Technique for compensating an X-ray grating shadow using an actuator

Abstract

A technique for controlling a compensation actuator to compensate for a shadow cast by an X-ray diffraction grating in an X-ray image acquired with a swiveling X-ray unit, particularly in real time, comprises a method with a step involving the reception of motion state data indicating the movement state of the X-ray unit. The X-ray unit is mounted on a swiveling arm. The movement state is associated with a specific point in time when an X-ray image is acquired by the X-ray unit. A quality measure of the X-ray image with respect to a shadow cast by an X-ray diffraction grating located at the detector of the X-ray unit is determined. A control signal for a compensation actuator to adjust the position of the X-ray diffraction grating is determined based on the received motion state data and the determined quality measure.Determining the control signal involves optimizing the quality factor. The determined control signal is then output to the compensation actuator to compensate for the shadow cast by the X-ray scattering absorption grating.

Description

[0001] The invention relates to a technology for controlling a compensation actuator for compensating for a shadow of an X-ray scatter absorption grid in an X-ray image acquired with a pivotably mounted X-ray device, particularly in real time. In particular, a method, a computing device, a system, and a computer program product are provided.

[0002] To generate an X-ray image, X-ray absorption gratings (also called X-ray gratings, scattered radiation absorption gratings, and / or scattered radiation grids) are typically inserted into the beam path between the X-ray source (also known as X-ray source and / or source) and the X-ray detector (short: detector). Thus, the scattered radiation is filtered out with a scattered radiation grid in front of the image receiver to improve image quality. Novel, particularly additive, processes enable the production of scattered radiation grids with high aspect ratios (and / or high structure height relative to the structure period), so-called super grids, "smart grids," or focused grids.

[0003] In the case of focused grids (also known as “smart grids”), the slats are not only higher than those of conventional anti-scatter grids in order to achieve better filtering, but are also directed towards the center of the radiator in order to retain a large proportion of the useful radiation.

[0004] In conventional angiography systems, however, the new focused grids (which can also be referred to as X-ray grids, or simply grids) cannot be used directly because, when movements occur due to elastic deformations (also: deformations) in the device or gantry, the focal point of the X-ray source migrates significantly, which leads to shadowing and artifacts visible in the imaging.

[0005] Conventional methods for active X-ray grid tracking use a calculation of the control commands for the correction drives based on an elastostatic model of a C-arm. However, the resulting tracking is not precise enough to ensure good image quality in an X-ray image.

[0006] It is therefore an object of the present invention to provide a reliable solution for improved image quality, particularly in real time, of an X-ray image accompanied by a reduced radiation dose.

[0007] This object is achieved by a method for controlling a compensation actuator for compensating a shadow of an X-ray scatter absorption grid in an X-ray image recorded with a pivotably mounted X-ray device, in particular in real time, by a computing device, by a system comprising the computing device, by a computer program (and / or a computer program product), and by a computer-readable storage medium according to the appended independent claims. Advantageous aspects, features, and embodiments are described in the dependent claims and in the following description, together with advantages.

[0008] In the following, the inventive solution is described with reference to the claimed method and with reference to the claimed computing device. Features, advantages, or alternative embodiments herein may be assigned to the other claimed subject matter (e.g., the system, the computer program, or a computer program product), and vice versa. In other words, the claims for the computing device may be enhanced by features described or claimed in connection with the method. In this case, the functional features of the method are embodied by structural units of the system, and vice versa.

[0009] According to one method aspect, a (particularly computer-implemented) method is provided for controlling a compensation actuator for compensating a shadow of an X-ray scatter absorption grid in an X-ray image that is recorded with a pivotably mounted X-ray device, in particular in real time.

[0010] The method comprises a step of receiving (in particular target and / or actual) motion state data indicating a motion state of an X-ray device. The X-ray device is mounted on a pivotable arm. The motion state is associated with a time at which an X-ray image is acquired by the X-ray device. The method further comprises a step of determining a quality measure of the X-ray image with respect to a shadow of an X-ray scatter absorption grid arranged on the detector of the X-ray device. The method further comprises a step of determining a control signal for the compensation actuator for adjusting the position of the X-ray scatter absorption grid depending on the received motion state data and depending on the determined quality measure. Determining the control signal comprises optimizing the quality measure.The method further comprises a step of outputting the determined control signal to the compensation actuator for compensating the shadow of the X-ray scattering absorption grating.

[0011] The inventive technology can also be referred to as smart grid X-ray imaging with intelligent control method.

[0012] The steps of the method can be executed repeatedly and / or iteratively (particularly at runtime and / or in real time), for example, when implemented using an iterative learning control (ILC), particularly as a function approximator. This makes it possible to provide a control loop that successively improves the quality of the X-ray image by adapting the control signal to the optimized quality measure.

[0013] ILC control is a method for tracking the control of systems that operate in an iterative or repetitive mode, such as the system for aligning and / or positioning the X-ray grid. In each alignment and / or positioning task, the system must perform the same action over and over again with high precision. This action is represented by the goal of accurately tracking a chosen reference signal (e.g., the figure of merit) in a finite time interval. Repetition allows the system to improve tracking accuracy from iteration to iteration by learning the inputs required to accurately track the reference. The learning process uses information from previous iterations to improve the control signal so that an appropriate control action can eventually be found iteratively.

[0014] Perfect tracking can, in principle, be achieved using an internal model (e.g., as a standard or essential result of control engineering from the 1970s). For example, it is well known that the controller must include an I component (integrating component) if constant disturbances are to be compensated (and, in particular, if the system does not already have an integrating effect). This can be attributed, among other things, to the fact that an integrator in the Laplace domain represents a model of a step signal.

[0015] When applied to the inventive technique, the result of an ILC control is not "perfect" (and / or no perfect model of the route for use is found). According to a first variant, an approximately selected model can be used that is approximately appropriate. The inventive method (in particular including the ILC control) will then still provide an improvement. The approximately selected model can be required (in particular only) for the design of the filters used in the ILC method.

[0016] Alternatively or additionally, a principle of an internal model can lead to conditions under which perfect tracking can be achieved (e.g., ideally and / or according to the model assumption). The design of the control algorithm can still allow for many decisions required for the particular application.

[0017] A (e.g., precisely defining) property (in particular the quality measure) of the method according to the invention can improve from execution to execution (e.g., in the case of ILC control) or (in particular, even) during an execution (e.g., when using a KAF).

[0018] A single implementation of the ILC controller can comprise a (e.g., linear and / or sequential) sequence of the steps of the inventive method for controlling the compensation actuators. Alternatively or additionally, the technical term "linear" vs. "nonlinear" can be independent of the property of whether or not there is a feedback loop (and / or repetition of process steps). In particular, a linear system can comprise a feedback loop (and / or repetition of process steps).

[0019] For example, the (e.g., ultimately achieved) quality measure depends on the (e.g., determined and output) control signal. The determined (and / or output) control signal depends, in particular, on the optimization of the quality measure.

[0020] The relationship (in particular between the quality measure and the control signal) can be described as a “feedback” property and with appropriate, in particular multiple, application of the method, the X-ray image acquisition is controlled, in particular the compensation actuators.

[0021] The feedback property can be essential (e.g. depending on the choice of the function approximator) for the success of compensating the shadow of the X-ray scattering absorption grating (short: X-ray grating), which ultimately determines the image quality.

[0022] The "feedback" property can be used between individual, repeatedly traversed motion state curves (also: motion sequences and / or trajectories) (e.g., when using ILC) to improve target value curves. Alternatively or additionally, the "feedback" property can provide feedback during (or within) the traversal of a motion state curve and / or change the course of the motion state curve (e.g., when a Kernel Adaptive Filter (KAF) is used in the process, especially as a function approximator).

[0023] A KAF is an adaptive kernel filter and thus a type of nonlinear adaptive filter. An adaptive filter is a filter that adapts its transfer function to changes in signal properties over time by minimizing an error or loss function that characterizes how far the filter deviates from ideal behavior. The adaptation process is based on learning from a sequence of signal samples and is thus an online algorithm. A nonlinear adaptive filter is a filter in which the transfer function is nonlinear. Adaptive kernel filters implement a nonlinear transfer function using kernel methods. In these methods, the signal is mapped to a high-dimensional linear feature space, and a nonlinear function is approximated as a sum over kernels whose domain is the feature space.When this happens in a reproducing kernel Hilbert space, a kernel method can be a universal approximator for a nonlinear function. Kernel methods have the advantage of having convex loss functions, no local minima, and are only moderately complex to implement.

[0024] By means of the technology according to the invention, an improved quality of the recorded X-ray image can be achieved, in particular in real time (and / or during operation of the X-ray device, in short: at runtime).

[0025] Using the inventive technology, relative shifts between the positions of the X-ray source and detector (and / or the X-ray grid attached to the detector) of the X-ray machine can be compensated using the compensation actuators (abbreviated to actuators). This allows shadows of the primary X-ray beam to be compensated for by shifting the position of the X-ray grid (also known as a grid) in the acquired X-ray image (also known as a radiography and / or X-ray exposure). Alternatively or additionally, the quality of an X-ray image, for example, sharpness and / or detail accuracy, can be improved using the inventive technology.

[0026] The X-ray grating can be designed as a predominantly two-dimensional element made of an X-ray-absorbing material. The base area of ​​the X-ray grating can preferably be adapted to the surface of the detector and / or coincide with it. At least one compensation actuator can be configured to adjust the X-ray grating, in particular for adjustment relative to the surface of the detector. The compensation actuator can comprise at least two actuators for translationally displacing and / or tilting the grating along the two main axes in the plane of the grating.

[0027] Effects due to manufacturing and assembly tolerances, which are conventionally not taken into account (or cannot be taken into account) in simulation models based on rigid body models and / or (e.g. elastic) multi-body dynamics models, can be compensated by means of the inventive control of the compensation actuators and thus the quality of a recorded X-ray image can be improved.

[0028] As mentioned above, the compensation actuator system can comprise one or more compensation actuators. In particular, the compensation actuator system can comprise two compensation actuators, wherein each compensation actuator can adjust the X-ray grid along a direction (e.g., along a detector plane and / or along a base area of ​​the X-ray grid). Adjusting the X-ray grid can comprise adjusting the detector (e.g., shifting and / or tilting the detector plane). Alternatively or additionally, the X-ray grid can be attached (in particular fixedly) to the detector and / or mounted (in particular, not adjustable relative to the detector).

[0029] Compensation (e.g., along a direction) may include tilting and / or shifting the X-ray grid and / or the detector (e.g., along the direction). Alternatively or additionally, compensation may include tilting and / or shifting the X-ray source.

[0030] The pivoting arm can also be referred to as a gantry (and / or scaffold). Alternatively or additionally, the pivoting arm can include a C-arm.

[0031] Compensating for the shadow by changing the relative position of the X-ray grid (and / or detector) and the X-ray source can also be referred to as active tracking, fine-tuning or adjustment (e.g. of the X-ray grid, the detector and / or the X-ray source).

[0032] Compensating the shadow allows for higher aspect ratios (and / or a greater height of slats, especially with non-zero inclination angle, relative to the length and / or width of a grid projected onto the base area, and / or relative to the distance between adjacent slats) than in conventional X-ray gratings.

[0033] The relative (especially positional) displacements between X-ray source and detector in the detector plane can also be referred to as deflections and / or outward migrations.

[0034] Improving the quality of the X-ray image can enable a better diagnosis and / or more targeted treatment of a patient. Alternatively or additionally, a reduction in the required radiation dose to achieve a given image quality can be achieved.

[0035] For example, in an X-ray image, particularly an angiography, the use of an X-ray grid can achieve a higher separation of the primary X-ray beam from scattered radiation, thus enabling a reduction in radiation exposure for patients and medical personnel while improving image quality.

[0036] In real time can also be defined as at runtime, especially of the X-ray image (and / or fluoroscopy process).

[0037] In real time can mean that the inventive method is executed so quickly that a real time sequence does not need to be slowed down to apply the method (e.g., for acquiring an X-ray image). Alternatively or additionally, in real time (and / or at runtime) can include that the inventive technique can be applied to compensate for the shadow of the X-ray grid during the acquisition of the X-ray image (and / or without, in particular, significant, delays) and in particular during image reconstruction.

[0038] For example, when determining the control signal using actual motion state data (e.g., using a KAF), real-time capability is advantageous (and / or, possibly, even mandatory).

[0039] Alternatively or additionally, the inventive technique can be applied before the X-ray image is acquired when determining the control signal (in particular only) based on target motion state data. For example, determining the control signal based on target motion state data can be performed prior to the X-ray image being acquired (e.g., in a planning phase) (e.g., using the ILC) and / or be less time-critical.

[0040] Depending on the technical design, such a real-time feature between image evaluation for quality determination and control of the actuators may be almost mandatory (e.g. in a variant with KAF), or not (e.g. in a variant with ILC).

[0041] Alternatively or additionally, (particularly active) compensation, particularly in real time, can enable high-quality X-ray images to be provided promptly, especially while the patient is still in the imaging position. Alternatively or additionally, improving the quality of the X-ray image, particularly in real time, enables a rapid assessment of whether further X-ray images are required. Furthermore, alternatively or additionally, compensating for the shadow, particularly in real time, can avoid future follow-up appointments (e.g., several days or weeks later) (e.g., for further images, to discuss the diagnosis and / or to discuss a treatment plan).

[0042] Panning can include translational and / or rotational movements.

[0043] The pivoting arm (and / or the pivotally mounted X-ray device) can comprise translational and / or rotational degrees of freedom. In particular, the X-ray device can comprise a C-arm, for example, with five independent degrees of freedom. Alternatively or additionally, the X-ray device (and / or the pivoting arm), in particular the C-arm, can also comprise more or fewer than five independent degrees of freedom.

[0044] The X-ray grid may also be referred to as a detector grid. The X-ray grid may comprise lamellae (e.g., made of metal, particularly lead) extending from a base surface.

