Scatter correction imaging using noisy scatter estimations and scatter interpolation

Abstract

The invention relates to a method for correcting initial images (10) acquired by means of radiation using an imaging device (2 to 6) with the aid of a trained machine learning algorithm (11). Several initial images (10) are acquired. In addition, a smaller number of first, high-quality scattered radiation images (14') are simulated from the initial images. Furthermore, a corresponding number of second, low-quality scattered radiation images (13') are simulated, wherein the simulation is performed with a number of photons reduced by at least one order of magnitude. The algorithm is trained with the second, low-quality scattered radiation images (13') as input data and the first, high-quality scattered radiation images (14') as output data.

Description

[0001] The present invention relates to a method for training a machine learning algorithm to correct images acquired using an imaging device. Furthermore, the present invention relates to a method for correcting images acquired using radiation. The present invention also relates to a corresponding imaging system and a computer program.

[0002] Modern imaging techniques, particularly X-ray-based imaging, are used for diagnostic purposes and to support interventions. Structures within the object under investigation with high X-ray absorption can produce high X-ray contrast. Tools or other devices inserted into or placed within the object under investigation, for example, within a vascular structure, can also produce high X-ray contrast. Objects with high X-ray absorption that produce high contrast are also referred to as high-contrast objects.

[0003] To support diagnostic purposes, achieving the highest possible image quality is desirable. To enable the most accurate tracking of a device, for example, with regard to a vascular structure, and thus the most precise guidance of the device within the object, achieving the highest possible image quality is also desirable. Particularly in the context of X-ray-based imaging procedures, it can sometimes be difficult to clearly identify structures or the device itself and to distinguish them from other components of the image, such as depictions of tissue or bone structures, or even the vascular structure. The same applies to the distinguishability of the vascular structure from other tissue or the like.

[0004] Scattered X-ray photons play a crucial role in image quality for both 2D X-ray projections and 3D reconstructions, such as cone-beam computed tomography (CBCT). Without adequate compensation, scattered radiation drastically degrades achievable image quality, leading to fringe, cupping (highlighting in the center), and smearing artifacts, which can potentially result in incorrect diagnosis or treatment. Here and in the following, the terms radiation and X-rays are used interchangeably. To prevent image quality degradation caused by scattered radiation, known as scatter artifacts, modern commercial CBCT systems typically employ an anti-scatter grating in front of the detector, which physically blocks the incoming X-rays. Anti-scatter gratings are particularly effective at blocking scattered radiation. However, this grating also has adverse effects.First, it also blocks some of the non-scattered primary radiation, which can increase the applied dose (due to automatic adjustment). Second, improved image quality and a lower dose can be achieved in 2D imaging by removing the anti-scatter grid and relying solely on air gap technology. This is particularly relevant for neurovascular interventions in the brain, such as the treatment of aneurysms or embolic strokes. To avoid the use of anti-scatter grids, a dedicated software solution for scatter compensation in CBCT imaging is desirable.

[0005] It is known in the art to use deep neural networks for computed tomography image reconstruction (e.g., from WU, Dufan; KIM, Kyungsang; LI, Quanzheng: Computationally efficient deep neural network for computed tomography image reconstruction. In: Medical Physics, Vol. 46, 2019, No. 11, pp. 4763-4776. - ISSN 0094-2405. URL: https: / / aapm.onlinelibrary.wiley.com / doi / epdf / 10.1002 / mp.13627 [accessed 27.05.2025]).

[0006] Recently, methods based on so-called "deep learning" have been proposed to compensate for scatter in the projection area (Maier, J., Eulig, E., Vöth, T., Knaup, M., Kuntz, J., Sawall, S. and Kachelrieß, M. (2019), Real-time scatter estimation for medical CT using the deep scatter estimation: Method and robustness analysis with respect to different anatomies, dose levels, tube voltages, and data truncation. Med. Phys., 46: 238-249. https: / / doi.org / 10.1002 / mp.13274). While these methods are extremely fast, their overall robustness is questionable due to their dependence on the training data. Because of this dependence, it is difficult to cover all possible combinations of X-ray physics, collimation, focal spot size, and so on in a combined training cohort. Furthermore, such methods can only be trained with simulated pairs of input and output data, creating a significant gap between simulation and reality.

[0007] The de facto gold standard is the simulation of X-ray physics, either through a direct and deterministic solution of the Boltzmann transport equation or through a stochastic approximation of image formation using Monte Carlo methods. While these methods provide highly precise results, their application in the diagnostic and interventional setting is questionable due to their inherently high computational complexity and thus long computation times.

