Techniques for compensating for X-ray grating shadows with actuation mechanism

Abstract

The invention relates to a technique for controlling a compensation actuation mechanism to compensate for shadows of an X-ray scattering absorption grating in an X-ray image captured, in particular in real time, with a swingably mounted X-ray machine, comprising a method comprising the step of receiving motion state data indicative of a motion state of the X-ray machine. And the X-ray machine is arranged on the swinging arm. The movement state corresponds to a time point at which an X-ray image is captured by means of the X-ray machine. A quality criterion of the X-ray image is determined with respect to a shadow of an X-ray scattering absorption grating which is arranged on a detector of the X-ray machine. And determining a control signal for compensating the actuating mechanism according to the received motion state data and the determined quality standard so as to adjust the position of the X-ray scattering absorption grating. The determination of the control signal includes optimization of the quality criterion. The determined control signal is sent to a compensation actuation mechanism to compensate for the shadow of the X-ray scattering absorption grating.

Description

Technical Field

[0001] The invention relates to a technique for controlling a compensating actuator in order to compensate for the shadowing of an X-ray scatter absorption grating in an X-ray image which is taken in real time using a swingably mounted X-ray machine. In particular, a method, a computing device, a system and a computer program product are provided. Background Art

[0002] To generate an X-ray image, an X-ray scatter absorption grating (also called X-ray grating, scatter absorption grating and / or scatter grating) is usually introduced into the beam path between an X-ray radiator (also called X-ray source and / or radiator) and an X-ray detector (referred to as detector). As a result, the scattered radiation is filtered by the scatter grating before the image capturer in order to achieve an improved image quality. Novel methods, in particular additive methods, can produce scatter radiation gratings with a high aspect ratio (and / or a high structure height relative to the structure period), so-called super gratings, "smart gratings" or focusing gratings.

[0003] In such focusing gratings (also called “smart gratings”), the lamellae are not only designed higher than in conventional scattered radiation gratings in order to achieve better filtering, but are also aligned in the direction of the radiator center point in order to achieve a higher usable radiation fraction.

[0004] However, in conventional angiography systems, the new focusing grating (also called X-ray grating, or grating for short) cannot be used directly because the movement caused by the elastic deformation of the device or gantry will cause the focus of the X-ray beam to move significantly, resulting in shadows and obvious artifacts in the imaging.

[0005] Conventional X-ray grating active tracking methods use an elastic static model based on the C-arm to calculate the control instructions of the correction drive. However, the tracking obtained in this way is not accurate enough to ensure good image quality of X-ray shooting. Summary of the invention

[0006] Therefore, the technical problem to be solved by the present invention is to provide a reliable technical solution to improve image quality, in particular to perform X-ray photography in real time with a reduced radiation dose.

[0007] The technical problem is solved by a method for controlling a compensating actuator in order to compensate for the shadowing of an X-ray scatter absorption grating in an X-ray image, which is recorded in real time with a pivotably mounted X-ray machine, and 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. Advantageous aspects, features and embodiments are described in the dependent claims and in the following description in conjunction with the advantages.

[0008] The technical solution according to the invention is described below with reference to the protected method and with reference to the protected computing device. Features, advantages or alternative embodiments thereof may correspond to other protected subjects (e.g., systems, computer programs or computer program products), and vice versa. In other words: claims relating to computing devices may be improved by incorporating features described or claimed for the method. In this case, the functional features of the method are implemented by structural elements of the system, and vice versa.

[0009] According to a method aspect, a (particularly computer-implemented) method for controlling a compensating actuator is provided in order to compensate for shadows of an X-ray scatter absorption grating in an X-ray image which is recorded with a pivotably mounted X-ray machine, particularly in real time.

[0010] The method comprises a step of receiving (especially rated and / or actual) motion state data, which describes the motion state of the X-ray machine. The X-ray machine is mounted on a swingable arm. The motion state corresponds to the time point at which the X-ray image is taken with the aid of the X-ray machine. In addition, the method also comprises the step of determining a quality standard of the X-ray image with respect to the shadow of an X-ray scattering absorption grating, which is arranged on the detector of the X-ray machine. In addition, the method also comprises the step of determining a control signal for a compensating actuator mechanism based on the received motion state data and based on the determined quality standard to adjust the position of the X-ray scattering absorption grating. The determination of the control signal includes the optimization of the quality standard. In addition, the method also comprises the step of sending the determined control signal to the compensating actuator mechanism to compensate for the shadow of the X-ray scattering absorption grating.

[0011] The technology according to the present invention can also be called intelligent grating X-ray imaging with an intelligent control method.

[0012] The steps of the method can be performed repeatedly and / or iteratively (in particular during operation and / or in real time), for example by means of iterative learning control (ILC), in particular as a function approximator. This makes it possible to provide a control loop which improves the quality of the X-ray image by gradually adjusting the control signal according to an optimized quality criterion.

[0013] ILC control is a method for tracking the control of a system that works in an iterative or repetitive mode, such as a system for the orientation and / or positioning of an X-ray grating. At each orientation and / or positioning, the system must always repeat the same operation with high accuracy. The purpose of this operation is to accurately track a selected reference signal (e.g. a mass standard) within a limited time interval. Repetitions allow the system to continuously improve the tracking accuracy in that the system learns the inputs required to accurately track the reference signal. The learning process uses information from earlier repetitions in order to improve the control signal so that an iterative, suitable control action can ultimately be obtained.

[0014] In principle, perfect tracking can be achieved with the help of an internal model (e.g. as a standard or fundamental achievement in control technology in the 1970s). For example, it is known that when a constant controlled variable is to be controlled (and in particular the circuit itself does not work in an integrating manner), the controller must contain an I part (integral part). This can be attributed, for example, to the model of the jump signal represented by an integrator in the Laplace domain.

[0015] When applying the technique according to the invention, the result of the ILC control is not "perfect" (and / or does not reach a perfect model of the circuit). According to a first variant, it is possible to work with an approximately selected model, which is selected as an approximate fit. The method according to the invention (especially including ILC control) then continuously provides improvements. The model selected as an approximate fit can (especially only) be required for the design of the filter used for the ILC method.

[0016] Alternatively or additionally, the principles of the internal model may lead to conditions that enable perfect tracking (eg ideal and / or in accordance with model assumptions).The design of the regulation algorithm still allows for making a number of decisions as required for various applications.

[0017] The (eg accurately expressed) properties (in particular quality criteria) according to the invention can be improved implementation-by-implementation (eg in an ILC control) or (in particular even) during implementation (eg when using a KAF).

[0018] A separate implementation of the ILC control may comprise a (e.g. linear and / or sequential) sequence of method steps for controlling a compensating actuator according to the invention. Alternatively or additionally, the terms "linear" and "non-linear" have nothing to do with the nature of the presence or absence of feedback (and / or repetition of method steps). A linear system may in particular comprise feedback (and / or repetition of method steps).

[0019] This (eg ultimately achieved) quality criterion depends, for example, on the (eg determined and emitted) control signal. The determined (and / or emitted) control signal depends in particular on the optimization of the quality criterion.

[0020] The relationship (in particular between quality criteria and control signals) can be referred to as a “feedback” property and thus enables a control of the x-ray image acquisitions, in particular a control of the compensation actuator, in corresponding, in particular multiple uses of the method.

[0021] The "feedback" property may (eg, depending on the choice of function approximator) be important for successfully compensating for shading of the X-ray scatter absorption grating (X-ray grating for short), which ultimately determines the image quality.

[0022] The "feedback" property can be used between individual motion state curves (also called: motion sequences and / or trajectories) that are approached repeatedly (for example when using ILC) in order to improve the setpoint value curve. As an alternative or in addition, the "feedback" property can realize feedback and / or change the course of the motion state curve during the operation of the motion state curve (also called: internal) (for example when a kernel adaptive filter, KAF, is used as a function approximator in the method).

[0023] KAF is an adaptive kernel filter and thus a nonlinear adaptive filter. An adaptive filter is a filter that adapts its transfer function to the time-varying characteristics of the signal by minimizing an error or loss function that describes the degree to which the filter deviates from the ideal behavior. The adjustment process is based on learning a sequence of signal samples and is therefore an online algorithm. Nonlinear adaptive filters are filters that are nonlinear in their transfer function. Adaptive kernel filters implement nonlinear transfer functions with the help of kernel methods. In these methods, the signal is mapped to a high-dimensional linear feature space and the nonlinear function is approximated as the sum of a kernel whose domain is the feature space. Kernel methods can become a universal approximation method for nonlinear functions if performed in the recurring kernel Hilbert space. The advantages of kernel methods are convex loss functions, the absence of local minima, and moderate complexity in their implementation.

[0024] By means of the technique of the present invention, the quality of recorded X-ray images can be improved, in particular in real time (and / or during operation of the X-ray machine, in short: runtime).

[0025] By means of the technology according to the present invention, the relative movement between the position of the X-ray source and the detector (and / or the X-ray grating arranged on the detector) of the X-ray machine can be compensated by means of a compensating actuator (hereinafter referred to as an actuator). As a result, the shadow (also referred to as a shadow) of the primary X-ray caused by the position of the X-ray grating (also referred to as a grating) in the recorded X-ray image (also referred to as a radiographic image and / or an X-ray recording) can be compensated. As an alternative or in addition, the quality of the X-ray image, for example the sharpness and / or the detail accuracy, can be improved by means of the technology according to the present invention.

[0026] The X-ray grating can be constructed as a substantially two-dimensional element from an X-ray absorbing material. The base surface of the X-ray grating can preferably be adjusted to and / or matched to the surface of the detector. For adjusting the X-ray grating, a compensating actuator is configured, which can in particular be configured for adjustment relative to the surface of the detector. The compensating actuator can include at least two actuators for translating and / or tilting the grating along two main axes of the plane of the grating.

[0027] Effects due to manufacturing and assembly tolerances which have traditionally not been taken into account (or cannot be taken into account) in simulation models based on rigid body models and / or (e.g. elastic) multibody dynamics models can be compensated by means of the control of the compensating actuator mechanism according to the invention and thereby improve the quality of the captured X-ray images.

[0028] As described above, the compensating actuator mechanism may include one or more compensating actuators. The compensating actuator mechanism may in particular include two compensating actuators, wherein each compensating actuator may be adjusted along a direction (e.g. along the detector plane and / or along the base surface of the X-ray grating). The adjustment of the X-ray grating may include an adjustment of the detector (e.g. a movement and / or a tilting of the detector plane). As an alternative or in addition, the X-ray grating may be arranged (in particular fixedly) on the detector and / or mounted (in particular non-adjustably relative to the detector).

[0029] Compensation (eg in a certain direction) may comprise a tilt and / or movement of the X-ray grating and / or the detector (eg in that direction). Alternatively or additionally, compensation may comprise a tilt and / or movement of the X-ray source.

[0030] The swingable arm may also be referred to as a gantry (and / or gantry). As an alternative or in addition, the swingable arm may comprise a C-arm.

[0031] Compensating for shadows by changing the relative positions of the X-ray grating (and / or detector) and the X-ray source may also be referred to as active tracking, fine tuning or calibration (eg of the X-ray grating, detector and / or X-ray source).

[0032] By compensating for shadows, higher aspect ratios (and / or higher plate heights, in particular non-gradient tilt angles, related to the length and / or width of the grating projected onto the substrate and / or to the distance between adjacent plates) can be achieved compared to conventional X-ray gratings.

[0033] The relative (position) movement between the X-ray source and the detector in the detector plane may also be referred to as offset and / or migration.

[0034] By improving the quality of the X-ray images, an improved diagnosis and / or a more targeted treatment of the patient can be achieved. Alternatively or additionally, the radiation dose required to achieve a predetermined image quality can be reduced.

[0035] For example, in X-ray recordings, especially angiography, the use of X-ray gratings can achieve a higher separation sharpness of primary X-rays and scattered radiation, thereby reducing the radiation burden on patients and medical staff while improving image quality.

[0036] Real time may also refer to runtime, in particular runtime of an X-ray exposure (and / or X-ray fluoroscopy procedure).

[0037] Real-time may mean that the technique according to the invention is performed so fast that there is no need to slow down the process of real time in order to apply the technique (e.g. take an X-ray image). In addition, real-time (and / or runtime) may also include that during the taking of an X-ray image (and / or without particularly noticeable delay), especially during image reconstruction, the technique according to the invention may be applied to compensate for the shadowing of the X-ray grating.

[0038] For example, when determining control signals using actual movement state data (eg, with the aid of KAF), real-time capability is advantageous (and / or possible or even mandatory or necessary).

[0039] As an alternative or in addition, when the control signal is determined (especially only) based on the desired movement state data, the technology according to the invention is used before the X-ray image is recorded. For example, the determination of the control signal based on the desired movement state data is carried out before the X-ray image is recorded (for example in the planning phase) (for example with the aid of ILC) and / or is less time-critical.

[0040] Depending on the technical design, the real-time nature of this image evaluation for determining the quality and controlling the actuators may be almost mandatory (such as in solutions with KAF) or not (such as in solutions with ILC).