[0045] The base area of ​​the X-ray grating can correspond to a detector plane. Alternatively or additionally, the position of the X-ray grating along the detector plane can be parameterized using two-dimensional (2D) coordinates, for example, (x,y).

[0046] A rest position (also referred to as neutral position or unshifted position) of the X-ray grid can comprise that the X-ray source and detector of the X-ray device are arranged vertically one above the other (or one below the other), and that a focal point (also: focus point) of an X-ray beam emitted by the X-ray source is imaged onto a predetermined point (preferably the center point) of the X-ray grid.

[0047] The X-ray image may also be referred to as an X-ray photograph and / or radiography. Alternatively or additionally, the X-ray image may comprise a temporal sequence of individual (and / or instantaneous) X-ray images.

[0048] The distance between the X-ray source and the X-ray grid can be between 90 centimeters (90cm) and two meters (2m), for example 1.2m.

[0049] The detector on which the X-ray grid is mounted can comprise a weight of, for example, 50 kg. Alternatively or additionally, the pivoting arm can be elastically deformed by the movements recorded in the motion state data and / or the motion state.

[0050] The width and / or diameter of the X-ray grid can be between ten centimeters (10 cm) and fifty centimeters (50 cm), in particular between 20 cm and 45 cm. Alternatively or additionally, the area of ​​the X-ray grid can be between 100 cm 2 and 2500cm 2 , in particular between 20cm × 20cm and 45cm × 45cm.

[0051] The X-ray grid may have a rectangular, in particular a square, base. Alternatively or additionally, the X-ray grid may have a round, oval, or circular base.

[0052] The height of the X-ray grid (and / or the height of the X-ray grid's slats) can be between 5 mm and 2 cm. Alternatively or additionally, the height-to-base ratio of the X-ray grid can be one to ten (1:10) or a value of one to a number greater than ten.

[0053] In a conventional X-ray grid, the slats can extend perpendicular to the base surface. Alternatively or additionally, in an X-ray grid (also referred to as a "smart grid" and / or focused grid), the central slats can extend (at least approximately) perpendicularly, with angles to the perpendicular to the base surface increasing toward the edge of the X-ray grid. Alternatively or additionally, the slats can be aligned to a common focal point. The common focal point can correspond to the position of a point X-ray source (especially without relative shifts of the X-ray source and detector caused by position adjustments).

[0054] The outer lamellae of the (e.g. focused) X-ray grating can deviate from the perpendicular to the base of the X-ray grating by up to an angle of 20° (and / or π / 9).

[0055] The X-ray grid can be manufactured using additive manufacturing. Alternatively or additionally, the X-ray grid can comprise an X-ray-absorbing material, in particular a metal (especially as a manufacturing material), preferably lead.

[0056] The slats of the X-ray grid can form a (e.g., rectangular) grid and / or a (e.g., hexagonal) honeycomb structure. The grid and / or honeycomb structure can be filled with an X-ray-transparent material, such as aluminum. Filling the grid and / or honeycomb can improve the stability of the X-ray grid.

[0057] The received movement state data (in short: movement data) can comprise sensor data and / or control data of the X-ray device. The movement state data received as sensor data (in particular, including measurement data) can also be referred to as actual movement state data. Alternatively or additionally, the actual movement state data can be assigned to a current and / or actual movement state. Furthermore, alternatively or additionally, the target movement state data can be assigned to a planned movement state. Alternatively or additionally, the target movement state data can be received as control data for the X-ray device.

[0058] The motion state data can represent a motion state of the X-ray device and / or its components (C-arm). Alternatively or additionally, the motion state data can include a position and / or orientation of the pivotable arm. Furthermore, alternatively or additionally, the motion state data can include a speed and / or acceleration of a change in position and / or orientation of the pivotable arm.

[0059] When using the ILC, for example, target motion state data can be received (e.g., directly). The target motion state data (also: target variable curves) can be changed in each repetition of the process steps (and / or in each iteration). Alternatively or additionally, the actual motion state data (and / or actual motion states) can be indirectly incorporated into the achieved quality measure (and / or not be explicitly used and / or determined in the process).

[0060] When using a KAF, both actual motion state data (and / or actual variables) and target motion state data (and / or target variables) can be used, in particular for determining the control signal and / or the quality measure.

[0061] The shadow of the X-ray grid can be determined by a shifted position (also: relative shift) of the detector relative to the X-ray source of the X-ray device depending on the received motion state data.

[0062] The position of the X-ray grid arranged on the detector, determined on the basis of the received motion state data, may comprise a relative displacement of the predetermined point (preferably the center point) of the X-ray grid relative to the focal point of the X-ray beam emitted by the X-ray source.

[0063] The motion state can also be referred to as the system state or device state (especially of the X-ray device).

[0064] The relative displacement can be caused by static and / or dynamic forces. For example, the pivoting arm of the X-ray device can comprise an elastically deformable connection (in particular a C-arm) between the X-ray source and the detector. Alternatively or additionally, the pivoting arm (and / or the X-ray device) can comprise an elastically deformable mount for the detector (in particular dependent on the state of motion), comprising the X-ray grid.

[0065] When the X-ray device moves, particularly the pivoting arm (especially the C-arm), deformations can occur due to acting forces (e.g., inertial forces of the different moving masses, Coriolis forces), which are represented in the motion state data. In particular, the pivoting arm (and / or the connection) can deform elastically if it deviates from its rest position, for example, due to gravity (also: gravitational force; in short: gravitation). Alternatively or additionally, inertia, centrifugal force, and / or Coriolis force can cause deformation of the pivoting arm (and / or the elastically deformable connection) of the X-ray source and detector.

[0066] Gravity can comprise a static, position-dependent force acting on the pivoting arm and / or the X-ray device, in particular its detector and X-ray source. Alternatively or additionally, dynamic forces (e.g., inertial force, centrifugal force, and / or Coriolis force) can act on the pivoting arm and / or the X-ray device, in particular its detector and X-ray source, in a movement-dependent manner (in particular, depending on the speed and / or acceleration of a change in position and / or orientation).

[0067] The quality measure can also be referred to as an error signal and / or quality signal. The quality measure reflects the quality of the X-ray image, particularly with regard to shadow minimization. In particular, a quality measure value of zero can correspond to an optimal (also: ideal) quality measure and / or optimal quality of the X-ray image. Alternatively or additionally, an optimal (e.g., minimal) value of the quality measure can correspond to a minimal shadow of the X-ray grid on the X-ray image.

[0068] Optimizing the quality measure may include minimizing the shadow (especially on an acquired x-ray image).

[0069] Determining the control signal (and / or optimizing the quality measure) may involve accessing a database. The database may include a table and / or a dictionary (also: dictionary; in technical terms: dictionary). The table and / or the dictionary may include a plurality of data sets (also: - in particular vectorial - data points). For each data set, motion state data of the X-ray device can be assigned to a control signal of the compensation actuator and a quality measure. For example, the determined quality measure can be assigned to a value of zero for the control signal (and / or to a non-activation of the compensation actuator).

[0070] The table and / or the dictionary may have been created based on measurement data. Alternatively or additionally, the table and / or the dictionary may be expanded during operation of the X-ray device. For example, after each acquisition of a (e.g., final) X-ray image, a data set (in particular comprising motion state data of the X-ray device, a—for example, optimized—quality measure, and optionally a control signal, especially if the optimization of the quality measure was non-trivial) may be added to the table and / or the dictionary.

[0071] Determining and outputting the control signal based on the motion state data of the X-ray device can also be referred to as data-driven learning control (especially of the compensation actuators).

[0072] By compensating (particularly actively) the shadow of the X-ray grid, image artifacts caused can be compensated and / or the quality of the X-ray image acquired (particularly in real time) can be improved.

[0073] The, or any, X-ray image (and / or any radiography) can be recorded as digital image data.

[0074] In one embodiment, the step of determining the quality measure can be performed for a test X-ray image. For example, a test X-ray image (and / or a Fourier-transformed test X-ray image) can be evaluated with regard to its quality measure. After compensating for the shadow (e.g., visible in the test X-ray image), a (e.g., final) X-ray image can be acquired.

[0075] In a further embodiment, the quality measure is determined based on the received motion state data, for example based on the table and / or the lexicon.

[0076] In a further embodiment, the step of determining the quality measure (for example for a test X-ray image) may comprise capturing an image using an environment camera.

[0077] Determining the quality measure based on the test X-ray image and / or the image taken with the environment camera can also be referred to as generating a ground truth.

[0078] The output control signal may include a target value for a current and / or voltage at the compensation actuator. Alternatively or additionally, the output control signal may include (e.g., updated and / or improved) target motion state data.

[0079] The process can be carried out cloud-based.

[0080] Alternatively or additionally, the method can be executed on a local processor, a central processing unit (CPU), a neuromorphic processor (NPU) and / or a graphics processor (GPU).

[0081] The pivoting arm can include a C-arm or be designed as such.

[0082] The pivoting arm (and / or the elastically deformable connection) between the X-ray source and the detector can also be referred to as a gantry and / or in particular comprise a C-arm.

[0083] The X-ray device (e.g., with the elastically deformable connection configured as a pivoting arm, particularly as a C-arm) can comprise an M-dimensional space of (particularly independent) degrees of freedom, for example, with M=5 for a C-arm. Alternatively or additionally, the motion state, for example of the C-arm, can include joint angles of axes, velocities of axes, and / or accelerations of axes.

[0084] According to one embodiment, discrete values ​​of the degrees of freedom and / or the state of motion may be predetermined.

[0085] The position of the X-ray grid arranged on the detector (e.g. fixed) may include a position in a detector plane and / or in a plane perpendicular to a focus direction of the X-ray source of the X-ray device.

[0086] The position of the X-ray grid (and / or detector) can be determined relative to a center point (and / or focal point) of an X-ray beam emitted by the X-ray source.

[0087] The position of the X-ray device and / or the pivoting arm may include a location and / or orientation in space.

[0088] The movement state can comprise a position, in particular an orientation in space, and optionally a speed of a position adjustment (in particular per degree of freedom) of the pivotable arm and / or the X-ray device.

[0089] According to one embodiment, the motion state can be two-dimensional (2D), e.g. per degree of freedom, and / or include as coordinates position and speed (e.g. as position change per unit time) of the X-ray device (e.g. the C-arm).

[0090] According to a further embodiment, the movement state may comprise discrete position signals (and / or positions) sampled at equidistant times and / or a series of one-dimensional (1D) coordinates (e.g. per, in particular independent, degree of freedom) corresponding to the position of the pivotable arm of the X-ray device (e.g. the C-arm) at a predetermined time in a time series.

[0091] The set of all positions, and optionally velocities, can also be referred to as the state space of the X-ray machine and / or the pivoting arm.

[0092] Using the time series, a time series analysis and / or time series prediction can be carried out, in particular without an independent measurement of a (e.g. generalized) speed or (e.g. generalized) acceleration.

[0093] The position of the X-ray grid, which depends on the state of motion of the X-ray device, can also be referred to as displacement in the detector plane (or displacement of the detector plane).

[0094] Alternatively or additionally, the motion state of the X-ray device can be parameterized by generalized coordinates (e.g., including angles) and / or generalized velocities (e.g., including angular velocities).

[0095] The received movement state data may include target movement state data and / or actual movement state data (e.g., recorded, in particular measured, by means of sensors).

[0096] The target movement state data may include movement state data determined in advance and / or (in particular in advance and / or “offline”) control signals for the X-ray device.

[0097] The actual motion state data can include motion state data determined based on sensor data (e.g., actual and / or real). For example, a position and / or a speed can be measured. Alternatively or additionally, an acceleration can be determined based on a (e.g., measured) motor current of an actuator of the pivotable arm (in particular, the C-arm).

[0098] Compensating for the shadow may involve tilting and / or shifting the detector (particularly with an X-ray grid attached thereto). Alternatively or additionally, compensating for the shadow may involve tilting and / or shifting the X-ray grid. Furthermore, alternatively or additionally, compensating for the shadow may involve tilting and / or shifting the X-ray source.

[0099] The compensation actuator system can comprise two compensation actuators. A first compensation actuator can be configured to tilt and / or shift the detector and / or the X-ray grid along a first direction of a detector plane. A second compensation actuator can be configured to tilt and / or shift the detector and / or the X-ray grid along a second direction of a detector plane.

[0100] The first direction may be different from the second direction. In particular, the first direction may be perpendicular (90°) to the second direction.

[0101] Tilting along a direction can involve rotation around an axis. The rotation axis (or axis for short) can extend along the (e.g., x- or y-) direction. Tilting can change a detector plane.

[0102] In one embodiment, a relative displacement of the X-ray source and detector, for example, per direction, of 5 mm to 10 mm can be compensated (particularly actively) by means of the compensation actuator. Alternatively or additionally, in one embodiment, tilting can comprise a rotation of up to 10° (and / or up to π / 18), in particular between 5° and 10° (and / or between π / 36 and π / 18), for example per direction.

[0103] A suitable numerical range in which the compensation actuator operates may depend on the geometry of the pivoting arm (especially the C-arm) and / or the X-ray device.

[0104] Alternatively or additionally, the displacement along one direction may comprise a parallel displacement in the detector plane.

[0105] The method may further comprise a step of outputting the recorded X-ray image comprising compensating, in particular in real time, the shadow of the X-ray grid.

[0106] Outputting the acquired X-ray image may include an artificial intelligence (AI)-based image reconstruction (e.g., using a neural network, NN). Alternatively or additionally, the output X-ray image may include a "calculation out" and / or subtraction of the X-ray grid (and / or a pure image of the grid) from the acquired X-ray image (especially with a patient placed between the X-ray source and the detector).

[0107] The quality measure can comprise one quality measure per direction, in particular along the detector plane. The quality measure can comprise two (in particular independent) values ​​for the first direction of the detector plane and for the second direction of the detector plane. Alternatively or additionally, the quality measure per direction of the detector plane can be used to determine a control signal for a compensation actuator (e.g., the first and / or second).