[0008] The same applies to so-called empirical methods, which estimate and optimize X-ray scattering in the reconstructed image layers based on an image quality metric. These methods involve many reconstruction steps, which also leads to high computational complexity.

[0009] From publication US 2021 / 0330274A1, a computer-implemented method for correcting X-ray image data with respect to noise effects is known. A statistical physical model, parameterized with model parameters, is used to describe the noise effects.

[0010] Against this background, it is an object of the present invention to provide an improved concept for imaging, in particular for x-ray-based imaging, by which high image quality can be achieved with reduced computational complexity.

[0011] This problem is solved by the respective subject matter of the independent claims. Advantageous further developments and preferred embodiments are the subject matter of the dependent claims.

[0012] According to the invention, a method is provided for correcting initial images obtained by means of an imaging device using radiation and for training a machine learning algorithm (in particular a neural network) for the correction. The imaging device can, for example, be an X-ray imaging device in which the images are obtained using X-rays. In particular, it can be a CT scanner (computed tomography), such as a C-arm scanner.

[0013] In a first step of the inventive process, a plurality of source images with different acquisition coordinates are acquired using the imaging device; that is, images are obtained. The subsequent image processing is based on these source images. Different acquisition coordinates of the imaging device are used for acquiring these source images. For example, the images are acquired from different angles with respect to the object being imaged. In this case, the angular coordinate forms the corresponding acquisition coordinate.

[0014] The recording coordinates can also be linear coordinates or other types of coordinates.

[0015] In a further step of the inventive method, a smaller number of first scattered radiation images than the majority of the original images are simulated using a scattering model derived from the majority of the original images. The simulation is performed with a first number of photons, and each first scattered radiation image is assigned (at least) one corresponding acquisition coordinate. Thus, scattered radiation images are generated from the previously acquired original images by simulation, the quality of which depends on the first number of photons. The number of scattered radiation images generated or simulated is not the same as the number of original images, but rather a reduced number of first scattered radiation images compared to the majority of the original images. This reduces the simulation effort.For example, if 500 projection images or source images are acquired, only ten initial scattered radiation images are simulated using the scattering model. The number of scattered radiation images can be reduced by a factor of 2, 3, and so on, preferably by at least an order of magnitude, compared to the number of source images. The simulation can be performed using a Monte Carlo method or by deterministic calculation. Each simulated initial scattered radiation image is assigned a corresponding acquisition coordinate or set of acquisition coordinates. For example, a simulated initial scattered radiation image is assigned an acquisition angle of a C-arm device as its acquisition coordinate. From this coordinate, the angle of incidence of the radiation on the object under investigation is derived in the scattering model.

[0016] In a further step, an equal or greater number of second scattered radiation images are simulated using the scattering model from the majority of the original images. This simulation uses a second set of photons reduced by at least one order of magnitude compared to the first set. Because the second scattered radiation images are simulated with a reduced number of photons, both the computational effort required for their simulation and their quality are reduced. Quality, in this context, refers to accuracy and physical correctness. Therefore, the first scattered radiation images can be described as high-quality, while the second scattered radiation images can be described as low-quality.Here and in the following, the terms first scatter radiation images and high-quality scatter radiation images are used synonymously, as are the terms second scatter radiation images and low-quality scatter radiation images.

[0017] The acquisition coordinates of the second, low-quality scatter radiation images correspond to those of the first, high-quality scatter radiation images. Thus, a small number (fewer than the number of original images) of low-quality scatter radiation images are generated by simulation. The number of low-quality scatter radiation images equals the number of high-quality scatter radiation images.

[0018] In the final step, an equal number of secondary, low-quality scatter radiation images can be simulated, just as many as the original images. From this larger number of low-quality scatter radiation images, the required number of low-quality scatter radiation images can then be selected. Specifically, those low-quality scatter radiation images are selected, or generated through simulation, whose acquisition coordinates correspond to those of the simulated high-quality scatter radiation images. For example, the high-quality and low-quality scatter radiation images are each simulated for a small number of acquisition angles.

[0019] In the final step, second, lower-quality scattered radiation images can also be simulated using different methods or scattering models. The number of second scattered radiation images would then additionally depend on the number of simulation models used. For example, using two different simulation models would result in twice the number of second scattered radiation images being simulated.

[0020] Simulating the second, lower-quality scattered radiation images does not necessarily have to be done directly from the original images, nor does simulating the first, higher-quality scattered radiation images necessarily have to be done directly from the original images. Rather, as will be explained in more detail below, a 3D image can first be reconstructed from the original images. In this embodiment, the second, lower-quality scattered radiation images, and optionally also the first, higher-quality scattered radiation images, are then simulated from this 3D image using the scattered radiation model.

