiment of an artificial neural network, ANN, 800. The ANN 800 comprises nodes 820, ..., 832 and edges 840, ..., 842, where each edge 840, ..., 842 is a directed connection from a first node 820, ..., 832 to a second node 820, ..., 832. In general, the first node 820, ..., 832 and the second node 820, ..., 832 are distinct nodes 820, ..., 832. However, it is also possible for the first node 820, ..., 832 and the second node 820, ..., 832 to be identical. In Fig. For example, edge 840 is a directed connection from node 820 to node 823, and edge 842 is a directed connection from node 830 to node 832. An edge 840, ..., 842 from a first node 820, ..., 832 to a second node 820, ..., 832 is also called an incoming edge for the second node 820, ..., 832 and an outgoing edge for the first node 820, ..., 832.

[0113] In this example, the nodes 820, ..., 832 of ANN 800 can be arranged in layers 810, ..., 813, where the layers can have an intrinsic order introduced by the edges 840, ..., 842 between the nodes 820, ..., 832. In particular, the edges 840, ..., 842 can only exist between adjacent layers of nodes. In the example shown, there is an input layer 810 consisting only of nodes 820, ..., 822 with no incoming edges, an output layer 813 consisting only of nodes 831, 832 with no outgoing edges, and hidden layers 811, 812 between the input layer 810 and the output layer 813. In general, the number of hidden layers 811, 812 can be chosen arbitrarily. For a multilayer perceptron (MLP), this number is at least one. The number of nodes is 820, ..., 822 within the input layer 810 usually refers to the number of input values ​​of the artificial neural network 800, and the number of nodes 831, 832 within the output layer 813 usually refers to the number of output values ​​of the artificial neural network 800.

[0114] In particular, each node 820, ..., 832 of the artificial neural network 800 can be assigned a real number as its value. Here, x denotes... (n) iThe value of the i-th node 820, ..., 832 of the n-th layer 810, ..., 813. The values ​​of nodes 820, ..., 822 of the input layer 810 correspond to the input values ​​of the artificial neural network 800. The values ​​of nodes 831, 832 of the output layer 813 correspond to the output value of the artificial neural network 800. Furthermore, each edge 840, ..., 842 can have a weight, which is a real number. In particular, the weight is a real number within the interval [-1, 1] or within the interval [0, 1]. Here, w denotes (m,n) i,j The weight of the edge between the i-th node 820, ..., 832 of the m-th layer 810, ..., 813 and the j-th node 820, ..., 832 of the n-th layer 810, ..., 813. Furthermore, the abbreviation w (n) i,j for the weight w (n,n+1) i,jdefined. To calculate the output values ​​of neural network 800, the input values ​​are propagated through neural network 800. Specifically, the values ​​of nodes 820, ..., 832 of the (n+1)th layer 810, ..., 813 can be calculated based on the values ​​of nodes 820, ..., 832 of the nth layer 810, ..., 813 as xj(n+1)=f(∑ixi(n)wi,j(n)).

[0115] In this context, the function f is referred to as the transfer function or activation function. Well-known transfer functions include step functions, sigmoid functions (for example, the logistic function, the generalized logistic function, the hyperbolic tangent, the arctangent function), the error function, the smoothstep function, and rectifier functions. The transfer function is used, for example, for normalization. Specifically, the values ​​are propagated layer by layer through neural network 800, with the values ​​of input layer 810 being given by the input of neural network 800. The values ​​of the first hidden layer 811 can be calculated based on the values ​​of input layer 810 of neural network 800, the values ​​of the second hidden layer 812 can be calculated based on the values ​​of the first hidden layer 811, and so on.

[0116] To determine the values ​​w (m,n)i,j To define the edges, the neural network 800 must be trained with training data. The training data includes, in particular, training input data and training output data (denoted as t). i In a training step, the neural network 800 is applied to the training input data to generate computed output data. Specifically, the training data and the computed output data comprise a number of values ​​equal to the number of nodes in the output layer. A comparison between the computed output data and the training data is used to recursively adjust the weights within the neural network 800 (backpropagation algorithm). Specifically, the weights are modified according to the following formula. wi,j'(n)=wi,j(n)−γ δj(n)xi(n), where γ is a predefined learning rate, and the numbers δ (n) j can be calculated recursively as δj(n)=(∑kδk(n+1)wj,k(n+1))f'(xi(n)wi,j(n)) based on δ (n+1) j , if the (n+1)th layer is not output layer 813, and δj(n)=(xj(n+1)−tj(n+1))f'(xi(n)wi,j(n)), if the (n+1)th layer is output layer 813, where f' is the first derivative of the activation function, and t (n+1) j The comparison training value for the j-th node of the output layer is 813.