[0108] Determining the quality measure can comprise a comparison with one or more data sets in a lexicon. One or each data set in the lexicon can comprise a motion state, a quality measure, and optionally a control signal for compensating for the shadow of the X-ray grating. Alternatively or additionally, one or each data set in the lexicon can comprise a temporal sequence (also: time series) of motion states, quality measures, and optional control signals. The temporal sequence can comprise a (e.g., predetermined) number of time steps. The (e.g., predetermined) number of time steps can also be referred to as time embedding. Alternatively or additionally, a history of the relevant data, namely the motion state and associated quality measure, and optionally the control signal, in particular for each time step, can be stored in the lexicon.

[0109] Alternatively or additionally, the lexicon may include control signals defined as “zero”, where the compensation actuator (e.g. corresponding per direction of the detector plane) is not (in particular not actively) controlled.

[0110] Determining the quality measure may comprise evaluating contrasts in an X-ray image, in particular in a Fourier-transformed X-ray image.

[0111] The Fourier-transformed X-ray image, for example, can comprise a predetermined (in particular periodic) pattern of amplitude maxima (and / or maximum brightnesses; in technical terms also: peaks) and amplitude minima (and / or minimum brightnesses) of one or more low-frequency fundamental waves of the X-ray beam. Deviations from the predetermined (in particular periodic) pattern (also: periodicity) and / or attenuations of contrasts between amplitude minima and amplitude maxima in the Fourier-transformed X-ray image (and / or increasing homogeneity of the Fourier-transformed image, for example) can indicate a deterioration in the quality measure of the X-ray image. Alternatively or additionally, a sum over amplitudes of local maxima can be correlated with (e.g. inversely proportional to) the quality measure. Furthermore, alternatively or additionally, a reduction in a number of characteristic frequencies can indicate a deterioration in the quality measure.

[0112] The Fourier transform can include a Fast Fourier Transform (FFT).

[0113] The quality measure can be measured by means of a displacement vector using phantom images (and / or calibration images).

[0114] A phantom image and / or a calibration image can be an X-ray image taken of a phantom or object with known image properties using the X-ray grid, in particular without a patient between the X-ray source and the detector.

[0115] Alternatively or additionally, the quality measure can be determined (and / or calculated) from the image data after an evaluation to be determined (e.g., during development), particularly as a formula and / or algorithm. During development, phantom images (and / or phantoms) can be used to develop (and / or design) the determination specification (and / or calculation specification, particularly the formula and / or algorithm). Alternatively or additionally, the ground truth and / or the expected image depending on the phantom used may be unknown during the development phase.

[0116] Alternatively or additionally, phantom images (and / or phantoms) should be sufficiently accurate that the results should also be satisfactory in clinical use. Alternatively or additionally, an extension of the inventive technology (e.g., at least theoretically) includes the possibility of applying the inventive method for improvement or tracking even during device operation.

[0117] The displacement vector can include and / or parameterize a (particularly actual) deviation of the focal point on the X-ray grid compared to calibration data without outward migration. The partial shadowing resulting from the displacement due to "incorrect" angles compared to the design of the smart grid's louvres can be at least partially compensated by a tilted position.

[0118] The quality measure can be used as the basis for controlling the compensation actuators. Alternatively or additionally, the quality measure can be learned (e.g., based on the motion state data). Furthermore, alternatively or additionally, the quality measure can be adapted (especially optimized), particularly during the step of determining the control signal.

[0119] At least the steps of determining the quality measure of the X-ray image and determining the control signal for the compensation actuator can be performed repeatedly and / or iteratively. Preferably, the repetitions and / or iterations are performed until the determined quality measure lies within a predetermined value range (and / or a predetermined minimum quality of the X-ray image is achieved after compensation).

[0120] Alternatively or additionally, the step of receiving motion state data, outputting the determined control signal, and / or (particularly in the case of end-to-end learning) outputting the acquired X-ray image can be carried out repeatedly and / or iteratively.

[0121] Determining a control signal for the compensation actuator may include machine learning (ML) and / or reinforcement learning (RL).

[0122] Reinforcement learning (RL) can involve a software agent and one or more reward signals. The reward signals can correspond to an optimization of the quality measure.

[0123] A behavior (technically known as a policy), especially in RL, can comprise a collection of probability measures. The policy can be learned iteratively (also: stepwise). For example, an optimization of a quality measure can be learned iteratively using motion state data and control signals. Alternatively or additionally, a control signal can be learned iteratively based on the motion state data and (for example) a predetermined quality measure.

[0124] The RL can comprise one or more parameters of a control law as an action space. Alternatively or additionally, a state space (e.g., for the ML and / or RL) can be continuous. Furthermore, alternatively or additionally, the action space can be continuous.

[0125] Policy learning can also be referred to as "policy search" and / or involve an actor-critical algorithm. Alternatively or additionally, policy learning can involve proximal policy optimization (PPO) (e.g., optimization of the performance measure).

[0126] The Actor-Critic algorithm can be an Asynchronous Advantage Actor Critic (A3C) algorithm.

[0127] Alternatively or additionally, the ML and / or the RL can include concurrent learning, in particular to enable adaptation to unknown influencing factors (which can also be referred to as disturbances) on the shadow of the X-ray grating.

[0128] The disturbance may, for example, include a tilted and / or skewed attachment of the detector to the pivoting arm (particularly to the C-arm).

[0129] Alternatively or additionally, the disturbance may include wear and / or an aging process of the X-ray device.

[0130] Alternatively or additionally, the ML and / or the RL can include fuzzy logic.

[0131] Using the ML and / or the RL, an adjustment of target variables (e.g., a shift relative to a calibration image) can be learned using the control signals. Alternatively or additionally, one or more target variables can include the quality measure (particularly per direction of the detector plane).

[0132] The determination of the control signals (for example by means of the ML and / or RL) can be adaptive and / or repetitive.

[0133] Determining the control signal can involve iterative learning control (ILC).

[0134] The ILC is particularly suitable in cases where disturbances are periodic and / or repeatable.

[0135] In ILC, a control signal at one (particularly discrete) point in time can be described by a control signal and an error signal (also: quality signal and / or quality measure) at the previous (particularly discrete) point in time using a first filter (e.g., designated Q), which acts on both the control signal and the error signal, and a second filter (e.g., designated L), which acts on the error signal. For example, the filters Q and L can act multiplicatively on the error signal.

[0136] The ILC enables acausal filtering, especially since filters and / or control signals are changed from one iteration step to the next.

[0137] The ILC can include a control of a target variable (e.g., the control signals for the compensation actuator to achieve the optimal quality factor). Alternatively or additionally, an actuator of the compensation actuator can include an electric drive (technically known as a motor controller). The target variable can include a current and / or a voltage of the actuator's electric drive.

[0138] The ILC is particularly suitable for repeatable (and / or nonlinear) disturbances.

[0139] Alternatively or additionally, a basic concept model and / or a first principle model (technically: First Principle Model) can linearize a device dynamics and / or system dynamics (e.g. of the X-ray device) at (particularly suitably chosen) points, and / or an H ∞ -Design can be used.

[0140] H ∞Methods can be used in control engineering to synthesize robust stabilizing controllers through mathematical optimization that achieve robust performance with respect to weighted transfer functions (see, for example, Geering et al. in [1]).

[0141] The first-principle model can be created starting from a basic physical understanding (and / or based on basic physical principles) (e.g. regarding conservation of energy and / or inertia).

[0142] The control signal can be determined based on calculations (e.g. the quality measure and / or, in particular high-frequency, error signals) from image data.

[0143] In each embodiment of the inventive technique, a time-continuous signal waveform can be discretized (in particular as sampling at equal intervals, for example with sampling steps k).

[0144] Determining the control signal may include adaptive kernel identification feedforward extrapolation.

[0145] The adaptive kernel identification forward extrapolation may comprise learning a relationship between the control signal and the quality measure (and / or error signal) by a kernel adaptive filter (KAF), in particular by means of a time embedding.

[0146] Adaptive kernel identification forward extrapolation can be particularly suitable when the formation of the X-ray grid's shadow is dominated by highly nonlinear effects. Alternatively or additionally, adaptive kernel identification forward extrapolation is online-capable and / or can be applied in real time during X-ray image acquisition. Furthermore, alternatively or additionally, adaptive kernel identification forward extrapolation does not require repetitive and / or periodic signal waveforms.

[0147] In particular, due to the “kernel trick” commonly used in KAF, in which a determination of a similarity of defined points in a feature space using inner products in the feature space is used by evaluating a positive definite kernel, the Adaptive Kernel Identification Forward Extrapolation enables a more efficient and / or improved performance compared to conventional X-ray grating shadow compensation techniques.

[0148] The application of the “kernel trick” to the inventive solution of control signal generation for suppressing image artifacts is not conventionally known.

[0149] Training data of the KAF may include data sets in which the control signal is zero.

[0150] The adaptive kernel identification forward extrapolation may comprise an interpolation method in which a first calibration image is shifted stepwise (and / or iteratively) and a deviation between the current X-ray image and the calibration image is determined, wherein the quality measure comprises the amount of a shift with respect to a minimum deviation.

[0151] Using MR, RL, ILC and / or Adaptive Kernel Identification forward extrapolation, generalization to previously unseen data (and / or data not used for training) can be achieved.

[0152] Previously unseen data may include the fact that not all possible movements are (and / or cannot be) performed to generate the training dataset. Alternatively or additionally, a predetermined (and / or finite) number of data (and / or datasets) can be used, e.g., one hundred (100) different (and / or discrete) positions can be approached. The previously unseen data can include continuous positions (especially unequal to the approached positions) up to any number of positions up to the quantization resolution in the decimal place.

[0153] Using previously unseen data and (especially nevertheless) good generalization capability, a good result of the inventive technique for compensating the X-ray grid shadow can be achieved, especially if similar data have not yet been available for training.

[0154] Alternatively or additionally, the previously unseen data may be attributed to wear or other effects occurring over a (particularly long) period of time (and which, for example, change the motion state data). For example, gear properties (e.g. of the pivoting arm) may change due to altered friction coefficients, which may, for example, be temperature-dependent. Alternatively or additionally, the temperature in the gear may depend on how long the X-ray machine was in continuous use and / or whether a room in which the X-ray machine is located is heavily air-conditioned. The expectation of a good algorithm (particularly MR, RL, ILC and / or Adaptive Kernel Identification Forward Extrapolation) may include that the respective algorithm delivers usable return values ​​(e.g. for the quality measure and / or control signal) despite using previously unseen (particularly motion state) data.

[0155] A KAF can incorporate a kernel-based learning method. Alternatively or additionally, Gaussian Process Regression and / or Support Vector Machine (SVM) can be used to determine the control signal.

[0156] A CAF can be particularly suitable for solving nonlinear regression problems. Alternatively or additionally, a nonlinear function can map an input to an output by transforming the input into a high-dimensional feature space H.

[0157] For example, a Gaussian kernel, a recursive least squares method in the kernel (technically known as kernel recursive least squares, KRLS) and / or kernel least-mean squares (technically known as kernel least-mean squares, KLMS) can be applied.

[0158] In each embodiment, a function can comprise a sum over weighted kernels. The weights (and / or coefficients of the individual kernels) can be calculated at runtime. Alternatively or additionally, a lexicon can comprise training points (and / or weights), and optionally data points and / or changed weights added at runtime (e.g., due to newly occurring target size changes).

[0159] In one embodiment, a cardinality (and / or number of records) of the lexicon may be limited to enable execution of the method with a limited memory capacity, limited computing capacity and / or limited computing time.

[0160] A KAF can comprise a function that maps a (e.g., current) motion state and a control signal to a quality measure. Alternatively or additionally, the application of the KAF and / or the adaptive kernel identification forward extrapolation can comprise an inversely learned function to predict the control signal (particularly based on the quality measure and motion state).

[0161] In each embodiment, the motion state data, the quality measure, and / or the control signal can be sampled in a time-discrete manner, and time embedding can be applied over multiple time steps to account for dynamic effects. A prediction horizon can each be assigned to the motion state data, the quality measure, and / or the control signal, in particular up to an order of the temporal embedding of the motion state data, the quality measure, and / or the control signal.

[0162] An order of the temporal embeddings can be chosen independently (and / or differently) for the motion state data, the control signal and the quality measure.

[0163] The independent (and / or different) choice of the orders of the temporal embeddings has the advantage that the dimension can be kept as small as necessary. For physical quantities, the order can be two (2), for example, especially since the underlying differential equations are second-order. For the dynamics of the quality measure, the order can be chosen to be at least five (5), in particular 25.

[0164] The KAF may be ignorant of whether a quantity is physical or not. Alternatively or additionally, a developer (and / or designer) may know which quantities are physical (and / or which quantities are non-physical) and use this knowledge to specify the order (e.g., as small as possible, as large as necessary). For motions, e.g., when physical quantities from position to acceleration are relevant, second (2nd) derivatives may be relevant and / or order 2 may be specified (especially by the developer and / or designer).

[0165] In a further embodiment, the order of the temporal embeddings can be chosen identically for the motion state data, the control signal and / or the quality measure.

[0166] The predictive ability of the KAF can be determined by an error measure, for example, the mean squared error (MSE), between given initial data points and the output data points estimated using the trained KAF for identical input data points. For example, the predictive ability can be better the smaller the error measure, for example, the MSE.

[0167] Alternatively or additionally, the KAF should only be able to estimate and / or predict the appropriate output date as accurately as possible based on a new input date.

[0168] A predictive skill can be associated with one or more steps into the future. For example, a predictive skill can decrease with the number of steps to be predicted (e.g., starting from a fixed point in time and / or time step).

[0169] The prediction horizon may be one and / or include the next possible sampling step. Alternatively or additionally, the prediction horizon may include an integer greater than one, for example, ten (10), particularly if no determination of a control signal is provided (and / or not possible) in each sampling step.