[0021] Finally, the machine learning algorithm is trained using the second set of low-quality scattered radiation images as input data and the first set of high-quality scattered radiation images as output data. The machine learning algorithm can be a neural network. Hereinafter, the term "neural network" is also used to refer to other machine learning algorithms. The neural network is thus trained with the reduced number of low-quality scattered radiation images and their corresponding high-quality scattered radiation images. The low-quality scattered radiation images effectively form the "input layer," and the simulated high-quality scattered radiation images form the "output layer."The neural network is therefore trained using a reduced number of pairs of simulated low-quality and high-quality scattered radiation images. This reduces the computational effort required.

[0022] If multiple simulation models were used to simulate the second, lower-quality scattered radiation images, then several pairs of simulated low-quality and high-quality scattered radiation images can be used to train the neural network. For example, if two simulation models were used to simulate low-quality scattered radiation images, then for each high-quality scattered radiation image, one low-quality radiation image from each of the two simulation models can be used to train the neural network. In other words, in this example, for each high-quality scattered radiation image, there would be two different pairs, each with different low-quality scattered radiation images. The different simulation models could, for example, have different physically relevant properties.In this way, the training of the neural network can be further improved by taking into account the different physically relevant properties of the simulation models.

[0023] In addition to simulating some high-quality (HQ) noise-free scatter images, some or even an entire stack (corresponding to the number of source images) of noisy, low-quality (LQ) scatter images are also simulated. These LQ scatter images are simulated with a lower photon count compared to the HQ scatter images. Therefore, the additional simulation time required to estimate the LQ scatter images is negligible. These LQ scatter images can be used in various ways, for example, to train a neural network based solely on the LQ scatter images and a selection of source images (this is not claimed here).

[0024] In this case, LQ scattered radiation images are the sole input for the neural network (or algorithm or artificial intelligence) trained to derive HQ scattered radiation images. The neural network is trained on a few LQ-HQ pairs and then directly applied to (all) LQ scattered radiation images. In this way, the neural network can be interpreted either as a denoising algorithm or as a surface fitting algorithm (comparable, for example, to B-spline fitting, except that it has a data-specific component in the learned Kl weights).

[0025] In one embodiment, it is provided that for the simulation of the high-quality and low-quality scattered radiation images, an uncorrected 3D image is reconstructed from the majority of the source images, and the high-quality and low-quality scattered radiation images are simulated directly on the basis of the uncorrected 3D image.

[0026] First, an uncorrected 3D image of the object to be imaged is created from the uncorrected source images. This uncorrected 3D image forms the basis for the simulation of the scattered radiation images. Alternatively, the simulation could also be performed directly from the source images themselves, without the intermediate step of reconstructing a 3D image. However, the reconstructed 3D image has the advantage that further simulated scattered radiation images could be obtained without additional effort.

[0027] In a further embodiment, the imaging device is an X-ray device, and in particular a C-arm device, with which the initial images are acquired. The imaging device is thus based on X-ray technology. The imaging modality can therefore be designed, in particular, as an X-ray imaging modality, for example, as a digital X-ray imaging device, and in particular as a C-arm X-ray imaging modality (i.e., a C-arm device). The imaging modality includes, in particular, an X-ray source and a sensor unit. The sensor unit can, for example, include a detector array, in particular a two-dimensional detector array, of optical detectors, in particular photodiodes, which can generate the at least one sensor data set.

[0028] In the embodiment using the C-arm device, the simulated scatter radiation images can be evenly distributed across the detection range of the C-arm device with respect to the acquisition coordinates. For example, if a detection range of 200 degrees is selected for the C-arm device, a scatter radiation image can be simulated for every 20 degrees. This would result in eleven simulated scatter radiation images for the ten intervals, including an initial and final scatter radiation image. However, the even distribution can also be based on a different number of intervals and is not limited to ten. Alternatively, a deviation from an even distribution is possible if, for example, the scatter radiation changes particularly strongly in a specific angular range. In this case, more simulations can be performed for this range of increased change than in other ranges.

[0029] In a specific embodiment, the acquisition range of the C-arm device is greater than 180 degrees, and the number of simulated scatter radiation images is less than 30, and particularly less than 20. Typically, several hundred source images (e.g., 500) are acquired over an acquisition range of 180 degrees. For training purposes, a scatter radiation image is simulated for only every 25th source image. This allows the training to be performed significantly faster than if it were carried out with all source images and their corresponding scatter radiation images.