[0117] A convolutional neural network (CNN) is an ANN that uses a convolutional operation instead of general matrix multiplication in at least one of its layers. These layers are called convolutional layers. Specifically, a convolutional layer performs a dot product of one or more convolutional kernels with the convolutional layer's input data, where the entries of the one or more convolutional kernels are parameters or weights that can be adjusted through training. In particular, one can use the inner Frobenius product and the ReLU activation function. A convolutional neural network may include additional layers, such as pooling layers, fully connected layers, and / or normalization layers.

[0118] By using convolutional neural networks, input can be processed very efficiently. This is because a convolution operation based on different kernels can extract different image features, allowing the relevant image features to be determined during training by adjusting the weights of the convolution kernel. Furthermore, because weights are shared across convolutional kernels, fewer parameters need to be trained, preventing overfitting during the training phase and enabling faster training or more layers in the network, thus improving network performance.

[0119] Fig.Figure 4 shows an exemplary embodiment of a convolutional neural network 700. In the illustrated embodiment, the convolutional neural network 700 comprises an input node layer 710, a convolution layer 711, a pooling layer 713, a fully connected layer 714, and an output node layer 716, as well as hidden node layers 712 and 714. Alternatively, the convolutional neural network 700 can also include multiple convolution layers 711, multiple pooling layers 713, and / or multiple fully connected layers 715, as well as other types of layers. The order of the layers can be chosen arbitrarily; typically, fully connected layers 715 are used as the last layers before the output layer 716.

[0120] In particular, in a convolutional neural network 700, the nodes 720, 722, 724 of a node layer 710, 712, 714 can be viewed as a d-dimensional matrix or as a d-dimensional image. Specifically, in the two-dimensional case, the value of the node 720, 722, 724 indexed by i and j in the nth node layer 710, 712, 714 can be denoted as x(n)[i, j]. However, the arrangement of the nodes 720, 722, 724 of a node layer 710, 712, 714 has no influence on the computations performed within the convolutional neural network 700, as these are determined solely by the structure and the weights of the edges.

[0121] A convolution layer 711 is a connecting layer between a front node layer 710 with node values ​​x(n-1) and a back node layer 712 with node values ​​x(n). A convolution layer 711 is characterized in particular by the structure and weights of the incoming edges that perform a convolution operation based on a certain number of kernels. Specifically, the structure and weights of the edges of the convolution layer 711 are chosen such that the values ​​x(n) of the nodes 722 of the back node layer 712 are computed as a convolution x(n) = K * x(n-1) based on the values ​​x(n-1) of the nodes 720 of the front node layer 710, where the convolution * in the two-dimensional case is defined as x(n)[i,j]=(K∗x(n−1))[i,j]=∑i'∑j'K[i',j']⋅x(n−1)[i−i',j−j']

[0122] Here, the kernel K is a d-dimensional matrix, in this example a two-dimensional matrix that is usually small compared to the number of nodes 720, 722, for example a 3x3 matrix or a 5x5 matrix. This means, in particular, that the weights of the edges in the convolution layer 711 are not independent, but are chosen such that they yield the aforementioned convolution equation. Specifically, for a kernel that is a 3x3 matrix, there are only 9 independent weights, where each entry of the kernel matrix corresponds to an independent weight, regardless of the number of nodes 720, 722 in the front node layer 710 and the back node layer 712.

[0123] In general, convolutional neural networks use 700 node layers 710, 712, 714 with a multitude of channels, particularly due to the use of a multitude of kernels in the convolution layers 711. In these cases, the node layers can be viewed as (d+1)-dimensional matrices, where the first dimension indexes the channels. The effect of a convolution layer 711 is then defined in a two-dimensional example as xb(n)[i,j]=∑a(Ka,b∗xa(n−1)[i,j]=∑a∑i'∑j'Ka,b[i',j']⋅xa(n−1)[i−i',j−j'] where xa(n) corresponds to the a-th channel of the preceding layer 710, xb(n) corresponds to the b-th channel of the following nodal layer 712 and K a,b corresponds to one of the kernels. If a convolution layer 711 acts on a preceding nodal layer 710 with A channels and outputs a subsequent nodal layer 712 with B channels, there exist A·B independent d-dimensional kernels K.a,b .

[0124] In general, 700 activation functions can be used in convolutional neural networks. In this embodiment, ReLU (rectified linear unit) is used, with R(z) = max(0, z), so that the effect of the convolution layer 711 in the two-dimensional example xb(n)[i,j]=R(∑a(Ka,b∗xa(n−1)[i,j])=R(∑a∑i'∑j'Ka,b[i',j']⋅xa(n−1)[i−i',j−j']) It is also possible to use other activation functions, such as ELU (Exponential Linear Unit), LeakyReLU, Sigmoid, Tanh, or Softmax.