[0170] Adaptive kernel identification forward extrapolation can be performed iteratively (and / or stepwise), particularly in real time and / or during an X-ray acquisition process.

[0171] The choice of KAF algorithm can be changed during the lifetime of the X-ray machine. For example, the recursive kernel least squares (KRLS) method can be applied initially. Alternatively or additionally, a forgetting mechanism can be introduced if the X-ray machine deteriorates due to wear and / or aging.

[0172] Tuning hyperparameters of the KAF can influence the learning behavior of the chosen algorithm (e.g., improve convergence and / or avoid over-fitting), for example, when a periodicity is known and / or a size of expected deviations.

[0173] Adaptive kernel identification forward extrapolation enables multivariable control, for example tilting along M ≥ 2 degrees of freedom using M parallel (and / or independent) KAFs, e.g. separately along the (x,y) directions of the detector plane.

[0174] Conventional KAF-based inverse control methods are based on reference models, which, however, are not applicable to imaging applications (especially X-rays). Other conventional KAF-based methods do not use reference models at all. Alternatively or additionally, the inventive technique differs from conventional KAF-based methods through novel generalization capabilities for extrapolating new control signals.

[0175] Determining the control signal may include end-to-end learning.

[0176] According to one embodiment, a controller of the compensation actuator can learn rewards (in particular directly) from one or more X-ray images.

[0177] End-to-end learning can include deep reinforcement learning and / or a deep convolutional neural network (deep CNN).

[0178] A label (and / or quality measure) of an X-ray image can be determined based on a reference image and / or based on an assessment, in particular a quality or goodness assessment, by an expert (e.g. a radiologist).

[0179] In a further development of the invention, a control signal can be determined based on a quality measure (and / or label) and without acquiring motion state data. Alternatively or additionally, the generalizability of the inventive technique can be improved by end-to-end learning.

[0180] According to one aspect of the device, a computing device is provided for controlling a compensation actuator for compensating a shadow of an X-ray scattering absorption grid in an X-ray image recorded with a pivotably mounted X-ray device, in particular in real time.

[0181] The computing device comprises a receiving interface configured to receive motion state data indicating a motion state of an X-ray device. The X-ray device is mounted on a pivotable arm. The motion state is associated with a time at which an X-ray image is acquired by the X-ray device.

[0182] The computing device further comprises a quality measure determination unit which is designed to determine a quality measure of the X-ray image with respect to a shadow of an X-ray scattering absorption grid which is arranged on the detector of the X-ray device.

[0183] The computing device further comprises a control signal determination unit configured to determine a control signal for a compensation actuator for adjusting the position of the X-ray scattering absorption grating depending on the received motion state data and the determined quality measure. Determining the control signal includes optimizing the quality measure.

[0184] The computing device further comprises a control signal output interface configured to output the determined control signal to the compensation actuator for compensating the shadow of the X-ray scatter absorption grating.

[0185] Optionally, the computing device comprises an X-ray image output interface configured to output the acquired X-ray image including the compensation, in particular in real time, of the shadow of the X-ray scatter absorption grid.

[0186] The computing device may be configured to perform one or more steps of the method according to the method aspect. Alternatively or additionally, the computing device may comprise one or more features disclosed in connection with the method according to the method aspect.

[0187] According to one system aspect, a system for controlling a compensation actuator for compensating a shadow of an X-ray scatter absorption grid in an X-ray image recorded with a pivotably mounted X-ray device, in particular in real time, is provided.

[0188] The system comprises an X-ray device mounted on a pivotable arm with an X-ray scatter absorption grid. The system further comprises a computing device according to the device aspect. The system further comprises a compensation actuator. A receiving interface of the compensation actuator is configured to receive the specific control signal from the control signal output interface of the computing device.

[0189] The system may be configured to execute the method according to the method aspect.

[0190] According to a further aspect, a computer program product is provided. The computer program product comprises program elements that cause a computing device to execute the steps of the method for controlling a compensation actuator for compensating a shadow of an X-ray scatter absorption grid in an X-ray image acquired with a pivotably mounted X-ray device, in particular in real time, according to the method aspect when the program elements are loaded into a memory of the computing device.

[0191] According to yet another aspect, a computer-readable medium is provided on which program elements are stored that can be read and executed by a computing device to carry out steps of the method for controlling a compensation actuator for compensating a shadow of an X-ray scatter absorption grid in an X-ray image that is recorded with a pivotably mounted X-ray device, in particular in real time, according to the method aspect when the program elements are executed by the computing device.

[0192] The above-described properties, features, and advantages of the present invention, as well as the manner in which they are achieved, will become clearer and more understandable in light of the following description and the exemplary embodiments explained in more detail in conjunction with the drawings. This following description does not limit the invention to the embodiments included. Like components or parts may be provided with the same reference numerals in different figures. In general, the figures are not to scale.

[0193] It is understood that a preferred embodiment of the present invention may also be any combination of the dependent claims or the above embodiments with the respective independent claim.

[0194] These and other aspects of the invention will become apparent from and will be explained by reference to the embodiments described below. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a flowchart of a method for controlling a compensation actuator for compensating a shadow of an X-ray grid in an X-ray image acquired with a pivotably mounted X-ray device, in particular in real time, according to a preferred embodiment of the present invention; Fig. 2 is an overview of the structure and design of a computing device for controlling a compensation actuator for compensating a shadow of an X-ray grid in an X-ray image acquired with a pivotally mounted X-ray device, in particular in real time, according to a preferred embodiment of the present invention; Fig. 3A and Fig. 3B show a conventional X-ray grid with slats perpendicular to a base surface or a so-called "smart grid" with slats aligned towards an X-ray source, wherein the method according to the invention is suitable for compensating the shadow of the "smart grid" in particular; Fig. 4 and Fig. 5 show a prior art technique for avoiding a shadow of an X-ray grid by means of additional actuators; Fig. 6, Fig. 7 and Fig. 8 show schematic embodiments of the inventive technique in which, for example, iterative learning control (ILC), adaptive kernel identification forward extrapolation, reinforcement learning (RL) or end-to-end learning can be used to implement a function approximator; Fig. Figure 9 shows schematically the basic idea of ​​the “kernel trick” for determining data points; Fig. 10, Fig. 11 and Fig. 12 show embodiments of an agent that is used when using an RL as a function approximator; Fig. 13 schematically shows an iterative development of a control signal and a quality measure when an iterative learning controller (ILC) is used; Fig. Figure 14 schematically shows a state-of-the-art KAF control system operating with a reference model and without (in particular unknown) disturbances; Fig. 15 shows an X-ray image of a honeycomb X-ray grid, in particular without compensation; Fig. 16 shows a 2D FFT image of an X-ray image taken with an X-ray grid on the detector; and Fig. 17A, Fig. 17B and Fig. 17C shows an example of subtracting a calibration image from an acquired X-ray image to calculate the X-ray grid from the output X-ray image.

[0195] Any reference signs in the claims are not to be understood as limiting the scope of application.

[0196] Fig. 1 schematically shows an exemplary flowchart of a (particularly computer-implemented) method for controlling a compensation actuator for compensating for a shadow of an X-ray scatter absorption grid (X-ray grid for short) in an X-ray image acquired with a pivotably mounted X-ray device, particularly in real time. The method is generally designated by reference numeral 100.

[0197] The method 100 includes a step S102 of receiving motion state data indicating a motion state of the X-ray device. The X-ray device is mounted on a pivotable arm. The motion state is associated with a time at which the X-ray image is acquired by the X-ray device.

[0198] The method 100 further comprises a step S104 of determining a quality measure of the X-ray image with respect to a shadow of the X-ray grid arranged on the detector of the X-ray device.

[0199] The method 100 further comprises a step S106 of determining a control signal for a compensation actuator for adjusting the position of the X-ray grid depending on the received S102 motion state data and depending on the determined S104 quality measure. Determining S106 of the control signal comprises optimizing the quality measure.

[0200] The method 100 further comprises a step S108 of outputting the determined S106 control signal to the compensation actuator for compensating the shadow of the X-ray grid.

[0201] Optionally, the method 100 comprises a step S110 of outputting the recorded X-ray image comprising the compensation (e.g. by means of the determining step S106), in particular in real time, of the shadow of the X-ray grid.

[0202] At least some of the method steps S102, S104, S106, and S108 (and optionally S110) can be executed repeatedly, in particular iteratively. This allows the control of the compensation actuators to be successively improved.

[0203] An example is Fig. 1 schematically shows that steps S104 and S106 of determining the quality measure of the X-ray image and the control signal for the compensation actuator, respectively, can be performed repeatedly. For a specific control signal S106, the (e.g., expected) quality measure can be determined again S104.

[0204] In another in Fig. 1, after a sample X-ray image has been output S110, new motion state data can be received S102 and the determination S104 and S106 of the quality measure of the X-ray image or the control signal for the compensation actuator can be repeated.

[0205] Fig. Figure 2 schematically shows an exemplary architecture of a computing device for controlling a compensation actuator for compensating for a shadow of an X-ray grid in an X-ray image acquired with a pivotably mounted X-ray device, in particular in real time. The computing device is generally designated by reference numeral 200.

[0206] Computing device 200 includes a receiving interface 202 configured to receive motion state data indicating a motion state of the X-ray device. The X-ray device is mounted on a pivotable arm. The motion state is associated with a time at which the X-ray image is acquired by the X-ray device.

[0207] The computing device 200 further comprises a quality measure determination unit 204, which is designed to determine a quality measure of the X-ray image with respect to a shadow of the X-ray grid arranged on the detector of the X-ray device.

[0208] Computing device 200 further includes a control signal determination unit 206 configured to determine a control signal for a compensation actuator for adjusting the position of the X-ray grid depending on the received motion state data and the determined quality measure. Determining the control signal includes optimizing the quality measure.

[0209] The computing device 200 further comprises a control signal output interface 208, which is designed to output the determined control signal to the compensation actuator for compensating the shadow of the X-ray grid.

[0210] Optionally, the computing device 200 comprises an X-ray image output interface 210, which is designed to output the acquired X-ray image comprising the compensation (in particular in real time) of the shadow of the X-ray grid.

[0211] The receiving interface 202, the control signal output interface 208, and optionally the X-ray image output interface 210, can be arranged in an input-output interface 212.

[0212] The quality measure determination unit 204 and the control signal determination unit may be embodied by a processor 214 (for example, comprising a CPU and / or a GPU).

[0213] Computing device 200 may be configured to execute method 100. For example, receiving interface 202, quality measure determination unit 204, control signal determination unit 206, control signal output interface 208, and optionally x-ray image output interface 210 may be configured to execute steps S102, S104, S106, S108, and optionally S110.

[0214] A system (not shown) may include the computing device 200 and an X-ray device mounted on a pivotable arm, including an X-ray grid, and a compensation actuator. A receiving interface of the compensation actuator may be configured to receive the specific control signal from the control signal output interface 208 of the computing device 200.

[0215] Fig. 3A shows a schematic example of a conventional X-ray grid with perpendicular (and / or mutually parallel) blades 312. The X-ray grid includes a height, designated by reference numeral 308, and a width, or diameter, designated by reference numeral 310, of a base area 306. Starting from the X-ray source 302, the X-ray beam fans out, as shown schematically at reference numeral 304.

[0216] In the case of the conventional X-ray grid of Fig. 3A is the shadow caused by the (e.g. outer) lamellae independent of the position of the X-ray grating in the detector plane parallel to the base surface 306.

[0217] Fig. Figure 3B shows an example of a “smart grid” (also known as a focused grid). The same dimensions as in Fig. 3A are designated by identical reference numerals. Reference numeral 316 indicates, by way of example, a maximum angular deviation (e.g., 20° and / or π / 9) from a perpendicular of an outer lamella 314.

[0218] In the case of the “smart grid” of the Fig. 3B, ideally (in particular, without displacement and / or migration of the detector), no shadowing occurs. Due to a displacement of the detector relative to the X-ray source 302, shadows of varying intensity (depending in particular on the state of motion of the X-ray device) can occur from the slats 314 of the X-ray grid.

[0219] Fig. 4 shows an exemplary conventional method 400 in which an elastostatic model of the dynamic effects is created and calculated at runtime in order to correct the relative displacement between the radiator and detector by means of correction drives and a gimbal suspension.

[0220] At reference numeral 402, a trajectory (also: movement sequence and / or motion state curve) of a C-arm (as an example of a pivoting arm) is specified. At reference numeral 404, a real-time capable (e.g., elastostatic) model of the C-arm's mechanics is created and / or applied. At reference numeral 406, deformations of the C-arm are determined (e.g., calculated). At reference numeral 408, the relative displacement between the radiator and detector resulting from the deformation is determined. At reference numeral 410, a correction of the relative displacement is determined by two correction drives (and / or compensation actuators).

[0221] Fig. Figure 5 shows components used in the conventional process 400 of Fig. 4 can be used. Reference numeral 502 schematically shows a user interface of a computer (and / or a computing device). Reference numeral 504 shows, by way of example, a circuit board (and / or a mini-PC) of the computer (and / or the computing device) or an external computing device on which the real-time capable (e.g. elastostatic) model of the mechanics is executed or exported thereto (and / or a code is generated) 508. In the conventional (e.g. elastostatic) model of the mechanics, dynamic equations are generated from a simulation model and calculated on the mini-PC (and / or the circuit board) 504 at runtime, and comparisons are carried out with a control of one or more control units (e.g. programmable logic controllers, PLCs) 506 of the correction drives (and / or correction actuators) 512, as shown schematically at reference numerals 402 and 406.

[0222] By means of the computer (and / or computing device) 502, as indicated by reference numeral 510, the control is projected onto the control units (e.g., PLC) 506. The control unit (e.g., PLC) 506 can operate a subordinate control of one or more motors. A target value for the control unit (e.g., PLC) 506 is typically calculated by another control unit (e.g., in the schematic computer 502).