[0030] In another embodiment, the algorithm or neural network is calibrated to a physical model before training. Such calibration can be performed, for example, as described in the above-mentioned publication US 2021 / 0330274 A1. With this pre-calibration, the training does not have to start completely from scratch. Rather, the individual weights of the neural network are already given a sensible preset value based on the physical conditions (e.g., type of imaging area, patient type, X-ray dose, etc.).

[0031] According to another embodiment, from the majority of the source images, those are selected that have the same acquisition coordinates as the simulated low-quality and high-quality scattered radiation images, and these selected source images are also used to train the algorithm. Thus, real projections and their corresponding low-quality scattered radiation images are provided as additional input to the algorithm or AI during the (intraoperative) training phase, which is trained to derive high-quality scattered radiation images. In the (intraoperative) inference phase, all projections and all their corresponding low-quality scattered radiation images are fed into the now-trained AI to derive all their corresponding high-quality scattered radiation images.In this case, the (few) selected source images and (few) LQ scattered radiation images are used as input data, and the (few) HQ scattered radiation images are used as output data for training the algorithm. This allows the algorithm to be trained on the specific type of source images and their content.

[0032] In another embodiment, an expected value and a standard deviation are determined for each pixel of all high-quality scattered radiation images during simulation. These expected and standard deviation values ​​are then used to train the machine learning algorithm. This means that statistical values ​​such as the expected value and standard deviation can be used to obtain additional plausibility checks when applying the algorithm. For example, intraoperative training can be provided using high-quality (HQ) standard deviations as additional input data. Based on, for example, rejection samples, the resulting high-quality scattered radiation images and standard deviations can be adjusted range by range so that the low-quality (LQ) scattered radiation images represent probable samples from the distribution.

[0033] In a specific implementation example, the simulation is performed using Monte Carlo simulation. When conducting Monte Carlo simulations, the simulation results (e.g., HQ scattering images) are understood as an approximation of the arithmetic expected value. Therefore, each scattering pixel value can be understood as the expected value of that pixel. Furthermore, the variance or standard deviation for each pixel can be estimated simultaneously. Thus, a probability density function based on normal distributions (defined by expected value / mean and standard deviation) can be formulated for each pixel. Based on, for example, rejection sampling, the validity of artificially determined distributions (Kl-based prediction of mean and standard deviation, i.e., HQ scattering images and standard deviations) can be verified using real samples (i.e.,LQ or noisy scattered radiation images) are evaluated.

[0034] According to the invention, a method for correcting images obtained using (X-ray) radiation is provided by simulating a low-quality scattered radiation image for each initial image, creating a high-quality scattered radiation image for each initial image using an algorithm (e.g., a neural network) trained according to the above method, into which all initial images and all low-quality scattered radiation images are input, and correcting each initial image using the respective corresponding high-quality scattered radiation image. Thus, the neural network, trained with a few simulated (high- and low-quality) scattered radiation images, is provided to estimate corresponding scattered radiation images for all initial images, so that all initial images can be corrected with the respective scattered radiation images.To generate an individual scatter radiation image for each of the input images using the neural network, it is only necessary, according to the invention, to train the neural network with a smaller number of pairs of input images and scatter radiation images. Correcting each input image can be achieved, for example, by subtracting the corresponding scatter radiation image from the input image. Alternatively, further processing of the scatter radiation images, such as weighting or similar operations, can be performed for correction purposes.

[0035] When applying the algorithm or neural network to generate all high-quality scattered radiation images (corresponding to the number of source images), only the low-quality scattered radiation images corresponding to the number of source images are input to the algorithm if it has been trained solely on these low-quality scattered radiation images. Otherwise, if the algorithm has been trained on both the low-quality scattered radiation images and the source images, applying the algorithm to generate all high-quality scattered radiation images requires input of all source images as well as the corresponding low-quality scattered radiation images.

[0036] According to one embodiment, the number of simulated scattered radiation images used for training can be automatically increased until a predefined accuracy in pixel values ​​relative to the acquisition coordinates of adjacent scattered radiation images is achieved. For example, if the pixel values ​​of predetermined pixels in adjacent scattered radiation images fluctuate by more than a predefined amount, this can indicate an insufficiently trained neural network. Therefore, in this case, the number of training pairs of source images and scattered radiation images is increased. This increase can be carried out successively until the pixel values ​​of individual pixels in adjacent scattered radiation images no longer deviate excessively from each other, i.e., the error is below a predefined threshold.

[0037] In yet another embodiment, when generating each high-quality scattered radiation image, each pixel is checked to see if its corresponding pixel value lies within an interval defined by its respective expected value and standard deviation. If not, the pixel value is interpolated using pixel values ​​from immediately adjacent pixels. This allows specific distributions to be considered and applied individually to pixels. Using the simulated statistical values, which are also used to train the algorithm, the algorithm can then reliably generate high-quality scattered radiation images.