[0041] Alternatively or additionally, by means of the in particular real-time (in particular active) compensation, a high-quality X-ray image can be provided quickly, in particular while the patient is still in the exposure position. Alternatively or additionally, by means of the in particular real-time improvement of the quality of the X-ray image, a rapid assessment can be made as to whether further X-ray images are required. Further alternatively or additionally, by means of the in particular real-time compensation of the shadows, a return visit (for example for further exposures, discussion of the diagnosis and / or discussion of a treatment plan) in the future (for example after a few days or weeks) can be avoided.

[0042] Oscillation may include translational and / or rotational motion.

[0043] The swing arm (and / or the swingably mounted X-ray machine) may have translational and / or rotational degrees of freedom. The X-ray machine may in particular comprise a C-arm, for example with five independent degrees of freedom. As an alternative or in addition, the X-ray machine (and / or the swingable arm), in particular the C-arm, may also comprise more or less than five independent degrees of freedom.

[0044] The X-ray grating may also be referred to as a detector grating. The X-ray grating may include plates (for example made of metal, in particular lead) which extend outwards from a base surface.

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

[0046] The rest position (also called the neutral position) or unmoved position of the X-ray grating may include: the X-ray source and detector of the X-ray machine are arranged perpendicular to each other (or to each other), and the focus (or Fokuspunkt) of the X-rays emitted by the X-ray light source is displayed at a predetermined point (preferably the center point) of the X-ray grating.

[0047] An X-ray image may also be referred to as an X-ray recording and / or an X-ray video. As an alternative or in addition, an X-ray image may also include a time series of individual (and / or instantaneous) X-ray images.

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

[0049] The detector, on which the X-ray grating is arranged, may comprise a weight of, for example, 50 kg. Alternatively or additionally, the oscillating arm may be elastically deformed due to the movement and / or movement state detected in the movement state data.

[0050] The width and / or diameter of the X-ray grating can be 10 centimeters (10 cm) to 50 centimeters (50 cm), in particular 20 cm to 45 cm. As an alternative or in addition, the area of ​​the X-ray grating can be 100 cm 2 Up to 2500cm 2 , especially 20cm×20cm to 45cm×45cm.

[0051] The x-ray grating can include a rectangular, in particular a square, base surface. Alternatively or in addition, the x-ray grating can include a round, elliptical or circular base surface.

[0052] The height of the X-ray grating (and / or the height of the plates of the X-ray grating) can be 5 mm to 2 cm. Alternatively or additionally, the ratio of the height of the X-ray grating to the base surface is 1:10, or a value greater than 1:10.

[0053] In conventional X-ray gratings, the plates can extend perpendicularly to the base surface. Alternatively or additionally, in X-ray gratings (also called "intelligent gratings" and / or focusing gratings), the central plates can extend perpendicularly (at least approximately) relative to the normal of the base surface at an angle that increases towards the edge of the X-ray grating. Alternatively or additionally, the plates can be aligned with a common focus. The common focus can correspond to the position of a point X-ray source (especially if there is no relative movement between the X-ray source and the detector due to position adjustment).

[0054] The outer webs of the (for example focusing) x-ray grating can be offset by an angle of a maximum of 20° (and / or π / 9) relative to the normal to the base surface of the x-ray grating.

[0055] The X-ray grating can be produced by means of an additive manufacturing method. Alternatively or additionally, the X-ray grating can include an X-ray absorbing material, in particular a metal (in particular as production material), preferably lead.

[0056] The plates of the X-ray grating 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-transmissive material, such as aluminum. The stability of the X-ray grating can be improved by filling the grid and / or honeycomb structure.

[0057] The received motion state data (for short: motion data) may include sensor data and / or control data of the X-ray grating. The motion state data received as sensor data (especially including measurement data) may also be referred to as actual motion state data. As an alternative or in addition, the actual motion state data corresponds to the current and / or real motion state. Furthermore, as an alternative or in addition, the set motion state data may correspond to the planned motion state. As an alternative or in addition, the set motion state data may be received as control data for the X-ray machine.

[0058] The motion state data may represent the motion state of the X-ray machine and / or its components (C-arm). Alternatively or additionally, the motion state data may include the position and / or direction of the swing arm. Furthermore, as an alternative or additionally, the motion state data may also include the speed and / or acceleration of the position change and / or direction change of the swing arm.

[0059] When using ILC, for example, the desired motion state data can be received (e.g. directly). The desired motion state data (also referred to as: desired quantity curve) can be changed with each repetition of the method steps (and / or in each iteration). As an alternative or in addition, the actual motion state data (and / or the actual motion state) can also be indirectly incorporated into the quality standard that has been achieved (and / or is specifically used and / or determined in the method).

[0060] When using KAF, both actual movement state data (and / or actual variables) and also setpoint movement state data (and / or setpoint variables) are used, in particular for determining control signals and / or quality criteria.

[0061] The shadow of the X-ray grating can be determined according to the received motion state data by the position of the detector of the X-ray machine relative to the X-ray light source (also called: relative movement).

[0062] The position of the X-ray grating arranged on the detector determined according to the received motion state data may include the relative movement of a predetermined point (preferably a center point) of the X-ray grating relative to the focus of the X-rays emitted from the X-ray light source.

[0063] The movement state can also be referred to as a system state or device state (especially of an X-ray machine).

[0064] The relative movement may be induced by static and / or dynamic forces. For example, the swing arm of the X-ray machine may include an elastically deformable connection between the X-ray source and the detector. As an alternative or in addition, the swing arm (and / or the X-ray machine) may include an elastically deformable holding portion of the detector (particularly related to the state of motion), the holding portion including the X-ray grating.

[0065] When the X-ray machine, in particular the swing arm (in particular the C-arm) moves, deformations that are represented in the motion state data occur due to the acting forces (for example the inertial forces of the moving different weights, the Coriolis forces). In particular, the swing arm (and / or the connection) can be elastically deformed when it is offset relative to the rest position, for example due to gravity (also called: gravitational force: for short: Gravitation). As an alternative or in addition, inertia, centrifugal forces and / or Coriolis forces can lead to deformations of the swing arm (and / or the elastically deformable connection) of the X-ray source and the detector.

[0066] The gravitational force may include a position-dependent, static force acting on the swing arm and / or the X-ray machine, in particular its detector and X-ray source. Alternatively or additionally, dynamic forces (e.g., inertial forces, centrifugal forces, and / or Coriolis forces) may act on the swing arm and / or the X-ray machine, in particular its detector and X-ray source, in a motion-dependent manner (in particular, in relation to the speed and / or acceleration of the position change and / or direction change).

[0067] The quality criterion may also be referred to as an error signal and / or a quality signal. The quality criterion reflects the quality of the X-ray image, in particular the quality with regard to shadow minimization. A quality value of zero may correspond in particular to an optimized (also: ideal) quality criterion and / or optimized quality of the X-ray machine. Furthermore, as an alternative or in addition, an optimized (e.g. minimum) value of the quality criterion may correspond to a minimum shadow of the X-ray grating on the X-ray image.

[0068] Optimization of quality criteria may include minimization of shadows (particularly in the captured X-ray images).

[0069] The determination of the control signal (and / or the optimization of the quality criterion) may include accessing a database. The database may include a table and / or a word bank (also referred to as: dictionary; technically: dictionary). The table and / or word bank may or can include a large number of data records (also referred to as: - in particular vectorial - data points). For each data record, the movement state data of the X-ray machine may be assigned to the control signal of the compensation actuator and the quality criterion. For example, the determined quality criterion may correspond to a value of zero for the control signal (and / or to a "deactivation" of the compensation actuator).

[0070] The table and / or thesaurus can or can be created based on the measurement data. As an alternative or in addition, the table and / or thesaurus can or can be expanded during the continuous operation of the X-ray machine. For example, after each (e.g. final valid) X-ray image is taken, a data record (including in particular the movement state data of the X-ray machine, for example the optimized quality criteria and optionally the control signal, in particular when the optimization of the quality criteria is not easy) is added to the table and / or thesaurus.

[0071] Determining and issuing control signals based on the motion state data of the X-ray machine can also be referred to as data-driven learning control (especially of the compensation actuator).

[0072] By means of shadow compensation (in particular active compensation) of the x-ray grid, image artifacts can be compensated and / or the quality of x-ray images recorded (in particular in real time) can be improved.

[0073] The or each X-ray machine (and / or each arbitrary X-ray) may record as digital image data.

[0074] In one embodiment, the step of determining the quality criterion may be performed for a test X-ray image. For example, the test X-ray image (and / or the Fourier transformed test X-ray image) may be evaluated with respect to the quality criterion. After compensating for shadows (e.g. visible in the test X-ray image), a (e.g. final valid) X-ray image may be captured.

[0075] In a further exemplary embodiment, the quality criterion is determined based on the received movement state data, for example based on a table and / or a vocabulary.

[0076] In a further embodiment, the step of determining the quality criterion (eg for a test X-ray image) may comprise detecting the image by means of a surrounding camera.

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

[0078] The control signal emitted may include a setpoint value for a current and / or voltage on the compensating actuator. Alternatively or additionally, the control signal emitted may include (eg updated and / or improved) setpoint motion state data.

[0079] The method can be implemented based on the cloud.

[0080] As an alternative or in addition, the method can also be implemented on a local processor, a central processing unit (termed as: Central Processing Unit, CPU), a neural processor (termed as: Neuromorphic Processor Unit or Neural Processing Unit, abbreviated as: NPU) and / or a graphics processor (termed as: Graphics Processing Unit, GPU).

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

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

[0083] An X-ray machine (e.g., elastically deformable with a pivot arm, in particular as a C-arm) may include M-dimensional (in particular independent) degrees of freedom, for example, M = 5 for a C-arm. Alternatively or additionally, the motion state, for example the motion state of a C-arm, may include a deflection angle of an axis, a velocity of an axis and / or an acceleration of an axis.

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

[0085] The position of the X-ray grating arranged (eg fixedly arranged) on the detector may include a position in the detector plane and / or in a plane perpendicular to the focusing direction of the X-ray source of the X-ray machine.

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

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

[0088] The movement state can include the position, in particular the orientation, of the swivel arm and / or the x-ray machine in space and optionally the speed of their position adjustment (in particular in each degree of freedom).

[0089] According to one embodiment, the motion state may be two-dimensional (2D), for example in each degree of freedom, and / or include as coordinates the position and velocity (eg, change in position per unit time) of an X-ray machine (eg, a C-arm).

[0090] According to another embodiment, the motion state may include sampled discrete position signals (and / or positions) at equidistant time points and / or a series of one-dimensional (1D) coordinates (e.g., each degree of freedom, in particular independent degrees of freedom) corresponding to the position of a rotatable arm (e.g., a C-arm) of the X-ray machine at predetermined time points in the time series.

[0091] The set of all positions and velocities may also be referred to as the state space of the X-ray device and / or the swing arm.

[0092] Time series can be used to perform time series analysis and / or time series forecasting (termed as: time series prediction), in particular when there is no independent measurement of (e.g. generalized) velocity or (e.g. generalized) acceleration.

[0093] The position of the X-ray grating in relation to the motion state of the X-ray machine can also be referred to as the movement in the detector plane (or the movement of the detector plane).

[0094] As an alternative or in addition, the motion state of the X-ray machine can be parameterized by generalized coordinates (eg including angles) and / or generalized velocities (eg including angular velocities).

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

[0096] The setpoint movement state data may include movement state data determined in advance and / or control signals of an X-ray machine (in particular in advance and / or “offline”).

[0097] The actual motion state data may include (e.g. actual and / or real) motion state data determined from the sensor data. For example, the position and / or speed may be measured. Alternatively or in addition, the acceleration may be determined from the (e.g. measured) motor current of the actuating mechanism of the swing arm (in particular the C-arm).

[0098] The compensation of the shadow may include a tilt and / or a movement of the detector (in particular together with an X-ray grating arranged on the detector). As an alternative or in addition, the compensation of the shadow may include a tilt and / or a movement of the X-ray grating. Furthermore, as an alternative or in addition, the compensation of the shadow may include a tilt and / or a movement of the X-ray source.

[0099] The compensating actuator mechanism may include two compensating actuators. The first compensating actuator may be designed to tilt and / or move the detector and / or the X-ray grating in a first direction of the detector plane. The second compensating actuator may be designed to tilt and / or move the detector and / or the X-ray grating in a second direction of the detector plane.

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

[0101] The tilting in direction may comprise a rotation about an axis. The axis of rotation (abbreviated as: axis) may extend along a (eg x or y) direction. The detector plane may be changed by the tilting.

[0102] In one embodiment, the relative movement of the X-ray source and the detector in each direction (5 mm to 10 mm) can be compensated (especially actively) by means of a compensating actuator. In one embodiment, as an alternative or supplement, the tilting can include, for example, a rotation of no more than 10° (and / or no more than π / 18) in each direction, especially a rotation of 5° to 10° (and / or π / 36 to π / 18).

[0103] A suitable range of values ​​for compensating the action of the actuator depends on the geometry of the swing arm (especially C-arm) and / or the X-ray machine.