[0223] Fig. 4 and Fig. Figures 5 show a simple prototype of an actively actuated mechanism for tracking the X-ray grid during operation. The X-ray grid, which is attached to the detector, is mounted so that it can be tilted in two axes, allowing fine positioning to kinematically compensate for displacements that may occur, for example, due to the elasticity of the C-arm.

[0224] The conventional solutions, as shown schematically in Fig. 4 and Fig. 5, the basic procedure is based either on static measurements at support points and storing values ​​in a table (and / or a position-related calibration) or on model-based calculation of the control commands for the correction drives.

[0225] The second variant of the model-based calculation of the control commands is based on the creation (especially manually and / or by engineering) of an elastostatic model of the C-arm and the dynamic effects, which is capable of real-time calculation and is intended to calculate the relative displacement between the radiator and detector in order to compensate for the displacement by appropriately controlling the correction drives.

[0226] Due to its principle, the conventional method has the disadvantage that the achievable compensation accuracy for correcting the supergrid depends on the modeling quality of the underlying model. However, it is non-trivial to calculate real-time, yet highly accurate models of the respective deformation, so approximations and model assumptions or model simplifications must be used. Furthermore, there are effects on the real X-ray device that are not, or only with great difficulty, taken into account by the model and can therefore impair the validity of the results, e.g., if the smart grid is attached to the detector slightly differently than assumed in the model due to manufacturing and / or assembly tolerances.Preliminary evaluations have already shown that such effects (particularly the precision of the mechanical attachment of the X-ray grid to the detector) have a significant influence on the achievable quality of an X-ray image and can significantly worsen the result.

[0227] In particular, in contrast to conventional techniques, the inventive technique is (e.g. as far as possible) independent of the concrete hardware system architecture.

[0228] Alternatively or in addition to a PLC-based controller for the compensation actuators, an embedded electronics board can be used, on which the real-time control of the pivoting arm (especially the C-arm) and / or the control of the smart grid's correction actuators can be executed (especially simultaneously and / or in parallel). Alternatively or in addition, a GPU or a GPU-like component can be used to enable fast computing of a NN.

[0229] By means of the technology according to the invention (for example comprising the method 100 and / or the computing device 200), active tracking of the X-ray grid (and / or raster) is enabled (e.g., possibly necessary due to radiation physics). While it has conventionally been shown that active tracking of the X-ray grid is fundamentally possible (in particular mechanically and / or in terms of control technology) and enables the use of a (high-aspect) super-raster, the question of high-precision compensation of the radiator displacement and / or detector displacement continues to be a lever for more precisely compensating for shadowing, enabling the use of super-rasters with higher aspect ratios, and thus improving the achievable image quality with a reduced radiation dose. A data-driven learning control system also enables further advantages according to the invention.

[0230] In the technology according to the invention (for example comprising the method 100 and / or the computing device 200) for calculating the control signals for the actuators for compensation, a data-driven method is used which, as a quality measure, is based directly on the achieved image quality as the ultimately relevant variable (and, for example, not on an assumed shift of the beam focus, which indirectly influences the image quality).

[0231] The following symbols and sizes are used below: n ∈ ℕ: number of movable degrees of freedom of the pivoting arm and / or gantry, e.g. of a C-arm (with n ≥ M, where M is the number of independent degrees of freedom); q ∈ ℝ n : (actual) joint angles of the axes and / or generalized coordinates; q ∈ ℝ n : (Actual) speeds (and / or time derivatives) in generalized coordinates; q̈ ∈ ℝ n: (actual) accelerations in generalized coordinates; q d ∈ ℝ n : Target joint angles of the axes and / or generalized coordinates; q d ∈ ℝ n : Target speeds of the axes and / or generalized coordinates; y ∈ ℝ m : joint angle of the axes that actuate the tilt of the smart grid (usually with m = 2); y d ∈ ℝ m : Target joint angle of the axes that actuate the tilt of the smart grid; x¯=[qq˙]∈ℝ2n: (particularly dynamic and / or actual) movement state of the X-ray device (also: system); x¯d=[qdq˙d]∈ℝ2n: Target values ​​of the motion state; d ∈ ℝ nd : disturbances (especially unknown) that may result in a deterioration of image quality; (x, y) ∈ ℝ 2 : Coordinate directions in the detector plane; c i ∈ ℝ ≥0: Overall quality measure of image quality, with c i = 0 in case of perfect compensation of artifacts; c x ∈ ℝ ≥0 : Quality measure of the compensation in x-direction, where c x = 0 in case of perfect compensation of artifacts in this direction; c y ∈ ℝ ≥0 : Quality measure of the compensation in y-direction, where c y = 0 in case of perfect compensation of the artifacts in this direction; q[k] = q(kt s ): time-discrete signal obtained by sampling with the sampling time t s from the continuous signal q(t); and k ∈ ℕ0: discrete sampling time step.

[0232] One aim of embodiments of the inventive technology is to provide the control signals for the compensation actuators (and / or setpoints) y dto determine (and / or calculate) and output to the compensation actuators (and / or control unit of the smart grid) at runtime in such a way that the quality measure is optimized (in particular c i → 0), although (especially unknown) disturbing influences d act on the system.

[0233] In further embodiments of the inventive technology, the overall quality measure c i and / or the quality measures per direction of the detector plane (c x , c y ) can be determined (and / or calculated) from medical X-ray imaging data.

[0234] Fig. 6 shows a first embodiment of the method 100 according to the invention.

[0235] It can be referred to as an “intelligent method” because the control system and / or regulation system uses learning strategies and / or adaptation strategies, for example from the field of artificial intelligence (AI) (e.g. fuzzy logic, RL, NNs, and / or concurrent learning), in order to achieve adaptation to unknown influencing variables (cf. e.g. Schröder, Buss: Intelligent methods: identification and control of nonlinear systems, Springer 2017, [2]).

[0236] In one embodiment (for example as in Fig. 6) the following assumptions are made: - The positions of the degrees of freedom of the pivoting arm (and / or the gantry, especially the C-arm) q and their time derivatives q̇ are known at runtime. The positions q and velocities q̇ are provided by the pivoting arm's system controller, as are the target curves q d and q̇ d . The corresponding quantities become the state of motion x¯=[qq˙] summarized. - First, the compensation actuators of the smart grid are deactivated, i.e. y d = 0 (and / or no movement of the grid tracking). Depending on the movement state x, a more or less accurate X-ray image is obtained due to known or unknown interference d. Interference could include, for example, the deflection of the pivoting arm (especially the C-arm) and / or dynamic forces such as inertial or Coriolis forces, which cause the X-ray grid to no longer match the calibration, thus causing artifacts to become visible in the X-ray image. - Based on the image processing, a quantitative quality measure can be calculated that describes the overall quality of the obtained X-ray image with regard to the detectable artifacts due to the X-ray grid. The overall quality measure is defined as follows: i called, with c i= 0 in the case of perfect compensation of the artifacts. The quality measure can be described in the 2D X-ray image alternatively or additionally axis-dependently by the quantities c x and c y . The axis-dependent description is relevant, for example, when undesirable artifacts are present only in one dimension, e.g., horizontally, but no (or only a few) artifacts are visible vertically. The quality measures c i , c x and c y are provided, for example, in method step S104 and optimized in step S106. - In order to ensure deterministic performance and certifiability of the approach, it may be necessary not to use the (feedback) controller (e.g. the control of the swiveling arm, also: Gantry control, and / or the control of the smart grid (also known as smart grid control) through a learning component. It may be advantageous (especially instead) to leave the controller structure and parameters the same and only change the control signals.

[0237] In Fig. 6, reference numeral 602 indicates the movement state (also: system position and / or system movement) x and / or x dfor receiving in step S102 of the method 100. Reference numeral 604 shows a calculation of a control target variable. This calculation can take place in step S106 of the method 100. Reference numeral 606 schematically shows the compensation actuators of the smart grid. Reference numeral 608 shows an X-ray imaging process, including the smart grid, at the detector. The X-ray imaging is influenced by interference, as shown schematically at reference numeral 614. Image processing takes place at reference numeral 610, on the basis of which an X-ray image can be output as a useful medical image 612. Reference numeral 616 schematically shows the determination (and / or calculation) of the quality measure, which can take place, for example, in step S106 as optimization of the quality measure and / or in step S104, in this exemplary embodiment as feedback for an iterative optimization.

[0238] In the embodiment of the Fig. 6 the target values ​​y d the actuators that can actively move the supergrid are generated at runtime in such a way that the quality measure is optimized (e.g. c_i → 0).

[0239] The embodiment of the Fig. 6 can use different function approximators to optimize the quality measure and output one or more corresponding control signals.

[0240] A first embodiment of a function approximator includes an iterative learning control (ILC). ILCs (without applications for X-ray images) are known per se, see, for example, [3], [4].

[0241] The basic assumption of an ILC is that a repetitive task and / or a recurring disturbance is present. In the case of the X-ray machine mounted on a pivoting arm, this corresponds to the assumption that the disturbances d leading to image artifacts are distributed at a given x dare repeatable, although the curves may be non-linear.

[0242] Fig. Figure 13 shows an ILC schematic diagram based on a figure from [3]. Reference numeral 1302 indicates a sample time, reference numeral 1304 indicates a count of iteration steps, and reference numeral 1306 indicates a control signal y d (e.g., a time-varying amplitude of the control signal, as shown schematically per iteration). At reference numeral 1308, the quality measure c i (e.g., a time-varying amplitude of the quality measure). Reference numeral 1310 schematically shows a disturbance d, and reference numeral 1312 shows a reference image.

[0243] As in Fig. 13 at higher iteration steps, an adjustment of the control signal y is made from iteration to iteration by calculating the ILC d achieved, so that the quality measure c iis optimized (and / or decreases, e.g., c_i → 0). An iteration here refers to the repetitive task, i.e., a complete "sequence" of the control signal, e.g., moving a C-arm by 180 degrees for orbital 3D imaging.

[0244] In computer-implemented signal processing, the continuous-time signal waveforms of physical quantities are typically discretized. For example, the waveforms of the positions q(t) and / or velocities q̇(t) are sampled at times t0, t1, .... The discretized position signals {q[t]} = {q(t0), q(t1), ...} and / or velocity signals {q̇[t]} = {q̇(t0), q̇(t1), ...} are obtained as motion state data.

[0245] Since at a time interval t s equally sampled quantities are used, in short form q[k] = q(kt s). Alternatively or additionally, the control signals (and / or control variables) are calculated in a time-discrete manner, in particular by the computing device 200 (and / or a computer), and subsequently converted back into physically continuous variables (e.g., current, voltage) by a motor controller.

[0246] The simplest ILC adaptation law is given by ydi+1[k]=Q(ydi[k]+Lc1i[k]).

[0247] Here, ydi[t] the control signal magnitude curve in the i-th iteration of the repetitive task. The initial control signal yd0[t] can be calculated using a classic control design. Alternatively or additionally, yd0=0 be accepted. In case yd0=0 The complete control of the smart grid is learned through the ILC method.

[0248] The engineering task for applying the method is to calculate the filters Q and L. A model-based design can be used for this, advantageously using first-principle models and / or system models of the X-ray device (also known as the system). In this case, the dynamics (especially of the pivoting arm) of the X-ray device (also known as the system dynamics) can be linearized at suitable points, and a robust filter design (and / or a robustness design for the filters Q and L) can be used, e.g., comprising a ℌ ∞ Design. Common methods for this can be found in the literature on control engineering.

[0249] The use of classical ILC methods (e.g. from [3]) in technical applications may be difficult if the requirement of a quality measure c available in each sampling time step k is not met. i[k] cannot be fulfilled, since the determination and / or calculation is carried out from the X-ray image data and is a computationally intensive operation that cannot conventionally be carried out in the (especially fast) cycle of the control system (e.g. t s ≈ 1-10 ms). This technical difficulty can be circumvented by using more advanced ILC methods that can also operate with less high-frequency quality measures (and / or error signals). Such an advanced ILC method is described (and / or developed), for example, in [5].

[0250] ILC can be seen as a (particularly relatively simple) special form of a RL strategy for the case where a repetitive task is performed multiple times and the perturbations to the system are also repetitive and predictable (but especially unknown). If the assumptions do not hold, more general RL methods can be resorted to.

[0251] In another embodiment, adaptive kernel identification feedforward extrapolation is used as a function approximator. Adaptive kernel identification feedforward extrapolation is little or not published in the academic control literature and can best be derived from reference [6]. In particular, kernel-based adaptive filters (KAFs) are not known to solve the problem of control signal generation for suppressing image artifacts (e.g., in X-ray images).

[0252] In contrast to the ILC method, adaptive kernel identification forward extrapolation does not require repetitive and / or periodic signal waveforms (e.g., for the quality signal, the control signal, and / or the motion state data). The overall performance of adaptive kernel identification forward extrapolation can be better than ILC because, thanks to the use of the "kernel trick," the method learns efficiently and, on the other hand, enables generalization of relationships to previously unseen data (particularly using machine learning, ML).

[0253] For the inventive technique, a modification of a Kernel Adaptive Filter (KAF) [7] is used (e.g., a modification of a class of signal processing methods that are well described in [8] and based on ML) in order to advantageously extrapolate the effects occurring during operation (e.g., elastic deformation of the pivoting arm of the X-ray device due to static and dynamic forces, temperature dependences of gear behavior and / or wear) by means of the “kernel trick” and to introduce a counteraction into the control signal path.

[0254] KAF is well known (see e.g. [7], [8],

[10] ,

[11] ) and belongs to the kernel-based learning methods (which also include e.g. Gaussian Process Regression and / or Support Vector Machine, SVP) for solving nonlinear regression problems. This means that for an input / output stream of data points (x i , y i ) the nonlinear function f i(·) which corresponds to the relationship between input x i and output y i underlying, yi=ƒi(xi), and this estimate is updated with each new data pair.