[0038] In another embodiment, a corrected 3D image is reconstructed from the corrected source images. For example, the source images of a C-arm sequence were corrected according to the method described above, thereby reducing the scattering effects in the individual images. These corrected individual images, or source images, can now serve as the basis for reconstructing a 3D image to obtain a corrected 3D image in which the scattering effects are also correspondingly reduced.

[0039] According to the invention, a method as described above can be performed while (X-ray) images of an object are being acquired. This means that, in effect, intraoperative training of the neural network takes place. Specifically, it is trained while X-ray images of the object are being acquired, for example, during an intervention. The advantage of this method is that a neural network is trained specifically for a particular patient or image, thus enabling highly accurate correction. Currently, neural networks are trained based on a large dataset, in the hope of achieving generalizability to unseen data. While this eliminates the need for intraoperative simulations, the results are often suboptimal for individual patients.

[0040] According to the invention, the problem described above is also solved by an imaging system comprising an imaging device and an image processing device configured to perform one of the methods described above. In particular, the imaging system can include a control unit configured to train an algorithm or a neural network for correcting tasks acquired by means of the imaging device. The imaging device can be configured to acquire a plurality of source images with different acquisition coordinates.The imaging device can be configured to generate a smaller number than the majority of the original images of simulated high-quality and low-quality scattered radiation images from the majority of the original images using a scattered radiation model, whereby each pair of simulated scattered radiation images is assigned (at least) one respective acquisition coordinate. The imaging device or its control unit can, in particular, be configured to select those images from the majority of the original images whose respective acquisition coordinates correspond to those of the simulated scattered radiation images, and to train the neural network with the selected original images as input data and the simulated scattered radiation images as output data.

[0041] For use cases or application situations that may arise during the procedure and are not explicitly described here, it may be provided that, according to the procedure, an error message and / or a request for user feedback is issued and / or a default setting and / or a predetermined initial state is set.

[0042] Further embodiments of the imaging system according to the invention result from the various configurations of the methods according to the invention, and vice versa. In particular, the imaging system according to the invention can be configured to perform a method according to the improved training or correction concept according to the invention.

[0043] According to the improved concept of the invention, a computer program (product) with instructions is also provided. When the instructions are executed by an imaging system according to the improved concept, in particular by a control or computing unit of the imaging system, the instructions cause the imaging system to perform a method according to the improved concept.

[0044] The computer program product can be designed as a computer program containing the commands. Alternatively, the computer program product can be designed as a computer-readable storage medium or electronically readable data carrier that stores a computer program containing the commands.

[0045] The features and combinations of features mentioned above in the description, as well as those subsequently mentioned in the figure description and / or shown in the figures alone, are usable not only in the combinations specified but also in other combinations without departing from the scope of the invention. Embodiments and combinations of features that do not exhibit all the features of an originally formulated independent claim and / or that go beyond or deviate from the combinations of features set out in the cross-references of the claims are also to be considered disclosed.

[0046] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

[0047] The figures show: Fig. 1 a schematic representation of an exemplary embodiment of an imaging device according to the improved concept; Fig. 2 a schematic visualization of an embodiment of a training method according to the invention; Fig. 3 a visualization of a correction method according to the invention; Fig. 4 a schematic visualization of another embodiment of a training method according to the invention; Fig. 5 a visualization of another correction method according to the invention; Fig. 6 a schematic visualization of yet another embodiment of a training method according to the invention; and Fig. 7 a visualization of yet another correction method according to the invention.

[0048] The exemplary embodiments described in more detail below represent preferred embodiments of the present invention.

[0049] In Fig. Figure 1 schematically depicts an exemplary embodiment of an imaging system 1 according to the improved concept, which is designed, for example, as an X-ray imaging system. In the example of the Fig. Figure 1 depicts a construction of the X-ray imaging system or device based on the principle of a C-arm device with a rotatable and movable C-arm 6, which can be rotated and moved accordingly to image an object 4 from different directions, i.e., with different angles of view. However, an imaging device 1 according to the improved concept can also be constructed according to other designs. In particular, the improved concept is not fundamentally limited to X-ray-based imaging methods.

[0050] Imaging device 1 of the Fig. Component 1 includes, for example, an X-ray source 2, which is configured to generate X-rays and emit them towards object 4. On the opposite side of object 4 from the X-ray source 2, a sensor 3 of the imaging device 1 is arranged, which, for example, contains a detector array of photodiodes to detect X-ray quanta penetrating object 4. The sensor 3 can then transmit the corresponding detector signals, for example, to a control or processing unit 5 of the imaging system 1 for further processing. Components 2 to 6 constitute an imaging device that can transmit its images to an image processing device 9. This device can make any necessary corrections to the acquired images.