[0104] As an alternative or in addition, the directional movement may comprise a parallel movement in the detector plane.

[0105] Furthermore, the method may also comprise the step of emitting the recorded x-ray image, which contains a particularly real-time compensation of the shadow of the x-ray grating.

[0106] The emission of the captured image may include an image reconstruction based on artificial intelligence (AI) (e.g., by means of a neural network, NN). Alternatively or in addition, the emitted X-ray image may include a "knockout" and / or subtraction of the X-ray grating (and / or a pure image of the grid) from the captured X-ray image (particularly when the patient is located between the X-ray source and the detector).

[0107] The quality criterion may comprise a quality criterion in each direction, in particular along the detector plane. The quality criterion may comprise two (in particular independent) values, in particular for a first direction of the detector plane and for a second direction of the detector plane. Alternatively or additionally, the quality criterion in each direction of the detector plane is used to determine the control signal of the (e.g. first and / or second) compensation actuator.

[0108] The determination of the quality criterion may include a comparison with one or more data records in the vocabulary. One or each data record in the vocabulary may include a motion state, a quality criterion and, optionally, a control signal for compensating for shadows of the X-ray grating. As an alternative or in addition, one or each data record in the vocabulary may include a time series (also called: Zeitreihe) of the motion state, the quality criterion and, optionally, the control signal. The time series may include a (for example, a predetermined) number of time steps. The (predetermined) number of time steps may also be referred to as time embedding (termed: Time Embedding). As an alternative or in addition, a record of key data may also be stored in the vocabulary, in particular a record of key data for each time step, i.e. a record of the motion state and the associated quality criterion and, optionally, a control signal.

[0109] As an alternative or in addition, the vocabulary can include control signals defined as “zero”, under which the compensation actuator (in each respective direction of the detector plane) is not controlled (in particular not actively controlled).

[0110] The determination of the quality criterion may include an evaluation of the contrast in the x-ray image, in particular in the Fourier transformed x-ray image.

[0111] For example, a Fourier transformed X-ray image may include a predetermined (especially periodic) pattern consisting of amplitude maxima (and / or maximum brightness; terminology: Peaks) and amplitude minima (and / or minimum brightness) of one or more low-frequency fundamental waves of the X-rays. Deviations from the predetermined (especially periodic) pattern (also called periodicity) and / or a weakening of the contrast between amplitude minima and amplitude maxima in the Fourier transformed X-ray image (and / or, for example, an increase in the homogeneity of the Fourier transformed image) may indicate a decrease in the quality standard of the X-ray image. As an alternative or in addition, the sum of the amplitudes of the local maxima may also be related to the quality standard (for example, inversely proportional). Furthermore, as an alternative or in addition, a decrease in multiple characteristic frequencies indicates a decrease in the quality standard.

[0112] Fourier transform may include fast Fourier transform (termed as: Fast Fourier Transformation, FFT).

[0113] The quality criterion may be measured by means of displacement vectors using the phantom image (and / or the calibration image).

[0114] The welcome image and / or calibration image may be an X-ray image of a phantom or object with known image properties taken with an X-ray grating, in particular when there is no patient between the X-ray source and the detector.

[0115] In addition, the quality measure can also be determined (and / or calculated) from the image data based on an assessment to be determined (e.g. during development), in particular a formula and / or an algorithm. In the development process, the model images (and / or the model) can be used to develop (and / or design) the determination rules (and / or the calculation rules, in particular the formula and / or the algorithm). In addition, in the development phase, the ground truth and / or the image types that can be expected based on the model used may be unknown.

[0116] As an alternative or in addition, the phantom image (and / or model) should be good enough to achieve good results in clinical use. In addition, the extension of the technology according to the invention also includes (for example, at least in theory) that the method according to the invention can also be applied during the operation of the device to improve or track.

[0117] The displacement vector may include and / or parameterize the (particularly actual) deviation of the focus on the X-ray grating compared to the non-shifted calibration data. By means of the tilt adjustment, partial shadows due to the movement due to a "wrong" angle relative to the design of the smart grating plate can be at least partially compensated.

[0118] The quality criterion can be used as a basis for controlling the compensating actuator. As an alternative or in addition, the quality criterion can be learned (for example, based on the motion state data). In addition, as an alternative or in addition, the quality criterion is adapted (especially optimized) in particular in the step of determining the control signal.

[0119] At least the steps of determining the quality standard of the X-ray image and determining the control signal for compensating the actuating mechanism can be performed repeatedly and / or iteratively. Preferably, the repetition and / or iteration are performed until the determined quality standard is within a predetermined value range (and / or the predetermined quality standard of the X-ray image is reached after compensation).

[0120] As an alternative or in addition, the steps of receiving movement state data, emitting specific control signals and / or (particularly in end-to-end learning) emitting recorded X-ray images are performed repeatedly and / or iteratively.

[0121] Determining the control signal for compensating the actuation mechanism may include machine learning (ML) and / or reinforcement learning (RL).

[0122] Reinforcement or enhanced learning (termed as: Reinforcement Learning, RL) can include a software agent and one or more reward (termed as: Reward) signals. The reward signal can be equivalent to a quality standard.

[0123] A strategy, in particular a strategy in RL, may include a collection of probability criteria. The strategy may be learned iteratively (or stepwise). For example, the optimization of a quality criterion may be learned stepwise based on the motion state data and the control signal. Alternatively or additionally, the control signal may be learned iteratively based on the motion state data and, for example, a predetermined quality criterion.

[0124] RL can include one or more parameters of the control law as an action range. As an alternative or in addition, the state range (eg for ML and / or RL) can be continuous. Also as an alternative or in addition, the action space can also be continuous.

[0125] The learning of the policy may also be referred to as "policy search" and / or include an actuator-critical (terminologically: "Actor-Critic") algorithm. Alternatively or additionally, the learning of the policy may include proximal policy optimization (PPO) (eg optimization of a quality criterion).

[0126] The "actuator critical" algorithm may be an asynchronous advantage criteria (A3C) algorithm.

[0127] As an alternative or in addition, ML and / or RL may include parallel (and / or concurrent) learning (termed: Concurrent Learning), in particular to adapt to unknown influencing variables (also called disturbance variables) on the X-ray grating shadow.

[0128] The disturbance variables may include, for example, a tilted and / or skewed mounting of the detector on a swivel arm, in particular on a C-arm.

[0129] As an alternative or in addition, the disturbance variable may include wear and / or aging processes of the x-ray machine.

[0130] As an alternative or in addition, ML and / or RL may include fuzzy logic (termed as: Fuzzy Logic).

[0131] By means of ML and / or RL, an adjustment of a setpoint variable (eg a shift relative to a calibration image) can be learned with the control signal. Alternatively or additionally, one or more setpoint variables can include a quality criterion (in particular in each direction of the detector plane).

[0132] The determination of the control signal (eg, by means of ML and / or RL) may be adaptive and / or repetitive.

[0133] The determination of the control signal may include iterative learning control (termed: Iterative Learning Control, ILC).

[0134] ILC is particularly suitable when the disturbance variables are periodic and / or repeatable.

[0135] In ILC, the control signal at one (particularly discrete) time point can be described by the control signal and the error signal (also called quality signal and / or quality measurement) at the last (particularly discrete) time point, by means of a first filter (e.g., labeled Q) acting on both the control signal and the error signal and a second filter (e.g., labeled L) acting on the error signal. For example, filters O and L can produce a multiplicative response to the error signal.

[0136] ILC implements causeless filtering, in particular because the filter and / or control signal changes from one iteration step to the next.

[0137] ILC may include control of a nominal quantity (e.g., compensating a control signal of an actuator to achieve an optimized quality standard). Alternatively or additionally, one of the actuators in the compensating actuation mechanism may include an electric drive (technically referred to as a motor controller). The nominal quantity may include the current and / or voltage of the electric drive of the actuator.

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

[0139] As an alternative or in addition, it is also possible to linearize the basic concept model and / or the first principle model (terminology: First Principle Model), the device dynamics and / or the system dynamics (e.g. X-ray machine) at (particularly appropriately selected) points and / or use H ￥ design.

[0140] H ￥ The method can be applied in regulation technology in order to synthesize a robust stable regulator by mathematical optimization, thereby achieving robust performance in terms of the weighted transfer function (see, for example, Geering et al. in [1]).

[0141] A first principles model may be created based on a fundamental understanding of physics (and / or basic physical principles), such as conservation of energy and / or inertia.

[0142] The control signal may be determined based on calculation results of the image data, such as a quality factor and / or in particular a high-frequency error signal.

[0143] In any implementation of the present technology, the time-continuous signal curve can be discretized (especially sampled at equal intervals, for example, with a sampling step size of k).

[0144] The determination of the control signal may include an adaptive kernel identification feedforward extrapolation method (termed as: Adaptive Kernel Identification Feedforward Extrapolation).

[0145] The adaptive kernel identification feed-forward extrapolation method may include learning of the correlation between the control signal and the quality criterion (and / or the error signal) by means of a kernel adaptive filter (KAF), in particular by means of time embedding.

[0146] The adaptive kernel identification feedforward extrapolation method is particularly suitable if the formation of the X-ray grating shadow is mainly influenced by strong nonlinear effects. In addition, the adaptive kernel identification feedforward extrapolation method can be used online and / or applied in real time when the X-ray image is acquired. In addition, the adaptive kernel identification feedforward extrapolation method does not require repetitive and / or periodic signal patterns.

[0147] In particular, due to the commonly used "kernel trick" for KAF, where the similarity of certain points in the feature range is determined based on the inner product in the feature range by evaluating a positively defined kernel, the adaptive kernel identification feed-forward extrapolation method can achieve more efficient and / or higher performance compared to traditional X-ray grating shadow compensation techniques.

[0148] Applying the "kernel trick" to the control signal generation solution according to the present invention to suppress image artifacts is not conventionally known.

[0149] The training data for the KAF may include data records where the control signal is zero.

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

[0151] Feed-forward extrapolation with the help of MR, RL, ILC and / or adaptive kernel recognition can achieve generalizability to never-before-seen data (and / or data not used for training).

[0152] Previously unseen data may include: not all possible movements were performed (and / or could not be performed) for generating the training data record. Alternatively or additionally, a predetermined (and / or limited) number of data (and / or data records) is used, for example possibly approaching hundreds (100) of different (and / or discrete) positions. Previously unseen data may include any number of consecutive positions (particularly positions that are not equal to approximate positions), up to a quantized resolution after a decimal point.

[0153] By means of previously unseen data and (especially) good generalizability, the technique according to the invention can achieve good results in compensating X-ray grating shadows, especially when no similar data has been obtained for training.

[0154] Alternatively or additionally, previously unseen data may also be attributed to wear or other effects occurring over a (particularly long) period of time (e.g., changed kinematic state data). For example, transmission characteristics (e.g., a swing arm) may change due to changes in the coefficient of friction, which may be temperature-dependent. Additionally, the temperature in the transmission may depend on how long the X-ray machine has been in continuous use and / or whether the room in which the X-ray machine is located is heavily air-conditioned. Expectations of good algorithms (particularly MR, RL, ILC, and / or adaptive kernel identification feedforward extrapolation methods) may include that, despite the use of previously unseen (particularly kinematic state) data, the respective algorithms are still able to provide useful feedback values ​​(e.g., quality criteria and / or control signals).

[0155] The KAF may include a kernel-based learning method. Alternatively or additionally, Gaussian process regression and / or a support vector machine (SVM) may also be used to determine the control signal.

[0156] KAF is particularly suitable for solving nonlinear regression problems. As an alternative or supplement, the nonlinear function can also map an input (termed as: Input) to an output (termed as: Output) by converting the input into a high-dimensional feature range (termed as: Feature Space) H.

[0157] For example, a Gaussian kernel (termed as: Kernel Recursive Least-Squares, KRLS), kernel recursive least squares (KRLS) and / or kernel least squares (termed as: Kernel least-mean squares, KLMS) may be used.

[0158] In any embodiment, the function may include a sum of weighted kernels. The weights (and / or coefficients of individual kernels) may be calculated at runtime. Alternatively or additionally, the vocabulary may include training points (and / or weights), and data points may be added and / or weights may be modified at runtime (e.g., due to newly emerging target size distributions).

[0159] In one embodiment, the number of cards (and / or records) of the vocabulary can be limited so as to execute the method within limited memory capacity, limited computing power and / or limited computing time.

[0160] The KAF may include a function that maps (eg, current) motion state and control signals to quality criteria. Alternatively or additionally, the application of the KAF and / or the adaptive kernel identification feedforward extrapolation method may include a reverse learning function for predicting control signals (particularly due to quality criteria and motion state).

[0161] In each embodiment, the motion state data, the quality criterion and / or the control signal are sampled discretely in time, and time embedding (termed as: Time embedding) is applied in multiple time steps in order to take into account dynamic effects. The prediction range can be configured for the motion state data, the quality criterion and / or the control signal, in particular the time embedding order of the motion state data, the quality criterion and / or the control signal.

[0162] The temporal embedding order of the motion state data, the control signal and the quality criterion may be independently (and / or differently) selected.