[0255] Kernel trick based extrapolation methods (and described for example in [9]) are based on a nonlinear transformation of the data x i at reference number 902 in Fig. 9 into a high-dimensional (in particular implicit) feature space ℌ, as indicated at reference numeral 904 in Fig. 9. The calculation of the inner products in the feature space 904 required for the ML can be determined by evaluating a positive definite kernel, 〈Φ(xi),Φ(xj)〉=κ(xi,xj), where Φ is a non-linear mapping into the (infinite-dimensional) feature space, shown as an example at reference numeral 906, which is usually not explicitly known. It is only important that a similarity of data points in the feature space 904 can be determined. Neither the feature space (and / or data in the feature space) 904 nor the Fig. are (e.g. explicitly) calculated (and / or are not, especially explicitly, known).

[0256] According to the "kernel trick," only the kernel function on the right-hand side of equation (3) needs to be evaluated (cf. Mercer's Theorem). The "kernel trick" is particularly suitable for nonlinear relationships and unknown feature spaces, such as those encountered in the inventive technology for compensating an X-ray grating shadow using an adaptive and / or learning control of a compensation actuator.

[0257] The kernel κ can be defined as a metric of the high-dimensional feature space ℌ (e.g., at reference numeral 904 in Fig. 9). Alternatively or additionally, a distance of data points (e.g. in the set SA at reference numeral 902 in Fig. 9) in the feature space ℌ be large if the value of the kernel κ is large.

[0258] There are many different kernel functions κ, the most commonly used kernel is the Gaussian kernel: κ(xi,xj)=exp(−‖xi−xj‖22σ2).

[0259] Using the Gaussian kernel in equation (4), many non-linear relationships (cf. equation (2)) can be represented sufficiently accurately by a kernel expansion over training data, ƒKAF(x)=∑i=1Nαiκ(xi,xj).

[0260] KAF algorithms, such as KRLS and / or KLMS, calculate the coefficients (and / or weights) α at runtime i . The training points used x iare collected in a lexicon (technical term: dictionary) X. The cardinality of the lexicon X is usually limited in practice in order to keep storage time and / or computing time practical and / or limited.

[0261] For the present technical problem, namely the control signals (and / or setpoints) y d to be calculated in such a way (in particular by determining the similarity in a high-dimensional, usually infinite-dimensional, feature space (and / or Hilbert space) spanned by the transformation Φ) that the quality measure is optimized (e.g. c i → 0, algorithms for adaptive control computation using kernel identification can be applied.

[0262] A KAF is implemented, which defines the relationship f between the input data (also: inputs), including the (e.g. current) target motion state x d and the control signal y d, and the resulting output data, including the quality measure c i (or c x and / or c y ): ci=ƒ(x¯d,yd), inverse and suitable for prediction: yd,next=ƒKAF(ci,x¯d,yd).

[0263] Alternatively or in addition to the (e.g. current) target movement state x d the (e.g. current) actual motion state x can be used (especially in equations (6) and (7) as well as in the implementation of the KAF described below).

[0264] It can be operated without a special control signal, y d = 0, and the quality measure c i of the recorded images. In a next pass (and / or repetition and / or iteration), the control signal (and / or target value) y d be changed, so that the quality measure c i This change (e.g. of the control signal y dand / or the quality measure c i ) forms the basis for the similarity analysis of the ML in the KAF. The procedure is similar to the ILC method from Fig. 13, but the signal processing used is very different.

[0265] The ML is used to generate (especially at runtime) new, suitable control signals y d to predict. The data points x i (especially for the kernel trick) must therefore be formed in such a way that they represent a temporal progression (and / or a history) of the state x d used target values ​​y d and the associated error sizes c i Using the kernel trick according to equation (3), the similarity of the points defined in this way in the feature space ℌ can then be determined.

[0266] The result of the kernel trick can be used to calculate new target values ​​y with a KAF according to equation (7). d,next“query” them so that they can be assigned to the data points with future optimized quality measures (e.g. c i → 0) are as similar as possible in the feature space.

[0267] Since not (and / or not only) static effects contribute to the quality measure (e.g. c i ), it is important to train the KAF with a time embedding (and / or history) of the data, in which not only the current time step is used. A longer time segment (e.g., some sampling steps, and / or the time embedding) of the signals (in particular, including quality measures, motion states, and control signals) is summarized to a vectorial data point of higher dimension (e.g., in the space SA designated by reference numeral 902 of the Fig. 9), with which the KAF is trained.

[0268] For the implementation according to the invention, time-discrete sampled signals q[t], c i[t],... In order for the KAF to learn a mapping that adequately takes into account the dynamic effects occurring in the system, it is essential to construct the data pairs used in the filter (x i , x j ) not only the values ​​of the current sampling time step k, but also the time embedding over several time steps. For the inventive technique, the data point x i [k] (for example, as a data point in the set 902 of the Fig. 9) in time step k as follows: xi[k]=(ci[k+rc]ci[k+rc−1]⋮ci[k]ci[k−1]⋮ci[k−Tc]q[k]q˙[k]⋮q[k−Tq]q˙[k−Tq]⋮yd[k]⋮yd[k−Ty]).

[0269] Here r denotes c ∈ ℕ the prediction horizon (especially directed into the future) and T c ∈ ℕ, T q ∈ ℕ and T y ∈ ℕ the orders of the temporal embedding (especially directed into the past) of the quantities c i , q, q̇ or y d .

[0270] In a particularly simple variant, the prediction horizon r c = 1 corresponding to the next time step. According to other preferred variants, the prediction horizon r c > 1 corresponding to a plurality of time steps in a time series prediction.

[0271] For preprocessing the data point x i [k] both the control signal y d as well as the corresponding quality measure c i In order for the adaptive control system to advantageously suppress unknown disturbances d, it is also necessary to determine the motion state x¯=[qq˙] of the system into the data point x i [k]. This results in the form of (8).

[0272] The data point x i [k] of equation (8) is usually queried in each time step k, where in each step by r cTime steps into the future (and / or looking “forward”).

[0273] Using a KAF defined over the data according to equation (8), the problem is solved according to the invention (e.g., phased and / or stepwise).

[0274] In the first phase, no data is available and / or the lexicon is empty (X = Ø). A sufficiently meaningful lexicon (and / or dictionary) X (especially for training the KAF) is created. To generate the first data points, the (e.g. entire) X-ray device (also: imaging system) is moved to the desired positions q d (t) and a vanishing control signal for the compensation actuator (y d (t) = 0) of the smart grid. During this time, the actual motion states x(t) and the corresponding achievable quality measure c i(t) of the X-ray image (and / or imaging). Since the smart grid is not actuated, the quality measure is expected to be poor (c i » 0). From the collected signal waveforms of length T, max(T c , T q , T y ) ≤ k ≤ T - r c the data points x i [k] is formed according to equation (8) and added to the lexicon (and / or dictionary), X[k] = X[k - 1] U x i [k], for example, if the data points are sufficiently different. Appropriate procedures for assessing the significance of potentially added data points with respect to the existing lexicon (and / or dictionary) are known in the KAF literature (e.g., [6], [7],

[11] ).

[0275] The lexicon (and / or dictionary) X can be expanded over the course of its service life (e.g., of the X-ray machine). In practice, an expansion of the lexicon (and / or dictionary) X may only be possible to a limited extent due to computing time and / or storage time, and if the limit is reached, old data points in the lexicon (and / or dictionary) X may be overwritten. Details on this are known, for example, from [9] and / or

[11] .

[0276] In a second phase (and / or a next pass), the control signal is changed by changing max(T c , T q , T y ) ≤ k ≤ T - r c the predictive ability of the KAF according to equation (7) is used, with the aim of achieving the best possible quality measure (in particular c i = 0) when the next control signal (and / or the next control setpoint) y d [k + 1] is introduced, in particular yd,predict[k+1]=∑i=1|X|αiκ(xi[k],xquery[k]), where the “point to be queried” x query [k] (for example, as a data point in the set 902 of the Fig. 9) is built by xquery[k]=(00⋮ci[k]ci[k−1]⋮ci[k−Tc]q[k]q˙[k]⋮q[k−Tq]q˙[k−Tq]⋮yd[k]⋮yd[k−Ty]).

[0277] The data points x i [k] in equation (9) correspond to the entries of the lexicon (and / or dictionary) X.

[0278] Due to the zero entry in the first r c The KAF determines the value of the control signal y d (especially y d [k + 1]) which has the greatest possible similarity in the implicit feature space ℌ to a point containing the current motion state curve of the X-ray device, the current smart grid compensation control signal curve (and / or target value curve) and the future optimal quality measure c i = 0 (and / or perfect compensation of artifacts).

[0279] In equation (10) the prediction horizon r c used to instruct the KAF to act in all r c future steps to find data (especially control signals) in such a way that the quality measure (also: error measure) c i = 0. For example, if the prediction horizon were only r c = 1 step forward, control signals can be obtained so that the quality measure (also: error measure) c i (e.g. always) oscillates around the value 0. If alternatively or additionally several time steps (and / or r c >1, and / or simultaneously) on c i = 0, the signal is more likely to be forced into a stationary state.

[0280] Since the lexicon (and / or dictionary) X is initially filled with causally available data, the vector of equation (8) cannot be determined from the collected signal curves “all the way to the end T”, but the determination is made at (and / or after) r ctime steps. Alternatively or additionally, a start is only possible when at least the embedding depth of data is available, as expressed precisely by the restriction to max(T c , T q , T y ) ≤ k ≤ T - r c during data collection.

[0281] Alternatively or additionally, the KAF is trained with training sets comprising the data points (x i , c i, ) are trained. Then, using the trained KAF, estimates c est as c i,es t=KAF(x i ) and a prediction quality as MSE, Σ i (C i - c i,est ) 2 , determined. With c i,est =KAF(x i ) may mean the evaluation of equation (7) and / or (9).

[0282] Thus, control signal values ​​y that can be calculated at runtime are d[k + 1] by asking the KAF (which continues to learn, for example) at each time step k which choice of the control signal (and / or the control variable) y d the quality of the X-ray image will be optimized (and / or best) in the future. By using the new target size curves (e.g. for the motion state x d , the control signal y d and / or the quality measure c i ) other values ​​are obtained (e.g. for the state of motion x and / or the achieved quality measure c i , which can also be referred to as image quality), the KAF is further trained analogously to the first (and / or) phase. Ultimately, this amounts to an online change of the lexicon (and / or dictionaries) X and / or the weighting factors α i (e.g. in equation (5)).

[0283] The procedure of the second phase (and / or the next run) as well as the obtaining of the control signal values ​​that can be calculated at runtime can be iterated.

[0284] Alternatively or additionally, the changes in the control signal and the quality measure can be iterated, and with a correctly adjusted KAF, the quality measure (and / or the image quality) can gradually move further towards optimization (e.g. c i → 0), especially for all target motion states (and / or target gantry motion courses) q d (t) and / or x d (t).

[0285] Alternatively or additionally, a new target value can be queried from the KAF and applied, thereby generating a new data point, which may or may not be included in the KAF. The KAF can then be queried again, and the subsequent steps repeated.

[0286] Depending on the properties of the quality measure, different types of KAF can be used. For example, for a proof of concept, the original KAF (e.g., as disclosed in

[10] ) can be used, especially as a starting point. The original KAF includes a kernel variant of the well-known "Recursive Least Squares" filter.

[0287] If the quality measure (and / or the image quality, e.g. c i ) deteriorates significantly over the device lifetime (and / over the running time of the X-ray device) (e.g. due to aging effects), a KAF can be used which has a forgetting mechanism built in (in technical terms: “forgetting”) (e.g. a KAF with a forgetting mechanism is known from

[11] ).

[0288] Alternatively or additionally, memory limitations and / or real-time capability, especially the limited computation time per iteration, play a role in practice. Therefore, the number of points in the lexicon (and / or the set) X is limited to an appropriate cardinality. For this purpose, KAFs could be advantageously used, whose performance is limited, for example, in exchange for computation time. One such KAF is described in

[12] .

[0289] For the KAF to function as planned (and / or desired), it is usually necessary to adjust hyperparameters during development that influence the learning behavior of the chosen algorithm. These include, in particular, the kernel parameters κ, σ, the prediction horizon r c , and / or the orders T c , T q , T y the temporal embedding.

[0290] If expected signal properties (e.g., a periodicity) are known a priori for the kernel function κ (e.g., a predetermined and / or certain number of), the expected signal properties can be encoded into the KAF by (particularly appropriate) selection of a corresponding kernel. If no expected signal property is known (and / or no suitable KAF is known for an expected signal property), the Gaussian kernel of equation (4) is selected (e.g., generically).

[0291] The variant (and / or the scalar) σ > 0 in equation (4) is preferably approximately in the order of magnitude of the expected deviations ∥x i - x j ∥ chosen so that the kernel is not too tight, no overfitting of the training points occurs, and / or a (e.g. limiting and / or certain) generalization is possible.

[0292] The prediction horizon r ccan (for example, usually) be set to one (1). However, if the computation time is so slow that the KAF cannot be calculated in every control cycle, then a multi-step prediction (and a prediction horizon r c >1) possible, e.g. r c = 10.

[0293] The time embedding is chosen as large as necessary and / or as small as possible. A suitable value for the time embedding can be determined experimentally, for example. In experiments, small orders T c , T q , T y of the time embeddings. One or more values ​​of orders T c , T q , T y is advantageously increased only if the achieved results (e.g., quality measures) are not satisfactory. Values ​​(e.g., approximately) of T c = 5, T q = 2, T y= 5 can be used as a starting point for the inventive technique of compensating the X-ray grid shadow.

[0294] The standard CAFs allow (e.g., only) to learn and / or predict the scalar relationships of equations (5), (7), and (9). However, to tilt the smart grid (and / or high-aspect X-ray grating) in such a way that the X-ray beam drift is fully compensated, (e.g., however) m (usually at least m = 2) degrees of freedom are required. In practice, therefore, m CAFs are run in parallel according to equation (9), with each of the CAFs generating the control signal (and / or control variable) of a single degree of freedom. Generating a control signal per degree of freedom can be referred to as multivariable control.