[0051] Imaging system 1 can be configured, for example, to perform a rotational angiography procedure, such as one based on the principle of subtraction angiography. In this case, processing unit 5 can, for example, generate a large number of two-dimensional projections (also called source images) acquired from different angles, and processing unit 5 can then calculate a three-dimensional reconstruction from these projections.

[0052] The following section explains in more detail the functioning of the imaging system 1 with reference to various embodiments of a training and / or correction procedure according to the improved concept of the invention, in particular with reference to the Fig. 2 to 7.

[0053] The ones related to the Fig. The three exemplary implementations described in sections 2 to 7 can each be divided into two essential sub-steps: a training procedure and the actual correction procedure. Collectively, they represent neural scatter interpolations (NSI). They utilize, for example, the advantages of deep learning and simulation techniques. Although it is generally known that deep learning models perform exceptionally well in the training cohort, this effect is rarely exploited intentionally. NSI can be implemented as a purely interventional technique that does not depend on prior information.

[0054] The following describes the two essential steps in the exemplary embodiment according to Fig. 2 and Fig. 3 explained in more detail. In the first step according to Fig. In step 2, a plurality of output images 10 are acquired by the imaging device 2 to 6. Each output image 10 is acquired, for example, by a C-arm at a different angle. In the present example, each output image 10 represents an X-ray image of a chest. Each individual output image 10 is characterized by a primary intensity Ip of the primary (X-ray) radiation and a secondary intensity Is of the secondary (X-ray) radiation. In each individual output image 10, the two intensities Ip and Is are initially superimposed (Ip+Is). The aim is to remove the secondary intensity Is from each individual output image 10. To this end, according to Fig. 2 a neural network 11 is trained to eventually according to Fig. 3 to use for a correction of the stack of source images 10.

[0055] In the first step (training phase) according to Fig. 2 Optionally, a first, uncorrected reconstruction 12 (3D image) is calculated using the entire stack of projections, i.e., the initial images 10. Based on this uncorrected reconstruction 12, low-quality scatter radiation images 13 can be generated based on a first photon count. (Intensity IsLQ) The acquisition coordinates of the initial images 10 are simulated, which result (uniformly) along the acquisition trajectory (e.g., CBCT trajectory). For example, for every tenth acquisition position, a selection is made from the set of low-quality scattered radiation images 13, resulting in a few selected low-quality scattered radiation images 13'. Furthermore, based on the uncorrected reconstruction 12 using a second photon number greater than the first photon number, a number of selected high-quality scattered radiation images 14' are generated. (Intensity IsHQ) The simulation uses the same acquisition coordinates as the selected low-quality scatter radiation images 13'. The number of selected high-quality scatter radiation images 14' is therefore also smaller (e.g., by at least an order of magnitude) than the total number of original images 10. The simulation is performed using a scattering model, for example, based on the uncorrected reconstruction 12 or the stack of original images 10.

[0056] Experiments have shown that, for example, ten to twenty scattering simulations, i.e., simulated selected scatter radiation images 13' or 14', are sufficient to cover the angular range of a short scan (200 degrees).

[0057] The selected scatter radiation images 13', 14' are simulated for specific acquisition coordinates (e.g., acquisition angle) of the acquisition trajectory. Preferably, the acquisition coordinates, or the corresponding simulated selected scatter radiation images 13', 14', are uniformly distributed along the acquisition trajectory. The number of simulated selected low-quality scatter radiation images 13' and the number of simulated selected high-quality scatter radiation images 14' are equal.

[0058] Now, algorithm 11 for machine learning (e.g., the neural network) is trained using the simulations, i.e., the simulated, selected low-quality scattered radiation images 13' and the corresponding selected high-quality scattered radiation images 14'. For the training, the selected low-quality scattered radiation images 13' serve as the so-called "input layer" and the selected high-quality scattered radiation images 14' as the so-called "output layer".

[0059] The architecture of the neural network or algorithm 11 is freely selectable, however, a combination with a physically informed neural network, as disclosed in publication US 2021 / 0330274 A1, is recommended.

[0060] Training of neural network 11 can be a purely interventional or intraoperative technique and preferably does not depend on prior information.