[0125] In the illustrated embodiment, the input layer 710 comprises 36 nodes 720 arranged in a two-dimensional 6x6 matrix. The first hidden node layer 712 comprises 72 nodes 722 arranged as two-dimensional 6x6 matrices, each of which is the result of convolution of the values ​​of the input layer with a 3x3 kernel within the convolution layer 711. Equivalently, the nodes 722 of the first hidden node layer 712 can be interpreted as a three-dimensional 2x6x6 matrix, the first dimension corresponding to the channel dimension.

[0126] One advantage of using convolutional layers 711 is that a spatially local correlation of the input data can be exploited by enforcing a local connectivity pattern between the nodes of neighboring layers, in particular by connecting each node only to a small range of the nodes of the preceding layer.

[0127] A pooling layer 713 is a connecting layer between an preceding node layer 712 with node values ​​x(n-1) and a subsequent node layer 714 with node values ​​x(n). A pooling layer 713 can be characterized, in particular, by the structure and weights of the edges and the activation function, which perform a pooling operation based on a nonlinear pooling function f. For example, in the two-dimensional case, the values ​​x(n) of the nodes 724 of the subsequent node layer 714 can be calculated based on the values ​​x(n-1) of the nodes 722 of the anterior node layer 712 as follows. xb(n)[i,j]=f(xb(n−1)[id1,jd2],..., xb(n−1)[(i+1)d1−1,(j+1)d2−1]).

[0128] In other words, by using a pooling layer 713, the number of nodes 722, 724 can be reduced by replacing a number of d1-d2 neighboring nodes 722 in the preceding node layer 712 with a single node 722 in the subsequent node layer 714, which is calculated as a function of the values ​​of the aforementioned number of neighboring nodes. The pooling function f can, in particular, be the max function, the mean, or the L2 norm. Specifically, in a pooling layer 713, the weights of the incoming edges are fixed and are not changed by training.

[0129] The advantage of using a pooling layer 713 is that it reduces the number of nodes (722, 724) and the number of parameters. This leads to a reduction in computational overhead on the network and helps control overfitting.

[0130] In the illustrated embodiment, the pooling layer 713 is a max-pooling layer, in which four adjacent nodes are replaced by only one node, where the value is the maximum of the values ​​of the four adjacent nodes. The max-pooling is applied to each d-dimensional matrix of the preceding layer. In this embodiment, the max-pooling is applied to each of the two-dimensional matrices, thereby reducing the number of nodes from 72 to 18.

[0131] In general, the last layers of a convolutional neural network can be fully connected layers. A fully connected layer is a linking layer between a preceding node layer and a subsequent node layer. A fully connected layer can be characterized by the presence of a majority, in particular all, edges between the nodes of the preceding node layer and the nodes of the subsequent node layer, and the weight of each of these edges can be individually adjusted.

[0132] In this embodiment, the nodes 724 of the leading node layer 714 of the fully connected layer 715 are represented both as two-dimensional matrices and additionally as non-connected nodes, displayed as a line of nodes, the number of which has been reduced for better clarity. This process is also called flattening. In this embodiment, the number of nodes 726 in the subsequent node layer 716 of the fully connected layer 715 is less than the number of nodes 724 in the preceding node layer 714. Alternatively, the number of nodes 726 can also be equal to or greater than the number of nodes 724 in the preceding node layer 714.

[0133] Furthermore, in this embodiment, the softmax activation function is used within the fully connected layer 715. By applying the softmax function, the sum of the values ​​of all nodes 726 of the output layer 716 is equal to 1, and all values ​​of all nodes 726 of the output layer 716 are real numbers between 0 and 1. In particular, when using the convolutional neural network 700 to categorize input data, the values ​​of the output layer 716 can be interpreted as the probability that the input data falls into one of the various categories.

[0134] In particular, convolutional neural networks 700 can be trained based on the backpropagation algorithm. To prevent overfitting, regularization techniques can be used, such as omitting nodes 720, ..., 724, stochastic pooling, using artificial data, weight reduction based on the L1 or L2 norm, or max-norm constraints.