[0295] For training and / or inference, the multidimensional control signal y dused in equations (8) and (10) respectively, together with the quality measure c i , which reports the overall image quality. If the learning algorithms do not work in this setup, the two KAFs can alternatively be filled separately in the x- and y-direction of the X-ray grid with data that describe the direct relationship in the respective direction. This can be done analogously to the procedure described above, whereby (e.g., however) in each step, instead of the overall quality measure c i the respective quality measure measuring only one in the coordinate direction (e.g. c x or c y ) is used.

[0296] The technology according to the invention differs fundamentally from conventional signal processing and control methods (in particular those cited in the prior art).

[0297] Kernel adaptive filters per se have been the subject of research for several years, see, for example, Steven Van Vaerenbergh, Ignacio Santamaria: A Comparative Study of Kernel Adaptive Filtering Algorithms, 2013 IEEE Digital Signal Processing and Signal Processing Education Meeting (DSP / SPE), pp. 181-186,

[13] . However, the application areas of KAF are not described. In the inventive technology, the KAF (particularly as a function approximator) is used to ultimately generate an adaptive control signal.

[0298] Control within the meaning of the inventive technology includes "feedforward control." Alternatively or additionally, the KAF depends on the achieved output (e.g., the quality measure). Alternatively or additionally, a control signal (feedforward control) is generated, which is tracked by a subordinate control system, in particular the compensation actuator.

[0299] The KAF can act similarly to a system inversion, whereby the system refers to the entire chain starting from the control signal of the compensation actuators of the smart grid up to the final relevant quantity, the quality measure (and / or the image quality, e.g. c i ). Adaptive system inversion has been researched in control engineering for decades, primarily using NNs, but no applications for improving image quality (e.g., X-ray) are known.

[0300] In [6], a KAF is used to describe an inverse control system in the context of MRAC (Model Reference Adaptive Control). Fig. Figure 14 schematically shows the architecture 1400 of the KAF control system from [6]. A model M at reference numeral 1406 generates, based on the reference variable ykr the target values ykm. The output of the KAF 1402 is u k , i.e. the control of the route (“Plant”) 1404.

[0301] The method according to the invention differs from the method known from [6] in that the reference value for the quality measure (e.g. c i ) the (particularly optimal) value 0 is used. Furthermore, the method according to the invention does not use a reference model (and / or M=0).

[0302] In contrast to [6], according to the invention, further control variables intervene in the path, namely the target movement states (and / or movement variables) of the pivoting arm (and / or the gantry) q d (and / or x d). Furthermore, in contrast to [6], the inventive technology therefore does not only have additive disturbances at the output. The biggest, obvious difference is that an imaging method (in particular X-ray) is used, the controlled system ("plant") is therefore technically much more comprehensive and cannot be described with a reference model M. Instead, the KAF must have a better generalization capability in order to extrapolate new control signals based on the data points of the lexicon (and / or dictionaries) X. This inventive use of the KAF is referred to as adaptive kernel identification forward extrapolation.

[0303] In a further embodiment, a class of AI (also known as Artificial Intelligence, AI) methods, in particular reinforcement learning (also known as reinforcement learning, RL), is used to determine the control signal (also known as control variable) and / or as a function approximator.

[0304] RL is basically known, e.g. from [9]. Fig. 10 schematically outlines the basic idea of ​​RL, according to which a learning controller is designed as an agent 1002, whose actions 1010 have an influence on the environment 1004 of the agent 1002, which in turn is measurable in its quality by a reward signal 1008 and changes in its state 1006 due to the actions 1010 of the agent 1002. The goal of the learning agent 1002 is to iteratively learn a behavior (technical term: policy) so that the expected rewards 1008 in the future are maximized and / or the agent 1002 behaves optimally.

[0305] The Agent 1002 of the Fig. 10 embodies the function approximator according to the invention in the RL (for example, as part and / or function of the control signal determination unit 206). The environment 1004 can embody all other components of the computing device 200 and / or the system comprising the X-ray device. The quality measure (e.g., c i , cx and / or c y ) can be embodied by a reward signal 1008 (and / or be understood as a reward signal 1008).

[0306] Fig. 6 can be regarded in the sense of the RL as an agent at reference numeral 604 and surroundings (all other reference numerals).

[0307] In the context of the RL, determining 604 the control signal in Fig. 6 be replaced by a policy of an agent 1002 according to Fig. 11, or according to Fig. 12, a parameter Θ of a control law can be understood as a possible action space of an agent 1002.

[0308] From the diverse class of RL algorithms, the methods most suited to the characteristics of the inventive application are selected. For the inventive compensation of an X-ray grating shadow, this is restricted as follows: The state space (especially the motion space) is continuous. The action space is advantageously also viewed as continuous to enable precise compensation actuators (e.g., in contrast to RL with discrete actions).

[0309] Off-policy and / or on-policy learning can be used. Since this is an iterative, roll-out-based learning method, alternative or complementary "policy search" and / or "actor-critic" algorithms may be advantageous. Alternatively or complementarily, value-based algorithms (e.g., Least Square Policy Iteration, LSPI) may be disadvantageous in terms of learning time and / or computation time for inference.

[0310] In a prototype implementation of RL, it is reasonable to start with the state of the art, for example, PPO

[14] or A3C

[15] for the application to the active compensation of the X-ray grid shadow. The applicability of the learning methods of all embodiments (especially of RL) is given by the fact that the control signal generated by the learning controller can be (e.g., relatively) easily checked for plausibility and (especially if necessary) further restricted (e.g., to a permissible adjustment range of the compensation actuators).

[0311] Since the X-ray grid (and / or raster) has a predetermined height, a tilted X-ray beam not only widens the bars but also shifts the slats (also: bars) on the X-ray image.

[0312] The displacement of the slats relative to a calibration image can be determined (and / or measured) by systematically shifting the initial calibration image in all directions in small steps (especially with a step length much smaller than a pixel size) using an interpolation method and measuring the deviation (e.g., the standard deviation of the difference between the shifted image and the current projection). The magnitude of the displacement vector that minimizes the difference can determine the quality measure (and / or the return quantity, e.g., c i , c x and / or c y ) (for example, per shift direction x or y).

[0313] An alternative method involves analysis of 2D Fourier transform images.

[0314] Fig. 15 shows an example of a recorded X-ray image 1500 of a honeycomb X-ray grid (also: raster image), in particular without compensation (and / or correction).

[0315] Fig. 16 shows a 2D Fourier-transformed (in particular FFT-transformed) image 1600 of an X-ray image taken with an X-ray grid.

[0316] The more visible local maxima (and / or patterns) of the amplitudes are in the FFT image, the more disturbing periodic structures (especially inclined lamellae of a smart grid) may be in the X-ray image. The sum of the amplitudes of the local maxima can be recorded. The sum of the amplitudes can determine the quality measure (and / or the return value, e.g., c i ) and / or be proportional to the quality measure (e.g. inversely).

[0317] The FFT image can appear similar with and / or without a patient in the X-ray beam (and / or between the X-ray source and the detector), as the FFT affects regular and / or periodic signal components. A patient in the X-ray beam typically results in irregular signal components and / or frequency components, which are (e.g., merely) smoothed and / or visible in the FFT image. The X-ray grid, on the other hand, typically contributes to a regular and / or periodic signal component.

[0318] The FFT image is based on the fact that every curve and / or every 2D image can be decomposed into a unique sum of fundamental waves with different frequency and amplitude, as in Fig. 16 is shown as an example. The amplitudes are shown. The brighter the FFT image (e.g. in Fig. 16), the greater the amplitude of the corresponding fundamental. The low-frequency fundamentals are located in the center of the FFT image. The further you go to the edge, the higher the frequency of the fundamental. The direction reflects the orientation of the fundamental.

[0319] Fig. 17A, Fig. 17B and Fig. 17C show an example of a subtraction (and / or division) of a calibration image. From the X-ray image 1700-A of a patient taken with an X-ray grid in Fig. 17A shows a calibration image 1500 of the X-ray grid in Fig. 17B is deducted (and / or subtracted, and / or divided) so that the Fig. 17C, the subtraction image 1700-C (and / or division image, which may also be referred to as a medical utility image) is created, which shows structures of the patient.

[0320] If a calibration image (and / or pure image) of the X-ray grid (and / or pattern) without a patient is available that matches the acquired X-ray image (especially with a patient) (e.g., in size, orientation, and type of X-ray grid), the acquired X-ray image (especially with a patient) and the calibration image can be subtracted (and / or divided) from each other. This eliminates the X-ray grid (and / or pattern) in the subtraction image (and / or division image).

[0321] The subtraction principle (and / or division principle) of X-ray images can be considered analogous to a digital subtraction angiography, in which only vessels are to be displayed and the anatomy is “subtracted away” (and / or “divided away”) by subtraction (and / or division) with a pure anatomical image.

[0322] The more deviations from the periodicity (e.g., due to tilting of the X-ray grating and / or grid) are visible in the FFT image, the more the peaks in the Fourier image disappear and it becomes increasingly homogeneous – there are fewer and fewer characteristic frequencies. This change can be measured using the displacement vector (and / or quality measure, e.g., c i ) using phantoms and later apply them inversely.

[0323] By analyzing the periodicity of the FFT image, the extent of artifacts can be determined and / or compensated (also: eliminated).

[0324] In another embodiment, end-to-end learning is used to learn the control signals of the smart grid's compensation actuators and / or the movement of the pivoting arm (and / or gantry movement). With end-to-end learning, all steps between input and final result are learned independently by AI.

[0325] For example, it is known to generate a control system for a robot arm that allows the robot to place books on a shelf directly from raw images provided by a camera directed at the scene, as described, for example, in

[16] . The robot thus observes itself and learns the necessary movements to obtain a desired camera image.

[0326] Transferred to X-ray imaging and the control of the compensation actuators of the smart grid, Fig. 7 the separately shown determination processes for the control signal and the quality measure at reference numerals 604 and 616, respectively, are replaced by the end-to-end learning at reference numeral 605 (in particular as a function approximator).

[0327] At reference number 618 in Fig. 7, a possible input of reference images for end-to-end learning is further shown at reference numeral 605.

[0328] To implement the end-to-end learning block 605, a Deep RL method can be used. In addition to an RL policy π for calculating the control signals, the required rewards are calculated directly from the acquired X-ray images (also: in particular, medical, useful images) 612 by training a Deep CNN classifier using RL. The acquired (also: obtained in use) X-ray images 612 are compared (and / or labeled) with reference images 618.

[0329] Alternatively or additionally, as in Fig. 8, the acquired X-ray images 612 are labeled with an evaluation by one or more experts (e.g., a radiologist), so that the policy π is iteratively improved with respect to the rewards thus calculated. As a result, the control of the compensation actuators will gradually function in such a way that X-ray images are obtained that exhibit fewer artifacts overall.

[0330] End-to-end learning, as exemplified in Fig. 8, can be so generalizable that even the motion state data (and / or system motion) is generated from the desired X-ray images.

[0331] End-to-end learning of X-ray grid shadow compensation offers significant advantages. Alternatively or in addition, end-to-end learning can also utilize an environment camera that externally monitors the relative position of the X-ray source and detector (and / or the C-arm) and (e.g., additionally) generates ground truth for the learning process (particularly with regard to the motion state data and / or the associated quality measure, especially when the control signal disappears).

[0332] Alternatively or additionally, for the end-to-end learning according to the invention, starting from the medical image, only the appearance of an artifact-affected image compared to an artifact-free one can be relevant, and the relationship with the positioning of the pivoting arm (and / or the actuators) that generates the artifacts can be learned. In particular, other influencing factors can be eliminated. Alternatively or additionally, the control signals generated in this way can be combined in a subsequent (and / or second) step with a monitoring function that monitors what is happening in the room and / or the environment using conventional methods, e.g., based on an environment camera.

[0333] The inventive technique is advantageously agnostic with regard to the specific mechanical structure and physical reasons (e.g., shadowing and / or focus drift) that result in a deterioration of X-ray image quality, since the quality measure can be determined (and / or measured) directly in the resulting X-ray image, and the compensation is determined (and / or calculated) (in particular exclusively) data-driven. Thus, the inventive technique can be easily transferred to novel imaging (in particular X-ray) systems, new system versions, and / or system variants, in contrast to conventional model-based methods, which require additional development effort each time.

[0334] In conventional methods, statically approached positions could be utilized, since only for this purpose is the difference calculation using calibration images possible, or the modeling would otherwise be too complex and / or not computable within runtime. Using the inventive technology, compensation for dynamic travel is possible (e.g., in principle).

[0335] Since the inventive technology achieves highly precise actuation and enables the compensation of highly nonlinear and / or unknown (and / or poorly measurable) effects, greater mechanical deformations than conventional methods can be tolerated. Thus, the pivoting arm (and / or gantry) does not need to be made more rigid; in fact, it can even be made less rigid.

[0336] This makes it possible to save material in the design, as many requirements, e.g. with regard to gears and / or bearings, which were previously necessary in order to prevent mechanical vibrations from occurring in the first place, can be reduced.

[0337] Also, variations in series production could be better tolerated using the inventive technique than with a conventional model-based approach based on a single fixed model.

[0338] Tilted and / or skewed detector attachments to the pivoting arm (especially the C-arm) are conventionally unacceptable. The inventive technology enables compensation, particularly by means of compensation actuators, so that manufacturing waste can be reduced.

[0339] From the perspective of achievable performance, the inventive technology has the advantage that learning mechanisms (particularly for the function approximator) can remain active even throughout the life cycle of the X-ray machine, counteracting (particularly automatically) slowly changing processes that degrade imaging. For example, poorer or different dynamics in the mechatronic drive train may occur due to wear.

[0340] The technology according to the invention can also be used if the compensation actuator is mounted on the radiator side to compensate for the mechanical deficiencies.