[0061] In the second essential step (application phase) according to Fig. 3. For each of the low-quality scattered radiation images 13 (obtained in the first step), a high-quality scattered radiation image 14 is now calculated using the trained neural network 11. While the original images 10 have the intensity Ip+Is, the calculated high-quality scattered radiation images have the simulated intensity or intensity distribution. Intensity IsHQ, which corresponds to the unnoticed scattered radiation intensity Is. They thus represent the The scattering distribution of the secondary radiation at the respective acquisition coordinate is represented. The scattering distributions of the individual high-quality scattering images 14 can be used directly to correct all acquired projections or original images 10. By correcting the original original images 10 with the same number of high-quality scattering images 14, corrected original images 16 of the same number are obtained. In the correction, for example, the corresponding high-quality scattering image 14 is subtracted from an original image 10 to obtain a respective corrected original image 16. With respect to the intensity distribution, the correction can be expressed as follows: (Ip + Is) - Is = Ip. Thus, each corrected original image 16 is free of the respective specific scattering radiation.From the corrected source images 16, which correspond in number to the original source images, a corrected 3D image 17 can now be reconstructed. This corrected 3D image 17 is free of dispersion errors.

[0062] Fig. Figure 4 shows another embodiment for training algorithm 11 for machine learning. The training is essentially the same as that of the embodiment shown in Figure 4. Fig. 2. Therefore, for the description of this embodiment, reference is made to the description of Fig. 2. Additionally, a set of selected initial images 10' is chosen from the total set of initial images 10. The selected initial images 10' have the same acquisition coordinates as the selected low-quality scattered radiation images 13' and the selected high-quality scattered radiation images 14'. For training purposes, the selected initial images 10' are provided to algorithm 11 as input data, in addition to the selected low-quality scattered radiation images 13'. The selected high-quality scattered radiation images 14' continue to represent the output data for training algorithm 11.

[0063] The one according to the exemplary embodiment of Fig. 4 trained algorithm 11 can be used according to the example of Fig. 5. The application is analogous to the example of Fig. 3. However, in addition to all low-quality scatter radiation images 13, all original images 10 are also fed into algorithm 11 as input data in order to obtain all high-quality scatter radiation images 14. From the original images 10, which were already provided for input to algorithm 11, the high-quality scatter radiation images 14 are now subtracted for correction to obtain the corrected original images 16. The latter then form the basis for the reconstruction of the 3D image 17.

[0064] Another example of the training of algorithm 11 is in Fig. Figure 6. It is based on the embodiment of Fig. 4, which is why reference is made to the description there. The only difference is that in the simulation of the selected high-quality scattered radiation images, the standard deviations are also calculated pixel by pixel. σ(IsHQ) are calculated. This results in selected, extended high-quality scatter radiation images 15', which are based on both the high-quality scatter radiation intensities. IsHQ as well as the associated standard deviations σ(IsHQ) These selected, extended high-quality scatter radiation images 15' are used in this example as input data for training algorithm 11. The input data for training corresponds to that of the example from Fig. 4.

[0065] The application of the trained algorithm 11 is exemplified in Fig. 7 is shown. It is based on the application according to the example of Fig. 5, which is why reference is made to the description there. In contrast to the example of Fig. 5 is used in the example of Fig. 7. Using the trained algorithm 11, a complete set of enhanced high-quality scattered radiation images 15 is calculated from the low-quality scattered radiation images 14 and the original images 10. The number of these enhanced high-quality scattered radiation images 15 corresponds to the number of low-quality scattered radiation images 13 and the number of original images 10. All enhanced high-quality scattered radiation images 15 contain, in addition to the respective intensities, IsHQ (arithmetic means) also the standard deviations σ(IsHQ). This application can also be performed intraoperatively.

[0066] In a post-processing or iterative refinement step 18, the artificial high-quality scattered radiation images can be iteratively adjusted based on the standard deviations, so that the actually simulated low-quality scattered radiation images 13 are meaningful samples from these normal distributions.

[0067] The exemplary embodiment according to Fig. Method 2 can also be easily implemented as a pre-trained AI that can be trained on any type of dispersion simulation. Mapping LQ dispersions to HQ dispersions is inherently simpler than mapping real-world projection images to HQ dispersions, as it is merely a denoising or surface-fitting method based on imperfect samples.

[0068] In an extended embodiment, the neural network 11 can be pre-trained or initialized, for example, with a physical model. This has the advantage that the neural network initialized in this way is already ready for intraoperative training according to the model. Fig. 2 can be used.