[0163] The advantage of independently (and / or differently) choosing the temporal embedding order is that the dimensionality can be kept as small as desired. For example, for physical quantities, the order may be two (2), in particular because the underlying differential equations are of second order. For dynamic variations of quality criteria, the order may be at least five, in particular 25.

[0164] The KAF may not know whether a quantity is physical or not. Alternatively, the developer (and / or designer) may also know which quantities are physical (and / or which quantities are non-physical) and use this knowledge to determine the order (e.g., as small as possible, as large as necessary). In motion, for example, when physical quantities from position to acceleration are critical, the motion is quadratic. Derivatives may be relevant and / or may be defined (particularly by the developer and / or designer) to be of order 2.

[0165] In another embodiment, the same time embedding order may be selected for the motion state data, the control signal and / or the quality criterion.

[0166] The predictive power of a KAF can be determined by an error measure, such as the mean squared error (MSE), between a predefined output data point and the output data point estimated using the trained KAF for the same input data point. The smaller the error metric, such as MSE, the stronger the predictive power.

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

[0168] Prediction power may be assigned to one or more steps into the future. For example, prediction power may decrease as the number of prediction steps increases (e.g., starting from a fixed point in time and / or time steps).

[0169] The prediction range may be one and / or include the next possible sampling step. Alternatively or additionally, the prediction range may also be an integer greater than 1, such as ten (10), especially when no control signal is provided (and / or is not possible) at each sampling step.

[0170] The adaptive kernel identification feed-forward extrapolation method may be implemented iteratively (and / or stepwise), in particular in real time and / or during X-ray recording.

[0171] The choice of KAF algorithm can be changed during the life of the X-ray machine. For example, the kernel least squares (KRLS) method can be used first. In addition, a forgetting mechanism can be introduced as the performance of the X-ray machine degrades due to wear and / or aging.

[0172] Setting the hyperparameters of KAF can influence the learning behavior of the chosen algorithm (e.g., improving convergence and / or avoiding overfitting), for example, if the periodicity is known and / or the size of the expected deviation is known.

[0173] Adaptive kernel identification feedforward extrapolation can be used to achieve multi-scale control, for example using M parallel (and / or independent) KAFs to tilt along M32 degrees of freedom, for example, along the (x, y) directions of the detector plane.

[0174] Conventional KAF-based inverse control methods are based on reference models, but are not applicable to imaging processes (especially X-rays). Other conventional KAF-based methods do not use reference models at all. As an alternative or in addition, the technology of the present invention is different from conventional KAF-based methods in that it has novel generalization capabilities and can infer new control signals.

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

[0176] According to one embodiment, a controller of the compensating actuation mechanism may learn a reward (particularly directly) from one or more X-ray images.

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

[0178] The labeling (and / or quality criteria) of the X-ray image may be determined based on reference images and / or an evaluation, in particular a quality or grade evaluation, by an expert (eg, a radiologist).

[0179] In an improved solution of the present invention, the control signal can be determined based on the quality standard (and / or the mark) without detecting the motion state data. As an alternative or supplement, end-to-end learning can also be used to improve the versatility of the technology of the present invention.

[0180] According to the device aspect, a computing device is provided for controlling a compensating actuator to compensate for the shadow of an X-ray scattering absorption grating in an X-ray image taken by a rotatably mounted X-ray machine, especially in real time.

[0181] The computing device comprises a receiving interface, which is designed to receive motion state data indicating the motion state of the X-ray machine, which is mounted on a swingable arm, and the motion state corresponds to a point in time when an X-ray image is taken by means of the X-ray machine.

[0182] Furthermore, the computing device comprises a quality criterion determination unit which is designed to determine a quality criterion of the x-ray image with regard to a shadow of an x-ray scatter absorption grating which is arranged on a detector of the x-ray machine.

[0183] Furthermore, the computing device further comprises a control signal determination unit, which is designed to determine a control signal for compensating the actuator mechanism to adjust the position of the X-ray scatter absorption grating according to the received motion state data and according to the determined quality standard. The determination of the control signal comprises the optimization of the quality standard.

[0184] In addition, the computing device also includes a control signal output interface for outputting the determined control signal to the compensation actuator to compensate for the shadow of the X-ray scattering absorption grating.

[0185] Optionally, the computing device comprises an X-ray image output interface designed to output the captured X-ray image, which comprises compensation of X-ray scatter absorption grating shadows, in particular real-time compensation.

[0186] The computing device may be designed to perform one or more steps of the method according to the method aspect. Alternatively or additionally, the computing device may include one or more features related to the method according to the method aspect.

[0187] According to a system aspect, a system is provided for controlling a compensating actuator to compensate for the shadow of an X-ray scatter absorption grating in an X-ray image taken by a rotatably mounted X-ray machine, especially in real time.

[0188] The system includes an X-ray machine mounted on a swing arm, which includes an X-ray scattering absorption grating. In addition, the system also includes a computing device according to the device aspect. The system also includes a compensating actuator. The receiving interface of the compensating actuator is designed to receive a determined control signal from a control signal output interface of the computing device.

[0189] The system may be designed to carry out a method according to the method aspect.

[0190] According to another aspect, a computer program product is provided. The computer program product comprises program instructions which, when loaded into a memory of a computing device, enable the computing device to execute the method for controlling a compensating actuator according to a method aspect in order to compensate for shadows of an X-ray scatter absorption grating in an X-ray image, which is recorded with a pivotably mounted X-ray machine, in particular in real time.

[0191] According to another aspect, a computer-readable medium is provided in which program instructions are stored, the program instructions being readable by a computing device and being designed to, when the program instructions are executed by the computing device, perform the steps of the method for controlling a compensating actuator according to method aspects in order to compensate for the shadow of an X-ray scattering absorption grating in an X-ray image that is taken, in particular in real time, using a swingably mounted X-ray machine.

[0192] The above characteristics, features and advantages of the present invention and their implementation methods are more clearly understood and easier to understand in connection with the following description of the embodiments of the present invention, which are described in detail in conjunction with the accompanying drawings. The following description does not limit the present invention to the embodiments described. The same components or parts are marked with the same reference numerals in different views. Figure 1 Generally not in proportion.

[0193] It should be understood that preferred embodiments of the present invention may be any combination of the dependent claims or the above embodiments with the independent claims.

[0194] This and other aspects of the invention are illustrated by the embodiments described below and by means of parameters thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0195] Figure 1 A flow chart showing a method for controlling a compensating actuator according to a preferred embodiment of the present invention for compensating for shadows of an X-ray grating in an X-ray image recorded by a pivotably mounted X-ray machine, in particular in real time;

[0196] Figure 2 A general diagram showing the structure and configuration of a computing device for controlling a compensating actuator according to a preferred embodiment of the invention for compensating for shadows of an X-ray grating in an X-ray image taken by a pivotably mounted X-ray machine, in particular in real time;

[0197] Figure 3A and Figure 3B A conventional X-ray grating with blades perpendicular to the base surface and a so-called "smart grating" are shown, which comprises blades pointing in the direction of the X-ray source, wherein the method according to the invention is suitable for compensating shadows in particular of the "smart grating";

[0198] Figure 4 and Figure 5 A technique known in the prior art for reducing X-ray grating shadows by means of an auxiliary actuation mechanism is shown;

[0199] Figure 6 , Figure 7 and Figure 8 Schematically illustrating an embodiment of the technology according to the present invention, wherein the function approximator can be implemented using an iterative learning controller (ILC), an adaptive kernel identification feed-forward extrapolation method, reinforcement learning (RL), or end-to-end learning;

[0200] Fig. 9 Schematically illustrates the basic idea of ​​the "kernel trick" for determining data points;

[0201] Fig.10 , Fig.11 and Fig.12 An embodiment of an agent used when using RL as a function approximator is shown;

[0202] Fig.13 Schematically illustrates the iterative progression of control signals and quality criteria when using iterative learning control (ILC);

[0203] Fig.14 A KAF control system according to the prior art is schematically shown, which operates with a reference model without (especially unknown) disturbance variables;

[0204] Fig.15 An X-ray image showing a honeycomb-shaped X-ray grating, in particular without compensation;

[0205] Fig.16 shows a 2D FFT image of an X-ray image taken by an X-ray grating on a detector; and

[0206] Fig.17A , Fig. 17B and Fig. 17C An example is shown in which a calibration image is subtracted from a captured X-ray image in order to calculate an X-ray raster from an output X-ray image. DETAILED DESCRIPTION

[0207] Certain reference signs in the claims should not be construed as limiting the field of implementation.

[0208] Figure 1 An exemplary flow chart of a method for controlling a compensating actuator for compensating for the shadowing of an X-ray scatter absorption grating (X-ray grating for short) in an X-ray image, which is recorded in real time with a pivotably mounted X-ray machine, is schematically shown (in particular calculated and executed). The method is generally denoted by reference numeral 100.

[0209] The method 100 comprises step S102: receiving motion state data, wherein the motion state data indicates the motion state of an X-ray machine installed on a swingable arm, and the motion state corresponds to a time point when an X-ray image is taken by means of the X-ray machine.

[0210] Furthermore, the method 100 comprises a step S104 of determining a quality criterion of the x-ray image with respect to a shadow of an x-ray grating which is arranged on a detector of the x-ray machine.

[0211] In addition, the method 100 further comprises step S106: determining a control signal for compensating the actuator mechanism to adjust the position of the X-ray grating according to the received motion state data S102 and according to the determined quality standard S104. Determining the control signal S106 comprises optimizing the quality standard.

[0212] In addition, the method 100 further comprises step S108: sending the control signal determined in step S106 to the compensation actuator mechanism to compensate for the shadow of the X-ray grating.

[0213] Optionally, the method 100 comprises a step S110 of emitting the recorded X-ray image, including compensation (eg by means of a determination step S106 ) for shadows of the X-ray grating, in particular in real time.

[0214] At least part of the method steps S102 , S104 , S106 and S108 (and optionally S110 ) can be performed repeatedly, in particular iteratively, so that the control of the compensating actuator can be gradually improved.

[0215] For example, Figure 1 As schematically shown in FIG. 1 , steps S104 and S106 of determining the quality standard of the X-ray image and determining the control signal for compensating the actuator mechanism can be repeatedly performed. To determine the control signal S106 , the (eg expected) quality standard S104 can be re-determined.

[0216] In another Figure 1 In the embodiment shown in , new movement state data may be received S102 after issuing S110 a sample X-ray image, and the determination S104 and S106 of the quality criteria for the X-ray image and for the control signal for compensating the braking mechanism is repeated.

[0217] Figure 2 The schematic diagram shows an exemplary architecture of a computing device for controlling a compensating actuator for compensating for shadows of an X-ray grid in an X-ray image that is recorded in real time using a pivotably mounted X-ray machine. The computing device is generally denoted by reference numeral 200 in the drawing.

[0218] The computing device 200 comprises a receiving interface 202, which is designed to receive motion state data indicating the motion state of the X-ray machine. The computing device is mounted on a swingable arm. The motion state corresponds to the time point at which the X-ray image is taken by means of the X-ray machine.

[0219] Furthermore, the computing device 200 comprises a quality criterion determination unit 204 which is designed to determine a quality criterion of the x-ray image with regard to shadows of an x-ray grating which is arranged on a detector of the x-ray machine.

[0220] Furthermore, the computing device 200 also comprises a control signal determination unit 206, which is designed to determine a control signal for compensating the actuator mechanism to adjust the position of the X-ray grating based on the received motion state data and based on the determined quality criterion. The determination of the control signal includes the optimization of the quality criterion.

[0221] In addition, the computing device 200 further comprises a control signal output interface 208 for outputting a determined control signal to the compensation actuator to compensate for the shadow of the X-ray grating.

[0222] Optionally, the computing device 200 comprises an X-ray image output interface 210, which is designed to output the captured X-ray image, which comprises compensation (especially real-time compensation) of X-ray grating shadows.

[0223] The receiving interface 202 , the control signal output interface 208 and optionally the X-ray image output interface 210 can be arranged at the input-output interface 212 .

[0224] The quality standard determination unit 204 and the control signal determination unit may be implemented by a processor 214 (eg, including a CPU and / or a GPU).

[0225] The computing device 200 may be designed to execute the method 100. For example, the receiving interface 202, the quality standard determination unit 204, the control signal determination unit 206, the control signal output interface 208 and the optional X-ray image output interface 210 are designed to execute steps S102, S104, S106, S108 and optionally S110.

[0226] The system (not shown) may include a computing device 200 and an X-ray machine with an X-ray grating mounted on a swingable arm and including a compensation actuator. The receiving interface of the compensation actuator is designed to receive a determined control signal from a control signal output interface 209 of the computing device 200.

[0227] Figure 3AA schematic example of a conventional X-ray machine is shown with vertical (and / or mutually parallel) blades 312. The X-ray grating includes a base surface 306 with a height indicated by reference numeral 308 and a width or diameter indicated by reference numeral 310. X-rays fan out from an X-ray source 302 as schematically indicated by reference numeral 304.

[0228] for Figure 3A For a conventional X-ray machine, no matter where the X-ray grating is located, the shadow caused by the (external) plate is located in the base plane 306 parallel to the detector plane.