[0341] A further advantage of the ILC method is the acausal filtering, which is possible due to the learning iterative approach. This allows the precision control of the compensation actuators to be learned in such a way that it intervenes almost before the effects would later occur due to the dynamics. This is possible because the filters L and Q of the ILC method can even be acausal, which is a particular advantage over an online method. The reason for this is that the ILC changes the control signal of the next iteration by calculating between iterations, at which point the signal curves of the previous iteration are fully available. Forward and / or backward and / or acausal filtering is therefore possible.

[0342] A (further) advantage of the kernel-trick-based KAF is that (for example) highly nonlinear effects can be learned. Alternatively or additionally, the KAF is online-capable, in particular, it learns not only between movement executions, but also during them. A (further) advantage of RL is that it is more general than ILC (for example, ILC can be viewed as a relatively simple RL strategy for episodic settings). RL requires more computational effort, has better generalization capabilities, and / or can also be applicable when the prerequisites for ILC (for example, repetitive execution of movements and repeatable disturbance variables) are not met.

[0343] The inventive technique is particularly evident when a mass (e.g. 50kg) is attached to the detector, which increases the deflection of the pivoting arm (especially C-arm), and then the image quality nevertheless improves after a few runs.

[0344] Unless already explicitly described, individual embodiments or their individual aspects and features described with reference to the drawings may be combined or exchanged with one another without limiting or expanding the scope of the described invention, provided such a combination or exchange is reasonable and in the spirit of the present invention. Advantages described with reference to a specific embodiment of the present invention or with reference to a specific figure are, wherever applicable, also advantages of other embodiments of the present invention.

[0345] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

[0346] Cited prior art (the contents of which are incorporated herein by reference) [1] Hans P. Geering et al.: “Design of robust controllers with the H ∞ -Method“, Control Engineering, Bulletin SEV / VSE 3 / 99, pages 55-58 [2] Schröder, Buss: Intelligent Methods: Identification and Control of Nonlinear Systems, Springer 2017 [3] DA Bristow, M. Tharayil and AG Alleyne, “A survey of iterative learning control,” in IEEE Control Systems Magazine, vol. 26, no. 3, pp. 96-114, June 2006, doi: 10.1109 / MCS.2006.1636313. [4] D. Shen, “Iterative learning control with incomplete information: a survey,” in IEEE / CAA Journal of Automatica Sinica, vol. 5, no. 5, pp. 885-901, September 2018, doi: 10.1109 / JAS.2018.7511123. [5] Nard Strijbosch, Tom Oomen, „Iterative learning control for intermittently sampled data: Monotonic convergence, design, and applications,“ Automatica, Vol 139, May 2022. [6] A. Abelli, A. Ferrari, S. Monaco and C. Richard, „Adaptive inverse control using kernel identification,“ 2012 American Control Conference (ACC), 2012, pp. 202-207, doi: 10.1109 / ACC.2012.6315449. [7] Weifeng Liu, José C. Principe, Simon Haykin: Kernel Adaptive Filtering: A Comprehensive Introduction, Wiley Science, 2010. [8] JL Rojo-Alvarez, M Martinez-Ramon, J Munoz-Mari, G Camps-Valls: Digital Signal Processing with Kernel Methods, Wiley-IEEE Press, 2018. [9] Stefan R. Friedrich, Michael Schreibauer, Martin Buss, „Least-squares policy iteration algorithms for robotics: Online, continuous, and automatic,“ Engineering Applications of Artificial Intelligence, Volume 83, 2019, Pages 72-84, https: / / doi.org / 10.1016 / j.engappai.2019.04.001

[10] Y. Engel, S. Mannor, and R. Meir. „The kernel recursive least-squares algorithm“, IEEE Transactions on Signal Processing, volume 52, no. 8, pages 2275-2285, 2004.

[11] S. Van Vaerenbergh, M. Lazaro-Gredilla, and I. Santamaria, „Kernel Recursive Least-Squares Tracker for Time-Varying Regression,“ IEEE Transactions on Neural Networks and Learning Systems, vol.23, no.8, pp.1313-1326, Aug. 2012.

[12] S. Zhao, B. Chen, P. Zhu, J. C. Principe, „Fixed budget quantized kernel least-mean-square algorithm“, Signal Processing, Volume 93, Issue 9, September 2013, Pages 2759-2770.

[13] Steven Van Vaerenbergh, Ignacio Santamaria: A Comparative Study of Kernel Adaptive Filtering Algorithms, 2013 IEEE Digital Signal Processing and Signal Processing Education Meeting (DSP / SPE), pp. 181-186

[14] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.

[15] Mnih, Volodymyr, et al. “Asynchronous methods for deep reinforcement learning.” International conference on machine learning. PMLR, 2016.

[16] Avi Singh, et al: “End-to-End Robotic Reinforcement Learning without Reward Engineering”, Robotics: Science and Systems 2019, Freiburg im Breisgau, June 22-26, 2019 QUOTES CONTAINED IN THE DESCRIPTION

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

[0000] Schröder, Buss: Intelligent Methods: Identification and Control of Nonlinear Systems, Springer 2017, [2

[0235] Steven Van Vaerenbergh, Ignacio Santamaria: A Comparative Study of Kernel Adaptive Filtering Algorithms, 2013 IEEE Digital Signal Processing and Signal Processing Education Meeting (DSP / SPE), pp. 181-186 [0297, 0346] Hans P. Geering et al.: “Design of robust controllers with the H ∞ -Method“, Control Engineering, Bulletin SEV / VSE 3 / 99, pages 55-58

[0346] Schröder, Buss: Intelligent Methods: Identification and Control of Nonlinear Systems, Springer 2017

[0346] DA Bristow, M. Tharayil and AG Alleyne, “A survey of iterative learning control,” in IEEE Control Systems Magazine, vol. 26, no. 3, pp. 96-114, June 2006, doi: 10.1109 / MCS.2006.1636313

[0346] D. Shen, “Iterative learning control with incomplete information: a survey,” in IEEE / CAA Journal of Automatica Sinica, vol. 5, no. 5, pp. 885-901, September 2018, doi: 10.1109 / JAS.2018.7511123

[0346] Nard Strijbosch, Tom Oomen, „Iterative learning control for intermittently sampled data: Monotonic convergence, design, and applications,“ Automatica, Vol 139, May 2022

[0346] A. Abelli, A. Ferrari, S. Monaco and C. Richard, „Adaptive inverse control using kernel identification,“ 2012 American Control Conference (ACC), 2012, pp. 202-207, doi: 10.1109 / ACC.2012.6315449

[0346] Weifeng Liu, José C. Principe, Simon Haykin: Kernel Adaptive Filtering: A Comprehensive Introduction, Wiley Science, 2010

[0346] JL Rojo-Alvarez, M Martinez-Ramon, J Munoz-Mari, G Camps-Valls: Digital Signal Processing with Kernel Methods, Wiley-IEEE Press, 2018

[0346] Stefan R. Friedrich, Michael Schreibauer, Martin Buss, „Least-squares policy iteration algorithms for robotics: Online, continuous, and automatic,“ Engineering Applications of Artificial Intelligence, Volume 83, 2019, Pages 72-84, https: / / doi.org / 10.1016 / j.engappai.2019.04.001

[0346] Y. Engel, S. Mannor, and R. Meir. „The kernel recursive least-squares algorithm“, IEEE Transactions on Signal Processing, volume 52, no. 8, pages 2275-2285, 2004

[0346] S. Van Vaerenbergh, M. Lazaro-Gredilla, and I. Santamaria, „Kernel Recursive Least-Squares Tracker for Time-Varying Regression,“ IEEE Transactions on Neural Networks and Learning Systems, vol.23, no.8, pp.1313-1326, Aug. 2012

[0346] S. Zhao, B. Chen, P. Zhu, J. C. Principe, „Fixed budget quantized kernel least-mean-square algorithm“, Signal Processing, Volume 93, Issue 9, September 2013, Pages 2759-2770

[0346] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

[0346] Mnih, Volodymyr, et al. „Asynchronous methods for deep reinforcement learning.“ International conference on machine learning. PMLR, 2016

[0346] Avi Singh, et al: „End-to-End Robotic Reinforcement Learning without Reward Engineering“, Robotics: Science and Systems 2019, Freiburg im Breisgau, June 22-26, 2019

[0346]

Claims

[1] Computer-implemented method (100) for controlling a compensation actuator for compensating a shadow of an X-ray scattering absorption grid in an X-ray image which is recorded with a pivotably mounted X-ray device, in particular in real time, comprising the method steps: - receiving (S102) movement state data indicating a movement state of an X-ray device, wherein the X-ray device is mounted on a pivotable arm, and wherein the movement state is associated with a time of recording an X-ray image by means of the X-ray device; - determining (S104) a quality measure of the X-ray image with respect to a shadow of an X-ray scattering absorption grid arranged on the detector of the X-ray device; - determining (S106) a control signal for the compensation actuator for adjusting the position of the X-ray scattered radiation absorption grating depending on the received (S102) motion state data and depending on the determined (S104) quality measure, wherein the determining (S106) of the control signal comprises optimizing the quality measure; and - Outputting (S108) the determined (S106) control signal to the compensation actuator for compensating the shadow of the X-ray scattering absorption grating. [2] The method (100) of claim 1, wherein the pivotable arm comprises a C-arm. [3] Method (100) according to one of the preceding claims, wherein the position of the X-ray scattered radiation absorption grid arranged on the detector comprises a position in a detector plane and / or in a plane perpendicular to a focus direction of the X-ray source of the X-ray device. [4] Method (100) according to one of the preceding claims, wherein the movement state comprises a position, in particular an orientation in space, and optionally a speed of a position adjustment, in particular per degree of freedom, of the pivotable arm and / or the X-ray device. [5] Method (100) according to one of the preceding claims, wherein the received (S102) motion state data comprise desired motion state data and / or actual motion state data. [6] Method (100) according to one of the preceding claims, wherein compensating for the shadow comprises tilting and / or shifting the X-ray scatter absorption grid, and / or tilting and / or shifting the X-ray source. [7] Method (100) according to one of the preceding claims, wherein the compensation actuator system comprises two compensation actuators, and wherein a first compensation actuator is designed to tilt and / or shift the X-ray scattered radiation absorption grating along a first direction of a detector plane, and wherein a second compensation actuator is designed to tilt and / or shift the X-ray scattered radiation absorption grating along a second direction of a detector plane. [8] Method (100) according to one of the preceding claims, further comprising the step: - Outputting (S110) the recorded X-ray image comprising the compensation (S106), in particular in real time, of the shadow of the X-ray scattering absorption grid. [9] Method (100) according to one of the preceding claims, wherein the quality measure comprises a quality measure per direction, in particular along the detector plane. [10] The method (100) according to any one of the preceding claims, wherein determining (S104) the quality measure comprises a comparison with one or more data sets in a lexicon, wherein each data set in the lexicon comprises a motion state, a quality measure, and optionally a control signal for compensating the shadow of the X-ray scattering absorption grid; further optionally wherein each data set in the lexicon comprises a temporal sequence of motion states, quality measures, and optional control signals. [11] Method (100) according to one of the preceding claims, wherein determining (S104) the quality measure comprises evaluating contrasts in an X-ray image, in particular in a Fourier-transformed X-ray image. [12] Method (100) according to one of the preceding claims, wherein at least the steps of determining (S104) the quality measure of the X-ray image and determining (S106) the control signal for the compensation actuator are carried out repeatedly, in particular iteratively, preferably until the determined (S104) quality measure lies within a predetermined value range. [13] Method (100) according to one of the preceding claims, wherein the determining (S106) of a control signal for the compensation actuator - Machine learning, ML, and / or reinforcement learning, RL; - an iterative learning controller, ILC; - an adaptive kernel identification forward extrapolation; and / or - includes end-to-end learning. [14] Computing device (200) for controlling a compensation actuator for compensating a shadow of an X-ray scattering absorption grid in an X-ray image recorded with a pivotably mounted X-ray device, in particular in real time, comprising: - a receiving interface (202) configured to receive movement state data indicating a movement state of an X-ray device, wherein the X-ray device is mounted on a pivotable arm, and wherein the movement state is associated with a time of recording an X-ray image by means of the X-ray device; - a quality measure determination unit (204) which is designed to determine a quality measure of the X-ray image with respect to a shadow of an X-ray scattering absorption grid which is arranged on the detector of the X-ray device; - a control signal determination unit (206) which is designed to determine a control signal for a compensation actuator for adjusting the position of the X-ray scattered radiation absorption grating depending on the received motion state data and depending on the determined quality measure, wherein the determination of the control signal comprises optimizing the quality measure; and - a control signal output interface (208) which is designed to output the determined control signal to the compensation actuator for compensating the shadow of the X-ray scatter absorption grating. [15] Computing device (200) according to the immediately preceding claim, wherein the computing device (200) comprises one or more features of claims 2 to 13, and / or wherein the computing device (200) is configured to perform one or more of the steps of claims 2 to 13. [16] System for controlling a compensation actuator for compensating a shadow of an X-ray scattering absorption grid in an X-ray image recorded with a pivotally mounted X-ray device, in particular in real time, comprising: - an X-ray device mounted on a pivoting arm with an X-ray scatter absorption grid; - a computing device according to claim 14 or 15; and - a compensation actuator, wherein a receiving interface of the compensation actuator is designed to receive the specific control signal from the control signal output interface (208) of the computing device (200). [17] Computer program product with program elements which cause a computing device (200) to carry out the steps of the method for controlling a compensation actuator for compensating a shadow of an X-ray scattered radiation absorption grid in an X-ray image which is recorded with a pivotably mounted X-ray device, in particular in real time, according to one of the preceding method claims when the program elements are loaded into a memory of the computing device (200). [18] Computer-readable medium on which program elements are stored which can be read and executed by a computing device (200) in order to carry out steps of the method for controlling a compensation actuator for compensating a shadow of an X-ray scattered radiation absorption grid in an X-ray image which is recorded with a pivotably mounted X-ray device, in particular in real time, according to one of the preceding method claims when the program elements are executed by the computing device (200).

Images