[0135] The preceding description is intended to include persons of male, female or other gender identities, regardless of the grammatical gender of a particular term. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature

[0000] A. Preuhs et al.: “Appearance learning for image-based motion estimation in tomography”, IEEE Transactions on Medical Imaging, 39(11), 3667-3678 (2020

[0004] R. Frysch & G. Rose: “Rigid motion compensation in interventional C-arm CT using consistency measure on projection data.”, Medical Image Computing and Computer-Assisted Intervention, MICCAI 2015: 18th International Conference, Munich, 2015, Proceedings, Part I 18, 298-306

[0005] The CMA evolution strategy: A tutorial.” (arXiv:1604.00772

[0006] Z. Liu et al.: „Swin transformer: Hierarchical vision transformer using shifted windows.“, Proceedings of the IEEE / CVF international conference on computer vision, 10012-10022

[0007]

Claims

[1] Computer-implemented method for motion-compensated reconstruction in an imaging procedure, wherein - projection images (7) produced by the imaging process are obtained, which represent an object from a variety of different projection directions; - depending on the projection images (7) three-dimensional first image data (8) can be reconstructed; - Consistency data (9) are generated, which quantify data consistency for each pair of projection images (7); - depending on the consistency data (9) and the initial image data (8), input data (10) are generated; - Motion data (12), which characterizes a movement of the object and / or motion artifacts in the initial image data (8), are generated by applying a trained machine learning model, MLM, (11) to the input data (10); and - depending on the projection images (7) and the motion data (12), three-dimensional second image data, at least partially motion-compensated, can be reconstructed. [2] Computer-implemented method according to claim 1, wherein the motion data (12) include a reprojection error or at least a value (12a) that quantifies the strength of the motion artifacts in the first image data (8). [3] Computer-implemented method according to one of the preceding claims, wherein the motion data (12) include a direction of movement of a translational movement of the object and / or a direction of rotation of a rotational movement of the object. [4] Computer-implemented method according to one of the preceding claims, wherein the motion data (12) includes a motion trajectory (12c) of the motion of the object. [5] Computer-implemented method according to any one of the preceding claims, wherein - the first image data (8) are reconstructed depending on the projection images (7) based on a given initial motion model for the movement of the object; - depending on the motion data (12) and the initial motion model, an adapted motion model is generated; and - the second image data are reconstructed depending on the projection images (7) based on the adapted motion model. [6] Computer-implemented method according to claim 5 and claim 2, wherein the adapted motion model is generated within the framework of an optimization procedure in which the reprojection error or the at least one value (12a) that quantifies the strength of the motion artifacts in the first image data (8) is optimized. [7] Computer-implemented method according to one of claims 5 or 6 and claim 3, wherein the adapted motion model is generated depending on the direction of motion and / or the direction of rotation. [8] Computer-implemented method according to one of claims 5 to 7 and claim 4, wherein the adapted motion model is generated depending on the motion trajectory (12c). [9] Computer-implemented method according to one of the preceding claims, wherein the consistency data (9) includes an error matrix, wherein each entry of the error matrix is ​​assigned to a pair of the projection images (7) and quantifies an inconsistency between the projection images (7) of the respective pair corresponding to the motion artifacts. [10] Computer-implemented method according to one of the preceding claims, wherein a plurality of layer images is generated based on the three-dimensional first image data (8) and the input data (10) includes the plurality of layer images. [11] Computer-implemented method according to any of the preceding claims, wherein the imaging method is a computed tomography method or a cone-beam computed tomography method. [12] Computer-implemented method according to claim 11, wherein the three-dimensional first image data (8) include at least two partial angle reconstructions and the input data (10) include the at least two partial angle reconstructions. [13] Computer-implemented training method for training a machine learning model, MLM, (11) for use in a computer-implemented motion-compensated reconstruction method in an imaging procedure, wherein - by simulating the imaging process under the assumption of a predetermined movement of an object to be imaged, projection images are simulated that depict the object from a variety of different projection directions; - depending on the simulated projection images, three-dimensional training image data can be reconstructed; - Depending on the specified movement, annotation data for movement data is generated, which characterizes the movement of the object and / or movement artifacts in the training image data; - Consistency training data is generated, which quantifies data consistency for each pair of simulated projection images; - Training input data is generated depending on the consistency training data and the training image data; - the movement data are predicted by applying the MLM to the training input data; - a value of a given loss function is determined depending on the predicted data and the annotation data; and - the MLM is updated depending on the value of the loss function. [14] Data processing system (6) configured to perform a computer-implemented method according to any one of claims 1 to 12 and / or a computer-implemented training method according to claim 13. [15] computer program product - first commands which, when executed by a data processing system (6), cause the data processing system (6) to perform a computer-implemented method according to any one of claims 1 to 12; and / or - second commands which, when executed by a data processing system (6), cause the data processing system (6) to perform a computer-implemented training method according to claim 13.

Images