[0229] Figure 3B An example of a "smart grating" (also called: focusing grating) is shown. Figure 3A Identical quantities are labeled with the same reference numerals. Reference numeral 316 shows, for example, a maximum angular deviation from the normal to the outer blade 314 (eg, 20° and / or π / 9).

[0230] for Figure 3B For the "smart grating" of the embodiment, no shadows appear in an ideal situation (especially without the movement and / or migration of the detector). Due to the movement of the detector relative to the X-ray source 302, the shadows of the plates 314 of the X-ray grating may appear to varying degrees (especially depending on the movement state of the X-ray machine).

[0231] Figure 4 An exemplary conventional method 400 is shown in which an elastostatic model of dynamic effects is created and calculated at run time to correct for relative displacement between the radiator and the detector by correcting the actuator and the gimbal suspension.

[0232] At reference numeral 402, a trajectory (also referred to as: motion curve and / or motion state curve) of a C-arm (e.g., a swingable arm) is preset. At reference numeral 404, a real-time (e.g., elastostatic) model of the mechanism of the C-arm is established and / or applied. At reference numeral 406, a deformation of the C-arm is determined (e.g., calculated). At reference numeral 408, a relative displacement between the radiator and the detector caused by the deformation is determined. At reference numeral 410, the relative displacement is corrected by two correction drives (and / or compensation actuators).

[0233] Figure 5 Shown to be applicable to Figure 4Components of the conventional method 400 of . A user interface of a computer (and / or computing device) is schematically shown at reference numeral 502. A circuit board (and / or mini PC) of a computer (and / or computing device) or an external computing device is exemplarily shown at reference numeral 504, in which a real-time (e.g., elastic static) model (and / or generated code) 508 of the mechanism is executed or output to it. In a conventional (e.g., elastic static) model of the mechanism, dynamic equations are generated by a simulation model and calculated on the fly on the mini PC (and / or circuit board) 504, and compensation is implemented using control of one or more control units (e.g., programmable memory controller, SPS) 506 of a correction drive (and / or correction actuator) 512, as shown at reference numerals 402 and 406.

[0234] With the aid of a computer (and / or computing device) 502, control is mapped to a control unit (e.g., SPS) 506, as indicated by reference numeral 510. The control (e.g., SPS) 506 can perform slave regulation of one or more motors. The rated quantities for the control (e.g., SPS) 506 are typically calculated by another control unit (e.g., in the schematic computer 502).

[0235] Figure 4 and Figure 5 A simple prototype of an active activation device for tracking an X-ray grating during operation is shown. The X-ray grating mounted on the detector is tiltably mounted, for example, on two axes, so that fine adjustment can be made in order to kinematically compensate for displacements that occur, for example, due to the elasticity of the C-arm.

[0236] Traditional solutions, such as Figure 4 and Figure 5 As schematically shown in FIG. , the operating principle is based either on static measurements at the support points and storage of values ​​in tables (and / or position-dependent calibration) or on model-based calculation of control commands for the correction drives.

[0237] A second option for model-based calculation of control commands is based on an elastostatic model of the C-arm and dynamic effects (particularly manual and / or engineering), which can be calculated in real time or at runtime with the aim of calculating the relative displacement between radiator and detector in order to compensate for the displacement by actuating corresponding correction drives.

[0238] In principle, the disadvantage of conventional methods is that the compensation accuracy that can be achieved by correcting the supergrating depends on the modeling quality of the underlying model. However, it is not easy to calculate a high-precision model of the corresponding deformation in real time, so approximations, model assumptions or model simplifications have to be used. In addition, the model does not take into account or can hardly take into account the influence of the actual X-ray machine, which affects the validity of the model, for example, due to manufacturing tolerances and / or assembly tolerances, the connection of the smart grating to the detector is slightly different from what is assumed in the model. In preliminary evaluations, it has been shown that these influences (especially the accuracy of the mechanical connection of the X-ray grating to the detector) have a significant impact on the achievable quality of the X-ray image and can seriously impair the results.

[0239] In particular, compared with conventional techniques, the techniques of the present invention are (eg, as far as possible) independent of specific hardware system architecture.

[0240] As an alternative or in addition to the SPS controller of the compensation actuator, an embedded electronic board can be used, on which the swivel arm (especially the C-arm) can be controlled in real time (technical term: Real-Time) and / or the correction actuator of the smart light barrier can be controlled (especially simultaneously and / or in parallel). As an alternative or in addition, a GPU or a GPU-like component can also be used to realize a fast calculation of the NN.

[0241] With the aid of the technology according to the invention (e.g. comprising the method 100 and / or the computing device 200), active tracking of X-ray gratings (and / or gratings) (e.g. as may be required due to radiation physics) is achieved. Although it has been conventionally proven that active tracking of X-ray gratings is feasible in principle (particularly in terms of mechanics and / or control technology) and that (high aspect ratio) supergratings can be used, the problem of high-precision compensation of radiator displacements and / or detector displacements remains the key to more accurately compensate for shadows, thereby enabling the use of higher aspect ratio supergratings and thus improving the achievable image quality with a reduced radiation dose. In addition, data-driven learning control achieves further advantages.

[0242] In the technology according to the present invention (for example comprising method 100 and / or computing device 200), a data-driven method is used to calculate the control signals of the actuator mechanism used for compensation, which method is based on quality criteria of quantities that are ultimately directly related to the image quality achieved (and not, for example, the displacement of the captured radiation focus, which indirectly affects the image quality).

[0243] The following symbols and quantities are used below:

[0244] The number of degrees of freedom of movement of, for example, a swingable arm of a C-arm and / or a gantry (denoted n≥M, where M is the number of independent degrees of freedom);

[0245] (actual) rotation angles of axes and / or generalized coordinate systems;

[0246] (actual) velocity (and / or time derivative) in a generalized coordinate system;

[0247] (actual) acceleration in a generalized coordinate system;

[0248] Rated rotation angles of axes and / or generalized coordinate systems;

[0249] Rated speed of the axis and / or generalized coordinate system;

[0250] The rotation angle of the axis that realizes the tilt of the smart grating (usually m=2);

[0251] A rated rotation angle of the axis that realizes the tilt of the intelligent grating;

[0252] the (especially dynamic and / or actual) motion state of the X-ray machine (also: system); Rated quantity of motion state;

[0253] (especially unknown) interfering variables that may lead to a degradation of the image quality;

[0254] The orientation of the coordinate system in the detector plane;

[0255] The quality standard of image quality, when artifacts are perfectly compensated, is c i =0;

[0256] Quality criterion for compensation in the x-direction, where c is the value in which the artifacts are perfectly compensated in this direction. x =0;

[0257] Quality criterion for compensation in the y direction, where the artifacts are perfectly compensated in this direction c y =0;

[0258] q[k]=q(kt s ): time discrete signal, which is acquired at time t s is derived from samples of a continuous signal q(t); and

[0259] Discrete sampling time step.

[0260] An object of an embodiment of the technology according to the invention is to determine a control signal (and / or a setpoint variable) y for compensating an actuating mechanism. d , and sends it to the compensation actuator (and / or the control unit of the intelligent light barrier) during operation, thereby optimizing the quality standard (especially c i →0), although there are still (especially even unknown) disturbance variables d acting on the system.

[0261] In a further embodiment of the technology according to the present invention, the total quality standard c i and / or quality criteria in each direction of the detector plane (c x ,c y ) can be determined (and / or calculated) from medical X-ray imaging data.

[0262] Figure 6 A first exemplary embodiment of a method 100 according to the invention is shown.

[0263] “Intelligent methods” are discussed here, because learning strategies and / or adaptation strategies, such as those from the field of artificial intelligence (AI) (e.g. fuzzy logic, RL, NNs and / or concurrent learning), are used in control systems and / or control systems to achieve adaptation to unknown influencing variables (e.g. see Buss: Intelligente Verfahren: Identification und Regelung nichtlinearer Systeme, Springer 2017, [2]).

[0264] In the embodiments (e.g. Figure 6 The following settings can be made:

[0265] - The position q of the degree of freedom of the pivotable arm (and / or in particular the gantry of the C-arm) and its time derivative is known during operation. Position q and velocity Available for system control of swingable arms, also with nominal curve q d as well as The corresponding quantities are collectively called states of motion

[0267] - First, the compensation actuator of the smart light barrier should be deactivated, i.e. d = 0 (and / or no raster tracking activity). Then, depending on the operating state X-ray images of varying quality are then formed due to known or unknown interference variables d. Interference variables can be, for example, deflections of a pivotable arm (especially a C-arm) and / or dynamic forces, such as inertial forces or Coriolis forces, which cause the X-ray grating to no longer be suitable for calibration and thus show artifacts in the X-ray image.

[0268] Based on the image processing, a quantitative quality criterion can be calculated which describes the image quality of the acquired X-ray image with respect to recognizable artifacts caused by the X-ray grating. The overall quality criterion is referred to below as c i , where c is the perfect compensation for the artifacts i =0.

[0269] Alternatively or additionally, in 2D X-ray images, the quality criterion is determined by the amount c in relation to the axis. x

[0270] and c y Description. The axis-related description is critical, for example, when undesirable artifacts are present only in one dimension, for example horizontally, while the artifacts are not visible (or only slightly visible) in the vertical direction. Quality criteria c i 、c x and c y For example, it is provided in method step S104 and optimized in step S106 .

[0271] - In order to ensure deterministic performance and certifiability of the method, it may be necessary not to change the (feedback) controller (e.g. swivel arm control, also known as:

[0272] Gantrycontrol and / or intelligent light grid control, also called: Smartgrid Control). It is (particularly conversely) advantageous that the controller structure and controller parameters remain unchanged while only the control signal is changed.

[0273] exist Figure 6 In FIG. 1 , the motion state (also referred to as system position and / or system motion) is shown at reference numeral 602. and / or Reception in step S102 of method 100 is shown. Calculation of the control rated quantity is shown at reference numeral 604. This calculation can be performed in step S106 of method 100. A compensating actuator mechanism of the smart grating is schematically shown at reference numeral 606. X-ray imaging is performed at reference numeral 608, including a smart grating on the detector, and the X-ray imaging is affected by the interference variable d, as schematically shown at reference numeral 614. Image processing is performed at reference numeral 610, through which an X-ray image of a medical use diagram 612 can be issued. Determination (and / or calculation) of a quality standard is schematically shown at reference numeral 616, which can be implemented as an optimization of the quality standard, for example, in step S106, and / or as feedback for iterative optimization in this embodiment in step S104.

[0274] exist Figure 6 In this embodiment, the setpoint amount y of the actuator capable of actively moving the supergrating is also generated during operation. d , thereby optimizing the quality criterion (e.g. c_i→0).

[0275] Figure 6 Embodiments of the invention may use different function approximators in order to optimize the quality criterion and issue one or more corresponding control signals.

[0276] A first embodiment of the function approximator comprises iterative learning control (ILC). ILC (not for X-ray imaging) is known per se, see for example [3], [4].

[0277] The basic assumption of ILC is the existence of repetitive tasks and / or repetitive disturbances. This assumption is met when the X-ray machine is mounted on a swingable arm, i.e. the disturbance variable d causing the image artifact can be given Repeat, although the curve may be non-linear.

[0278] Fig.13 FIG. 1 shows a simplified diagram of an ILC principle based on [3]. The sampling time is shown at reference numeral 1302, the number of iteration steps is shown at reference numeral 1304, and the control signal y is shown at reference numeral 1306. d (e.g., the amplitude of the control signal changing over time, as schematically shown at each iteration). The quality criterion c is shown at reference numeral 1308. i (eg the magnitude of a quality criterion over time). At reference numeral 1310 a disturbance d is schematically shown and at reference numeral 1312 a reference image is shown.

[0279] like Fig.13As schematically shown in the higher iteration steps, the control signal y is realized by the calculation of the ILC in one iteration after another. d Adjustment of the quality standard c i Iterations are here referred to as repetitive tasks, that is to say complete "runs" of control signals, such as a 180 degree movement of a C-arm for orbital 3D imaging.

[0280] In computer-implemented signal processing, the temporally continuous signal curve of a physical variable is usually discretized. For example, the position q(t) and / or the velocity q(t) are acquired at the time points t0, t1, ... As motion state data, a discrete position signal {q[t]} = {q(t0), q(t1), ...} and / or a velocity signal is obtained.

[0281] Since the time interval t s The uniformly collected amount is also abbreviated in a simple manner as q[k]=q(kt s As an alternative or in addition, the control signal (and / or the control variable) is calculated discretely in time, in particular by the computing device 200 (and / or the computer), and is subsequently converted back into a physical continuous quantity (eg current, voltage) by the motor control (termed as: Motor Controller).

[0282] The simplest ILC adaptation rule is

[0283] Given.

[0284] Here, represents the control signal variable curve in the i-th iteration of the repetitive task. The initial control signal can be calculated by a typical control design. As an alternative or in addition, it is possible to set exist In this case, the complete control of the smart grating is learned through the ILC method.