[0069] According to another embodiment, an expected accuracy of the pixel values ​​of the simulated scattered radiation images 13 can be defined. If this accuracy cannot be achieved during intraoperative training, additional simulations can be performed until it is reached. This means that the number of simulated scattered radiation images 13 is automatically increased, for example, until the desired accuracy is achieved. The approach according to the invention makes it possible to combine online simulation, training, and the productive use (inference) of the neural network within a single intervention procedure. This leads to the following advantages: Firstly, compared to purely simulation-driven approaches, the computational complexity can be reduced by a factor of up to 50.Since the network training is performed with only about 10 data sets, for example, the runtime is negligible, as is its application effort.

[0070] Secondly, neural scatter interpolation (NSI) is inherently robust compared to pure deep learning methods. It inherently ensures that no data outside the distribution occurs. Furthermore, the expected accuracy can be monitored during network training based on the training data or additional validation data. If the expected accuracy does not meet a defined quality criterion, additional simulations can be performed until the criterion is met.

[0071] Advantageously, a method can thus be provided that incorporates low-quality dispersion estimates into the training and / or inference process for AI-based dispersion estimation. In particular, low-quality dispersion estimates can be used as known samples from an unknown distribution. This enables robust and fast dispersion estimations, as low-quality dispersion patterns serve as guardrails in the training and inference process.

[0072] The preceding description is intended to include persons of male, female or other gender identities, regardless of the grammatical gender of a particular term.

Claims

[1] A method for correcting initial images (10) obtained by means of radiation using an imaging device (2 to 6) by simulating a second scattered radiation image (13) for each initial image (10), creating a first scattered radiation image (14) for each initial image (10) using a trained machine learning algorithm (11) into which the initial images (10) and second scattered radiation images (13) to be corrected are input, and correcting the respective initial image (10) using the respective corresponding first scattered radiation image (14), thereby obtaining corrected initial images (16), wherein the training of the machine learning algorithm (11) is characterized by - Acquiring a plurality of source images (10) with different acquisition coordinates using the imaging device (2 to 6), - Simulating a number less than the plurality of source images (10) of first scattered radiation images (14, 14') using a scattering model from the plurality of source images (10), wherein the simulation is performed with a first number of photons, and wherein each first scattered radiation image (14, 14') is assigned a respective acquisition coordinate, - Simulating an equal or greater number of second scattered radiation images (13, 13') using the scattering model from the plurality of initial images (10), wherein the simulation is performed with a second number of photons reduced compared to the first number, in particular by at least one order of magnitude, and wherein the respective acquisition coordinates of the second scattered radiation images (13, 13') correspond to those of the first scattered radiation images (14, 14'), - Training the machine learning algorithm (11) with the second scatter radiation images (13, 13') as input data and the first scatter radiation images (14, 14') as output data. [2] Method according to claim 1, wherein for simulating the first and second scattered radiation images (13, 13', 14, 14') an uncorrected 3D image (12) is reconstructed from the majority of the source images (10), and the first and second scattered radiation images (13, 13', 14, 14') are simulated directly on the basis of the uncorrected 3D image (12). [3] Method according to any of the preceding claims, wherein the imaging device (2 to 6) is an X-ray device and in particular a C-arm device with which the initial images (10) are obtained. [4] Method according to one of the preceding claims, wherein from the majority of the output images (10) those are selected which have the same acquisition coordinates as the simulated second and first scatter radiation images (13, 13', 14, 14'), and the selected output images (10') are also used for training the algorithm (11). [5] Method according to one of the preceding claims, wherein an expected value and a standard deviation value are determined for each pixel over all first scattered radiation images (14, 14') during simulation, and the expected values ​​and the standard deviation values ​​are used for training the algorithm (11) for machine learning. [6] Method according to any of the preceding claims, wherein the simulation is carried out by a Monte Carlo simulation. [7] Method according to claim 5, wherein, in the creation of each first scattered radiation image (14, 14'), it is checked pixel by pixel whether a corresponding pixel value lies within an interval defined by the respective expected value and respective standard deviation value, and, if this is not the case, the pixel value is interpolated with pixel values ​​of immediately adjacent pixels. [8] Method according to one of the preceding claims, wherein a corrected 3D image (17) is reconstructed from the corrected initial images (16). [9] Imaging system (1) comprising an imaging device (2 to 6) and an image processing device (9) configured to perform a method according to any of the preceding method claims. [10] Computer program which can be loaded directly into a memory of a control unit of an imaging system (1) according to claim 9, comprising program means to execute the steps of a method according to any of the preceding method claims when the program is executed in the control unit of the imaging device (2 to 6). [11] Electronically readable data carrier with electronically readable control information stored thereon, comprising a computer program according to claim 10 and designed such that, when the data carrier is used in a control device of an imaging system (1) according to claim 9, it performs a method according to one of the preceding method claims.

Images