[0285] The engineering problem for applying the method is to calculate the filters Q and L. For this purpose, a model-based design is used, wherein a first-principle model (term: first-principle Modelle) and / or a distance model of the X-ray machine (also: system) is advantageously used. In this case, the dynamics (also: system dynamics) of the X-ray machine (especially of the oscillating arm) can be linearized at suitable points and a robust filter design (and / or a robust design for the filters Q and L) can be used, for example including Design. Commonly used methods for this purpose are found in the literature on regulation technology.

[0286] If the available quality criterion c is not met at each acquisition time step k i [k], the use of ILC methods (see, for example, [3]) in this technical field can also be difficult, since the determination and / or calculation is performed from the X-ray image data and computationally intensive operations are usually not performed in the control system (e.g., t s This technical difficulty can be circumvented by using more advanced ILC methods that also use less frequent quality measurements (and / or error signals). For example, such advanced ILC methods are described (and / or developed) in [5].

[0287] ILC can be viewed as a special form of (especially relatively simple) RL policy that is suitable for situations where a repetitive task is performed many times and the system perturbations are also repetitive and predictable (but especially unknown). If the above assumptions do not apply, more general RL methods can be used.

[0288] In another embodiment, an adaptive kernel identification feedforward extrapolation method (termed as Adaptive Kernel Identification Feedforward Extrapolation) is used as a function approximator. Adaptive kernel identification feedforward extrapolation methods are little or not known in the academic control and regulation literature and can be derived as early as reference [6]. In particular, kernel adaptive filters (KAFs) (e.g. based thereon) are not known for solving the problem of control signal generation to eliminate image artifacts (e.g. in X-ray images).

[0289] Compared to ILC methods, the adaptive kernel-identified feedforward extrapolation method does not require repetitive and / or periodic signal sequences (e.g., quality signals, control signals, and / or motion state data). The overall performance of the adaptive kernel-identified feedforward extrapolation method may be better than that of the ILC method because the method learns efficiently through the use of the "kernel trick" and can also generalize the correlations to previously unseen data (especially through machine learning, ML).

[0290] In the present technique, a modified version of the kernel adaptive filter (KAF) [7] (an improvement of a class of ML-based signal processing methods as described in detail in [8]) is used to infer the effects of the operating process (e.g., elastic deformation of the swing arm of an X-ray machine due to static and dynamic forces, temperature dependence of gear performance and / or wear) through the "kernel trick" and introduce counteracting effects in the control signal path.

[0291] KAF is a well-known method (see [7], [8],

[10] ,

[11] ) that belongs to the kernel-based learning method (including Gaussian process regression and / or support vector machine SVP, etc.) for solving nonlinear regression problems. This means that for a data point (x i ,y i ) input / output flow, estimating the nonlinear function f i (·), input x i With the output y i The relationship between

[0292] y i =f i (x i ), (2)

[0293] And the estimate is updated with each new data pair.

[0294] Extrapolation methods based on the kernel trick (e.g., the method described in [9]) are based on Fig. 9 The data x at reference numeral 902 in i Nonlinear transformation to Fig. 9 The high-dimensional (especially implicit) feature range shown at 904 in the figure The inner product required for ML in feature range 904 can be determined by evaluating the positively defined kernel,

[0295] <Φ(x i ),Φ(x j )>=k(x i ,x j ), (3)

[0296] In this case, in the (infinite-dimensional) feature range Φ is a nonlinear graph, which is shown by way of example at reference numeral 906 and which is usually not explicitly known. It is only important that the similarity of the data points in the feature range 904 can be determined. Both the feature range (and / or the data in the feature range) 904 and the graph 906 are (e.g. specifically) calculated (and / or are not, in particular not explicitly known).

[0297] According to the "kernel trick", only the kernel function on the right side of equation (3) needs to be evaluated (see Mercer's theorem). The "kernel trick" is particularly suitable for nonlinear relationships and unknown feature ranges, because according to the technology of the present invention, the X-ray grating shadow can be compensated by adaptive and / or learning control of the compensation actuator.

[0298] Kernel k can be called the high-dimensional feature range (For example, in Fig. 9 As an alternative or in addition, when the value of kernel k is large, in the feature range The spacing between data points (e.g. Fig. 9 The set SA at reference numeral 902 in the figure can be large.

[0299] There are many different kernel functions k, the most commonly used of which is the Gaussian kernel:

[0300]

[0301] With the Gaussian kernel in equation (4), it is possible to develop a kernel using training data to represent a wide variety of nonlinear relationships (see equation (2)) with sufficient accuracy,

[0302]

[0303] A KAF algorithm, such as KLSA and / or KLMS, calculates coefficients (and / or weights) α during operation. i The training points x used for this are i are collected in a dictionary X. In practice, in order to ensure the practicality and / or limitation of storage time and / or computing time, the number of dictionaries X is usually limited.

[0304] For the current technical problem, namely to calculate the control signal (and / or target variable) y d The quality metric is optimized (e.g., c) in a manner that is specific to the task (particularly by determining similarities in a high-dimensional (usually infinite-dimensional) feature range (and / or Hilbert space) that is stretched by the transformation Φ). i →0), an adaptive control calculation algorithm with kernel identification can be used.

[0305] Execute KAF, which will include (e.g. current) nominal motion state data and control signal y d Input data (also called: Inputs) and quality standards including i (or c x and / or c y ) including the generated result data (also called: Outputs):

[0306]

[0307] Inverse, and apply to predictive learning:

[0308]

[0309] As the (current) rated motion state Alternatively or in addition to (particularly in equations (6) and (7) and also in the execution of KAK described below) use of the (e.g. current) actual motion state

[0310] In particular, when there is no control signal, y d = 0, start, and determine the quality standard c of the captured image i In the next process (and / or repetition and / or iteration), the control signal (and / or setpoint) can be changed. d , thereby changing the quality standard c i . (For example, the control signal y d and / or quality standardsc i The changes in ) form the basis for similarity judgment of ML in KAF. Fig.13 The ILC method in is similar, but the signal processing method used is very different.

[0311] ML is used to predict new, appropriate control signals (especially at runtime). Therefore, a data point x i (especially for kernel tricks) must be formed in a way that includes state The nominal variable y used in d and the resulting error variable c i By using the kernel technique according to equation (3), it is possible to determine the characteristic range The similarity of the points within this definition.

[0312] The result of the kernel trick can be used to “interrogate” a new nominal variable y using KAF according to equation (7) d,next , so that it is consistent with the future optimization quality criteria in the feature range (such as c i →0) are as similar as possible.

[0313] Since not (and / or not only) static effects can affect quality criteria (e.g. i ), the KAF must be trained using a temporal embedding (and / or history) of the data, where not only the current time step is used. Longer time periods (e.g., several sampling steps and / or temporal embeddings) of the signals (especially including quality criteria, motion states, and control signals) are aggregated into higher-dimensional vector data points (e.g., in Fig. 9 In the space SA marked with reference numeral 902), KAF is trained by the data points.

[0314] For the embodiment according to the present invention, there is a signal q[t], c which is acquired discretely in time i [t],... In order for KAF to learn a mapping that fully takes into account the dynamic effects of the system, it is necessary not only to use the variables of the current sampling time step to construct the data pairs (x i ,xj ), it is also necessary to use time embedding with multiple time steps k. For the technique according to the present invention, the data point x i [k](e.g. as Fig. 9 The data points of the set 902 in time step k are established as follows:

[0315]

[0316] Here we will is called the forecast horizon (especially for the future), and as well as is the variable c i ,q, or d The order of temporal insertion (especially past-oriented).

[0317] In a particularly simple approach, the prediction horizon r c = 1 can be selected according to the next time step. According to another preferred solution, the prediction range r c >1 is chosen based on the number of time steps in the time series forecast.

[0318] In order to pre-process the data points, the control signal y d and the associated quality standards i In order for the adaptive control system to be able to effectively suppress the unknown disturbance variable d, it is also necessary to i [k] contains the motion state of the system This results in equation (8).

[0319] The data point x of equation (8) i [k] is usually asked at each time step k, where at each step, c time steps into the future and / or “look ahead”.

[0320] According to equation (8) with the KAF defined by the data, the problem can be solved according to the present invention (eg in stages and / or stepwise).

[0321] In the first stage, there is no data and / or the vocabulary is empty Create a vocabulary (and / or dictionary) X that is sufficiently convincing from the outset (especially for training KAF). To form the first data point, the (eg overall) X-ray machine (also called: X-ray imaging system) is placed at the desired position q d (t) Using this method, and using the missing control signal for the compensation actuation mechanism of the smart grating (y d (t)=0). At the same time, measure the actual motion state and the quality standards c assigned to the X-ray recordings (and / or images) and achieved thereby i (t). As the intelligent light barrier was not activated, the quality standard was expected to be poor (c i >>0). From the aggregated signal curve of length T, the maximum (T c ,T q ,T y )≤k≤Tr c Data point x i [k], and add it to the vocabulary (and / or Dictionary), X[k] = X[k-1] ∪ x i [k], for example, as long as the data points are sufficiently mobile. In the KAF literature (e.g., [6], [7],

[11] ), there are already appropriate methods for evaluating the information value of potential new data points relative to the existing vocabulary (and / or dictionary).

[0322] The vocabulary (and / or dictionary) may be expanded during use (e.g., during use of an X-ray machine). In practice, due to limitations in computing time and / or storage time, the expansion of the vocabulary (and / or dictionary) X may only be performed to a limited extent. If the limit is reached, old data points X in the vocabulary (and / or dictionary) may be overwritten. Details of this are known, for example, from [9] and / or

[11] .

[0323] In the second phase (and / or the next process), the control signal is changed, and in particular, when the next control signal (and / or the next control variable) y d [k+1] is introduced by using the maximum value (T of KAF) used in each acquisition step according to equation (7). c ,T q ,T y )≤k≤Tr c Predictability, thus achieving the best quality standards (especially c o =0), especially

[0324]

[0325] Among them, the "to be queried" point x query [k] (e.g. as Fig. 9 The data points in the set 902 are thus formed

[0326]

[0327] The data point x in equation (9) i [k] is equivalent to entry X in the vocabulary (and / or dictionary).

[0328] Since the first r of the vector c The zero entry in the entry will be predicted by KAF to control the signal y in the implicit feature range H. d The value of (especially y d [k+1]), which is related to the motion state curve of the current X-ray machine, the current intelligent grating compensation control signal curve (and / or the rated variable curve) and the future optimized quality standard c i = 0 (and / or perfect artifact compensation) has the greatest similarity.

[0329] Using the forecast history r in equation (10) c , in order to request KAF in all r c In the future steps, we look for data (especially control signals) so that the quality standard (also called error standard) c i = 0. If, for example, the forecast history only looks forward r c = 1 step, a control signal can be obtained so that the quality standard (also known as error standard) c i (e.g. always) oscillates around the value 0. When, as an alternative or in addition, multiple time steps (and / or r c >1, and / or simultaneously) adjusted to c i =0, the signal enters the static state earlier.

[0330] Since the vocabulary (and / or dictionary) X is initially filled with causally available data, the vector of equation (8) can be determined from the collected signal curve, i.e. not "all the way to T", but in (and / or after) time step r c As an alternative or in addition, it is possible to start only when data of at least the embedding depth is available, which can be achieved by limiting the maximum (T c ,T q ,T y )≤k≤Tr c It is expressed in technically precise terms.

[0331] Alternatively or additionally, KAF utilizes data points (x i ,c i ,). Then, the estimated value c is est Determined as c i,est =KAF(x i ), and the prediction quality is determined as MSE, ∑ i (c i -c i,est ) 2 Here, using c i,est =KAF(xi ) can represent the evaluation of equations (7) and / or (9).

[0332] This results in the control signal value y which can be calculated during operation d [k+1], in that the (e.g. progressively learned) KAF is “queried” at any time step k by evaluating the control signal (and / or control variable) y d The selection of the X-ray image quality standard will be optimized (and / or optimal) in the future. By using new setpoint variable curves (e.g. for motion states Control signal y d and / or quality standardsc i Forming other values ​​(e.g. for motion status and / or the quality standards achievedc i , also known as image quality), further train the KAF similarly to the first (and / or) stage. Ultimately, this is equivalent to modifying the vocabulary (and / or Dictionary) X and / or the weight coefficient α online i (such as the weight coefficient in equation (5)).

[0333] The program of the second phase (and / or the next run) and the acquisition of the control signal values ​​that can be calculated during the run can be repeated.

[0334] In addition, the changes of the control signal and the quality criterion can be repeated, and the quality criterion (and / or the image quality) can continue to move in the optimization direction (e.g., c i →0) continuous development, especially for all nominal motion states (and / or target gantry motion curves) q d (t) and / or

[0335] As an alternative or in addition, a new setpoint value can also be queried from the KAF and used to generate a new data point, which may or may not be contained in the KAF. The KAF can then be queried again and the subsequent steps repeated.

[0336] Depending on the characteristics of the quality criteria, different types of KAFs can be used. For example, for a proof of concept, the original KAF (as disclosed in

[10] ) can be used, in particular as a starting point. The original KAF includes a kernel scheme of a known "recursive least squares" filter.

[0337] If the quality criteria (and / or image quality, e.g. i) degrades significantly over the lifetime of the device (and / or the operating time of the X-ray machine) (e.g. due to aging effects), a KAF with a built-in forgetting mechanism can be used (e.g., a KAF with a forgetting mechanism known from

[11] ).

[0338] Alternatively or additionally, memory limitations and / or real-time capabilities, in particular limited computation time per iteration, also play a role in practice. Therefore, the number of points in the vocabulary (and / or set) X is limited to a moderate size. Performance-constrained KAFs (e.g. in exchange for computation time) can be more advantageous here. Such KAFs are described, for example, in

[12] .

[0339] In order for KAF to work as planned (and / or expected), it is usually necessary to set hyperparameters during development that affect the learning behavior of the chosen algorithm. These parameters include in particular the kernel κ, σ parameters, the prediction horizon r c and / or the order T of temporal embedding c ,T q ,T y .

[0340] If the expected signal property (e.g., periodicity) of the kernel function κ is known a priori (e.g., predetermined and / or a certain number), the expected signal property can be encoded into the KAF by (particularly appropriately) selecting the corresponding kernel. If the expected signal property is not known (and / or a suitable KAF for the expected signal property is not known), the Gaussian kernel of equation (4) is selected (e.g., in general).

[0341] The variables (and / or scalars) in equation (4) σ>0 are preferably calculated as expected deviation ∥x i -x j The order of magnitude of || is approximately chosen so that the kernel is not too narrow, does not overfit the training points, and / or allows for (e.g. restrictive and / or deterministic) generalization.

[0342] Prediction range r c can (eg, usually) be set to one (1). However, if the computation time is too slow to compute KAF in every control cycle, a multi-step prediction (prediction horizon r c 1), for example, c =10.

[0343] The time embedding can be selected as large as possible or as small as possible according to the needs. For example, the appropriate time embedding value can be determined through experiments. The experiment can be carried out from the smallest time embedding order T c ,T q ,T y Only if the results obtained (e.g. quality measurement) are not satisfactory, increase one or more orders T. c,T q ,T y Value. (e.g. approximately) T c =5, T q =2, T y A value of =5 may serve as a starting point for the technique according to the invention for compensating the X-ray grating.

[0344] A standard KAF can (e.g. only) learn and / or predict the scalar relationships of equations (5), (7) and (9). In order to tilt the smart grating (and / or high longitudinal X-ray grating) to fully compensate for the X-ray shift, (e.g.) m (usually at least m=2) degrees of freedom are required. Therefore, in practice, mKAFs are run in parallel according to equation (9), and each KAF generates a control signal (and / or control variable) for one degree of freedom. Generating one control signal for each degree of freedom can be called multivariable control.

[0345] In training and / or inference, the multidimensional control signal y in equation (8) or (10) can be used d (e.g., but still usable), and a quality criterion c that feeds back the overall image quality i If the learning algorithm does not work in this setting, it is also possible to fill in two KAFs in the x- or y-direction of the X-ray grating, respectively, and use the data describing the direct relationship in the respective direction. To this end, the procedure can be similar to the one described above, wherein (for example, however) in each step the respective mass measurement (e.g. c ) is used, which is measured in only one coordinate direction. x or c y ), rather than the total quality standard c i .

[0346] The technology of the present invention is essentially different from traditional signal processing or control technology (especially the technology cited in the prior art).

[0347] Kernel adaptive filters have themselves been a research topic for many years, see Steven VanVaerenbergh, 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 field of KAF is not described. In the technology of the present invention, KAF is used (particularly as a function approximator) to ultimately generate an adaptive control signal.

[0348] Control in the technical sense of the invention includes "feedforward control". In addition, the KAF also depends on the output achieved (e.g. quality standard). In addition, a feedforward control signal (feedforward control in English) can be optionally or additionally generated, which is tracked by the subordinate control system, in particular the compensating actuator.

[0349] KAF works like a system inversion, where the system refers to the control signal from the compensation actuation structure of the smart grating to the final variable of interest - the quality standard (and / or image quality, e.g. c i )--the whole chain. Adaptive system inversion has been studied in control engineering for decades, mainly using NNs, but has not yet been applied to improving the quality of (e.g., X-ray) images.

[0350] In [6], KAF is used to describe backstepping control systems in the context of MRAC (model reference adaptive control). Fig.14 The architecture 1400 of the KAF control system of [6] is shown. Here, the model M is derived from the reference variable at reference numeral 1406. Generate Rated Variables The output of KAF 1402 is u k , i.e. control of the production line ("factory") 1404.

[0351] The method according to the invention differs from the method known from document [6] in that it uses (in particular the optimal value) 0 as a reference variable for the quality criterion (e.g. c i ).

[0352] Furthermore, the method according to the invention does not use a reference model (and / or M=0).

[0353] Unlike [6], according to the present invention, there are more control variables on the track, namely the swing arm (and / or gantry) q d (and / or ) of the desired motion state (and / or motion variable). In addition, unlike document [6], the technology of the present invention has not only additional interference variables at the output. The biggest and most obvious difference is that according to the present invention, an imaging process (especially X-rays) is used, so the control system "(device") is much more technically comprehensive and cannot be described by the reference model M. On the contrary, according to the present invention, the KAF must have better generalization capabilities in order to infer new control signals from the data points of the vocabulary (and / or dictionary) X. According to the present invention, this use of the KAF is called adaptive kernel recognition feedforward extrapolation.

[0354] In another embodiment, a class of artificial intelligence methods (termed as: Artificial Intelligence, AI), in particular reinforcement learning (termed as: Reinforcement Learning, RL), is used to determine the control signal (also known as: control variable) and / or as a function approximator.

[0355] RL is generally known, for example from [9]. Fig.10 The basic concept of RL is schematically shown, according to which a learning controller is designed as an agent 1002, whose actions 1010 will have an impact on the environment 1004 of the agent 1002, and the quality of the environment can be measured by a reward signal 1008, and the state 1006 of the agent 1002 will change due to the action 1010 of the agent 1002. The goal of the learning agent 1002 is to iteratively learn behaviors (technically called: strategies) to maximize the expected future rewards 1008 and / or make the behavior of the agent 1002 reach the best state.

[0356] In RL, Fig.10 Agent 1002 in embodies the function approximator of the present invention (e.g., as part of and / or a function of control signal determination unit 206). Environment 1004 may embody computing device 200 and / or all other components of a system including an X-ray machine. Quality standards (e.g., c i ,c x and / or c y ) can be reflected by the reward signal 1008 (and / or understood as reward 1008).

[0357] In the sense of RL, Figure 6 It can be seen as the agent at reference numeral 604 and its surroundings (all other reference numerals).

[0358] In the context of RL, Figure 6 The control signal determination 604 may be based on Fig.11 replaced by the strategy of agent 1002, or according to Fig.12 , the parameter Θ of the control law can be understood as the possible action space of agent 1002.

[0359] From a variety of RL algorithms, the most suitable method is selected in detail according to the application characteristics of the present invention. According to the present invention, for the compensation of the shadow of the X-ray grating, the constraints are as follows. The state space (especially the motion space) is continuous. The action space is preferably also continuous in order to be able to perform the compensation action accurately (for example, in contrast to RL with discrete actions).

[0360] Offline policy and / or online policy learning can be used. Since this is based on an iterative, rolling learning process, "policy search" and / or "actuator critical" algorithms can be used as an alternative or supplement. In addition, value-based algorithms (such as least squares policy iteration, LSPI) may be disadvantageous to framework conditions in terms of learning time and / or inference computation time.

[0361] In a prototype implementation of RL, starting from the state-of-the-art techniques (e.g. PPO

[14] or A3C

[15] ), it makes sense to apply active compensation of the shadows of the X-ray grating.

[0362] The applicability of the learning method for all embodiments (especially RL) is that the control signals generated by the learning controller can be easily checked (e.g. relatively checked) to determine their plausibility and (especially if necessary) further restricted (e.g. within the permissible adjustment range of the compensating actuator).

[0363] Since the X-ray grating (and / or gratings) has a predetermined height, a tilted X-ray beam not only causes the stripes to become wider, but also causes the stripes (also called fringes) on the X-ray image to shift.

[0364] The displacement of the plate relative to the calibration image can be determined (and / or measured) by systematically moving the initial calibration image in small steps in various directions (particularly with step sizes much smaller than the pixel size) using interpolation and measuring the deviation (e.g. the standard deviation of the difference between the moved image and the current projection). The number of displacement vectors that minimize the difference can include quality measures (and / or return quantities, such as c i ,c x and / or c y ) (e.g. each displacement direction x or y).

[0365] Another approach involves analyzing the two-dimensional Fourier transformed image.

[0366] Fig.15 An example of an X-ray image 1500 taken with a honeycomb X-ray grating (also referred to as a raster image) is shown, particularly without compensation (and / or correction).

[0367] Fig.16 A two-dimensional Fourier transform (particularly FFT transform) image 1600 of an X-ray image captured using an X-ray grating is shown.

[0368] The more pronounced the local maxima (and / or patterns) of the amplitudes are in the FFT image, the more interfering periodic structures (especially tilted bands of smart gratings) may be present in the X-ray image. The sum of the amplitudes of the local maxima can be detected. The sum of the amplitudes can determine the quality criterion (and / or return value, e.g. c i) and / or is proportional to (e.g., inversely proportional to) the quality factor.

[0369] Since the FFT operates on regular and / or periodic signal components, the FFT images may look similar without and / or with a patient in the X-ray beam (and / or between the X-ray source and the detector). A patient in the X-ray beam typically results in irregular signal components and / or frequency components that are (e.g. only) smoothed and / or visible in the FFT image. On the other hand, an X-ray grating typically contributes to regular and / or periodic signal components.

[0370] like Fig.16 The basis of the FFT image is that each curve and / or each two-dimensional image can be decomposed into a unique sum of fundamental waves with different frequencies and amplitudes. The amplitude is shown in the figure. The brighter the FFT image (e.g. Fig.16 ), the larger the amplitude of the related fundamental wave.

[0371] The low-frequency fundamental wave is located in the center of the FFT image. The closer to the edge, the higher the frequency of the fundamental wave. The direction reflects the direction of the fundamental wave.

[0372] Fig.17A , 17B 17C show examples of subtracting (and / or dividing) calibration images. Fig. 17B The calibration image of the X-ray grating 1500 from Fig.17A The X-ray image 1700-A of the patient taken using the X-ray grating is removed (and / or subtracted, and / or segmented), thereby generating Fig. 17C The subtracted image 1700 -C (and / or segmented image, also referred to as a medically usable image) shown in FIG. 1 shows the structure of the patient.

[0373] If a calibration image (and / or pure image) of the X-ray grating (and / or grating) without a patient is matched (for example in terms of the size, orientation and type of the X-ray grating) to the recorded X-ray image (especially the image with the patient), the recorded X-ray image (especially the image with the patient) and the calibration image can be subtracted (and / or segmented). As a result, the X-ray grating (and / or grating) in the subtracted image (and / or segmented image) is removed.

[0374] The subtraction principle (and / or segmentation principle) of X-ray images can be seen as similar to digital subtraction angiography, where only blood vessels should be displayed and the anatomical image is "subtracted" (and / or "cut out") by subtracting (and / or segmenting) the pure anatomical image.

[0375] The greater the deviation from periodicity in the FFT image (e.g. due to the X-ray grating and / or the tilt of the grating), the more the peaks in the Fourier image will disappear and become more and more uniform - with fewer and fewer characteristic frequencies. This change can be detected in the measurement of the displacement vector (and / or quality criteria, such as c i ) with the help of the model and then apply it in reverse.

[0376] By analyzing the periodicity of the FFT image, the characteristics of the artifacts can be determined and / or compensated (and possibly calculated).

[0377] In another embodiment, end-to-end learning (termed as: End-to-End-Learning) is used to learn the control signals of the intelligent grating compensation actuator and / or the movement of the swingable arm (and / or the gantry movement). In end-to-end learning, all steps between the input and the final result are independently learned by artificial intelligence.

[0378] For example, it is known to generate robot arm controls that enable the robot to place a book on a shelf directly from the raw image provided by the camera, as described in

[16] . In this way, the robot can observe itself and learn the actions required to obtain the desired camera image.

[0379] When applied to X-ray imaging and compensation actuator control of intelligent gratings, Figure 7 The control signal and quality criterion determination processes shown at reference numerals 604 and 616 respectively may be replaced by end-to-end learning at reference numeral 605 (in particular as a function approximator).

[0380] In addition, Figure 7 At reference numeral 618 of FIG. , a possible input of a reference image for end-to-end learning at reference numeral 605 is also shown.

[0381] In order to realize the end-to-end learning module 605, a deep RL method can be used. In this case, in addition to the RL policy π for calculating the control signal, the required reward can also be calculated directly from the acquired X-ray image (also: in particular medically usable image) 612 by training a deep CNN classifier with the help of RL. The captured (also: obtained in use) X-ray image 612 is compared (and / or marked) with a reference image 618.

[0382] As an alternative or supplement, Figure 8 As schematically shown at 620 in the figure, the captured X-ray image 612 can be annotated by the evaluation of one or more experts (e.g., radiologists), so that the strategy π is iteratively improved according to the reward calculated in this way. Therefore, the control of the compensating actuator will gradually play a role, thereby obtaining an X-ray image with fewer artifacts overall.

[0383] like Figure 8 As shown, end-to-end learning can be general and can even generate motion state data (and / or system motion) based on the required X-ray images.

[0384] End-to-end learning of compensation of X-ray grating shadows has significant advantages. As an alternative or in addition, end-to-end learning can also use an environmental camera to monitor the relative position of the X-ray source and the detector (and / or C-arm) from the outside and (e.g. additionally) generate reference truth values ​​for the learning procedure (in particular with regard to motion state data and / or relevant quality criteria, in particular when the control signal disappears).

[0385] As an alternative or in addition, according to the invention, end-to-end learning based on medical images may only be concerned with how images with artifacts compare to images without artifacts and with the connection between learning and the positioning of the swingable arm (and / or the actuator) that produces the artifacts. In particular, other influencing factors can be omitted. As an alternative or in addition, the control signal generated in this way can also be combined in a subsequent (and / or second) step with a monitoring function that uses conventional methods, for example based on an environmental camera, to monitor what is happening in the room and / or the environment.

[0386] The technique according to the invention is independent of the specific mechanical design and physical causes of impaired X-ray image quality (such as shadows and / or focus shifts), since the quality criterion can be determined (and / or measured) directly in the generated X-ray image and the compensation can be determined (and / or calculated) (in particular exclusively) in a data-driven manner. The technique according to the invention can therefore be easily transferred to new imaging (in particular X-ray) systems, new system versions and / or system concepts, in contrast to conventional model-based methods, which each time involve development complexity.

[0387] In conventional methods, a static approximation of the position is available, since only then can a difference calculation be performed via the calibration images, which would otherwise be too complex to model and / or impossible to calculate in a practicable manner. With the technique of the invention, dynamic travel can (eg in principle) be compensated.

[0388] Since the technology according to the invention allows high-precision driving and compensation of highly nonlinear and / or unknown (and / or difficult to measure) influences, larger mechanical deformations can be tolerated than in conventional cases. This means that the swingable arm (and / or gantry) does not have to be more structurally stable, but on the contrary can even be made less rigid.

[0389] This allows material savings in the design, since many requirements, such as those on gears and / or bearings, which were previously necessary to prevent mechanical vibrations, can be reduced to a minimum.

[0390] Furthermore, using the technique of the present invention can better tolerate deviations in batch production compared to conventional methods based on a single fixed model.

[0391] The connection of the detector to the swingable arm (especially the C-arm) is usually not tolerated if it is tilted and / or deflected. The technology according to the invention can compensate for this, especially by means of a compensating actuation mechanism, thereby reducing production rejects.

[0392] From the perspective of achievable performance, an advantage of the technique according to the invention is that the learning mechanism (particularly the function approximator) can continue to work over the life of the X-ray machine to (particularly automatically) counteract slowly changing processes that degrade the image quality, such as poor or varying dynamics of the electromechanical drive system that may occur due to wear and tear.

[0393] The technique according to the invention can also be used if a compensating actuator is mounted on the radiator side in order to compensate for mechanical defects.

[0394] One (and more) advantage of the ILC method is that non-causal filtering can be achieved by the learning iterative procedure. It is thus possible in this way to learn the precise control of the compensating actuators so that intervention can be made later before an influence due to dynamic factors occurs. This is possible because the filters L and Q of the ILC method can even be non-causal, which is an advantage in particular compared to online methods. The reason for this is that the ILC changes the control signal for the next iteration by calculations between iterations, while the signal characteristics of the previous iteration are fully available. Therefore, forward and / or reverse and / or discontinuous filtering is possible.

[0395] A (e.g. further) advantage of the KAF based on the kernel trick is that (e.g. also) strong nonlinear effects can be learned. In addition, the KAF can also be learned online, in particular, it can learn not only during the action execution process, but also during the action execution process.

[0396] Another advantage of RL is that it is more general than ILC (e.g., ILC can be viewed as a relatively simple RL policy that is applicable in episodic settings). RL has higher computational complexity, better generalization, and / or is applicable in situations where the prerequisites of ILC (such as repeated actions and repeatable nuisance processes) do not apply.

[0397] The technique according to the invention becomes particularly evident when a counterweight (eg 50 kg) is mounted on the detector, thereby increasing the deflection of a swingable arm, in particular a C-arm, but after a few movements the image quality improves again.

[0398] Unless otherwise expressly stated, the various embodiments or various aspects and features described with reference to the accompanying drawings may be combined or interchanged without limiting or expanding the scope of the invention, as long as such combination or interchange is useful and in accordance with the spirit of the invention. The advantages described with respect to a certain embodiment of the invention or a certain drawing are, where applicable, also the advantages of other embodiments of the invention.

[0399] The cited prior art (the contents of which are respectively incorporated into this application by reference)

[0400] [1]Hans P.Geering et al.: "Entwurf robuster Regler mit der H ￥ -Methode", Regelungstechnik, Bulletin SEV / VSE 3 / 99, pp. 55-58.

[0401] [2] Buss: Intelligente Verfahren: Identification und Regelungnichtlinearer Systeme, Springer 2017.

[0402] [3] DABristow, M. Tharayil and AGAlleyne, "A survey of iterative learning control," in IEEE Control Systems Magazine, vol. 26, no. 3, pages 96-114, June 2006, doi: 10.1109 / MCS.2006.1636313.

[0403] [4] D.Shen, "Iterative learning control with incomplete information:asurvey," in IEEE / CAA Journal of Automatica Sinica, vol.5, no.5, pages 885-901, September 2018, doi:10.1109 / JAS.2018.7511123.

[0404] [5] Nard Strijbosch, Tom Oomen, “Iterative learning control for intermittently sampled data: Monotonic convergence, design, and applications,” Automatica, Vol 139, May 2022.

[0405] [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.

[0406] [7] Weifeng Liu, José C. Principe, Simon Haykin: Kernel Adaptive Filtering: A Comprehensive Introduction, Wiley Science, 2010.

[0407] [8] JL Rojo - Alvarez, M Martinez - Ramon, J Munoz - Mari, G Camps - Valls: Digital Signal Processing with Kernel Methods, Wiley - IEEE Press, 2018.

[0408] [9] Stefan R. Friedrich, Michael Schreibauer, Martin Buss, ″Least - square policy iteration algorithms for robotics: Online, continuous, and automatic,” Engineering Applications of Artificial Intelligence, Volume 83, 2019, pp. 72 - 84, https: / / doi.org / 10.1016 / j.engappai.2019.04.001.

[0409]

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

[0410]

[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, August 2012.

[0411]

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

[0412]

[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.

[0413]

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

[0414]

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

[0415]

[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.

Claims

1. A computer-implemented method (100) for controlling a compensating actuator to compensate for a shadow of an X-ray scatter absorption grating in an X-ray image, the X-ray image being recorded with a swingably mounted X-ray machine, in particular in real time, comprising the method steps of: - receiving (S102) motion state data, the motion state data indicating the motion state of the X-ray machine, in, The X-ray machine is mounted on a swingable arm, and wherein the movement state corresponds to a point in time at which an X-ray image is taken with the aid of the X-ray machine; - determining (S104) a quality criterion of the X-ray image in terms of a shadow of an X-ray scatter absorption grating, the X-ray scatter absorption grating being arranged on a detector of an X-ray machine; - determining (S106) a control signal for a compensating actuator mechanism for adjusting the position of the X-ray scatter absorption grating based on the received (S102) motion state data and based on the determined (S104) quality criterion, wherein the determining (S106) of the control signal includes an optimization of the quality criterion; and - sending (S108) the determined (S106) control signal to the compensation actuator mechanism to compensate for the shadow of the X-ray scatter absorption grating.

2. The method (100) according to claim 1, wherein: The pivotable arm comprises a C-arm.

3. The method (100) according to any one of the preceding claims, wherein: The position of the X-ray scatter absorption grating arranged on the detector includes a position in the detector plane and / or a position in a plane perpendicular to the focal direction of the X-ray source of the X-ray machine.

4. The method (100) according to any one of the preceding claims, wherein: The movement state includes the position, in particular the orientation, of the swivel arm and / or the x-ray machine in space and optionally the speed of its position adjustment, in particular in each degree of freedom.

5. The method (100) according to any one of the preceding claims, wherein: The received ( S102 ) motion state data includes setpoint motion state data and / or actual motion state data.

6. The method (100) according to any one of the preceding claims, wherein: Compensation of the shadow includes tilting and / or moving the X-ray scatter absorption grating, and / or tilting and / or moving the X-ray source.

7. The method (100) according to any one of the preceding claims, wherein: The compensation actuator mechanism includes two compensation actuators, wherein the first compensation actuator is designed for tilting and / or moving the X-ray scatter absorption grating along a first direction of the detector plane, and the second compensation actuator is designed for tilting and / or moving the X-ray scatter absorption grating along a second direction of the detector plane.

8. The method (100) according to at least one of the preceding claims, wherein: The method further comprises the steps of: - issuing (S110) the captured X-ray image, including compensation (S106) for the shadow of the X-ray scattering absorption grating, in particular in real time.

9. The method (100) according to any one of the preceding claims, wherein: The quality criteria include quality criteria in every direction, in particular in every direction along the detector plane.

10. The method (100) according to any one of the preceding claims, wherein: The determination of the quality standard (S104) includes a comparison with one or more data records in the vocabulary, wherein one or each data record in the vocabulary comprises a motion state, a quality criterion and optionally a control signal for compensating for shadows of the X-ray scatter absorption grating; furthermore, optionally, Therein, one or each data record in the vocabulary includes a time sequence of a motion state, a quality criterion and optionally a control signal.

11. The method (100) according to any one of the preceding claims, wherein: The determination of the quality criterion ( S104 ) includes an evaluation of the contrast in the X-ray image, in particular in the Fourier transformed X-ray image.

12. The method (100) according to any one of the preceding claims, wherein: At least one step of determining ( S104 ) a quality criterion for the X-ray image and determining ( S106 ) a control signal for compensating the actuator is performed repeatedly, in particular iteratively, preferably until the determined ( S104 ) quality criterion lies within a predetermined value range.

13. The method (100) according to any one of the preceding claims, wherein: The determination of the control signal for compensating the actuation mechanism (S106) includes - Machine Learning (ML), and / or Reinforcement Learning (RL); -Iterative Learning Control ILC; - Adaptive kernel identification feed-forward extrapolation; and / or -End-to-end learning.

14. A computing device (200) for controlling a compensating actuator mechanism to compensate for the shadowing of an X-ray scatter absorption grating in an X-ray image, the X-ray image being taken with a swingably mounted X-ray machine, in particular in real time, the computing device comprising: a receiving interface (202) configured to receive movement state data, the movement state data indicating the movement state of an X-ray machine, wherein the X-ray machine is mounted on a pivotable arm and wherein the movement state corresponds to a point in time at which an X-ray image is taken by means of the X-ray machine; a quality criterion determination unit (204) designed to determine a quality criterion of the X-ray image with respect to a shadow of an X-ray scatter absorption grating, the X-ray scatter absorption grating being arranged on a detector of the X-ray machine; a control signal determination unit (206) designed to determine a control signal for a compensating actuator for adjusting the position of the X-ray scatter absorption grating according to the received motion state data and according to the determined quality criterion, wherein the determination of the control signal includes an optimization of the quality criterion; and - a control signal output interface (208) designed to output a determined control signal to a compensation actuator to compensate for the shadow of the X-ray scattering absorption grating.

15. The computing device (200) according to the preceding claim, wherein: The computing device (200) comprises the features of one or more of claims 2 to 13 and / or wherein the computing device (200) is designed to perform the steps of the method of one or more of claims 2 to 13.

16. A system for controlling a compensating actuator in order to compensate for the shadowing of an X-ray scatter absorption grating in an X-ray image, the X-ray image being taken with a swingably mounted X-ray machine, in particular in real time, the system comprising: - an X-ray machine mounted on a swingable arm, comprising an X-ray scattering absorption grating; - A computing device according to claim 14 or 15 and A compensating actuator, wherein the receiving interface of the compensating actuator is designed to receive a determined control signal originating from a control signal output interface (208) of the computing device (200).

17. A computer program product with program instructions, which, when loaded into the memory of a computing device (200), allow the computing device (200) to control a compensating actuator according to any one of the above method claims in order to compensate for the shadow of an X-ray scattering absorption grating in an X-ray image, which is taken in real time using a swingably mounted X-ray machine.

18. A computer-readable medium in which program instructions are stored, the program instructions being readable by a computing device (200) and being designed to, when the program instructions are executed by the computing device (200), perform the steps of a method for controlling a compensating actuator according to any one of the above method claims in order to compensate for the shadow of an X-ray scattering absorption grating in an X-ray image, the X-ray image being taken in real time using a swingably mounted X-ray machine.

Images