COMPUTER-IMPLEMENTED METHOD FOR PROVIDING AN OUTPUT DATA SET WITH IMPROVED IMAGE QUALITY

Description

[0001] The invention relates to a computer-implemented method for providing an output data set with reduced noise compared to a selected input data set, wherein the selected input data set and at least one further input data set relating to the same subject under investigation are determined by a medical imaging method, and wherein the further input data set is or is registered with the selected input data set. The invention also relates to a provisioning device, a medical imaging device, a computer program, and a computer-readable data carrier.

[0002] In some medical imaging techniques, the same object or anatomy is acquired multiple times. This is also known as multiphase imaging. Multiple acquisitions can be used, for example, to capture different phases of organ movement, such as different phases of heart or respiratory activity. Additionally or alternatively, multiple images can be acquired with varying contrast agent concentrations, different image parameters, etc. This results in redundant information for at least some areas of the object, which can be used to reduce noise and / or dose.

[0003] Several approaches are already known in the art for combining image data from multiple imaging phases—that is, image data acquired at different times or with different parameters—to improve image quality. For example, the article by Brehm, M., et al. (2015), Cardiorespiratory motion-compensated micro-CT image reconstruction using an artifact model-based motion estimation. Med. Phys., 42: 1948-1958, doi:10.1118 / 1.4916083, proposes registering computed tomography images acquired at different times or phases of the imaging process by compensating for movements due to respiration and cardiac motion using a motion model. This provides a larger amount of image data for each individual reconstruction, as the separate phases do not need to be reconstructed separately, thus improving image quality.

[0004] The article by Sawall, S., et al. (2011), Low-dose cardio-respiratory phase-correlated cone-beam micro-CT of small animals. Med. Phys., 38: 1416-1424, doi:10.1118 / 1.3551993, utilizes a high-dimensional filter function to improve image quality. In addition to the three spatial dimensions of the individual image dataset, an additional dimension is used, in which the different cardiac and respiratory phases are arranged. Bilateral filtering is employed to prevent excessive blurring of edges in the image.

[0005] The use of bilateral filters to improve image quality is also described, for example, in the articles Ehman, EC, et al. (2012), Noise Reduction to Decrease Radiation Dose and Improve Conspicuity of Hepatic Lesions at Contrast-Enhanced 80-kV Hepatic CT Using Projection Space Denoising, AJR Am J Roentgenol. 2012 Feb;198(2):405-11, doi: 10.2214 / AJR.11.6987, and Tomasi, C., et al. (1998), Bilateral Filtering for Grey and Color Images, proceedings of the 1998 IEEE International Conference on Computer Vision, Bombay, India.

[0006] From the article Kang, E., et al. (2019), Cycle-consistent adversarial denoising network for multiphase coronary CT angiography, Med. Phys., 46: 550-562, doi:10.1002 / mp.13284, it is also known to use machine learning approaches to improve image quality.

[0007] DE 10 2012 206714 A1 discloses a method for reducing metal artifacts in CT image datasets.

[0008] EP 3 467 766 A1 discloses a medical image processing device comprising a memory configured to store multiple neural networks corresponding to multiple image target locations.

[0009] EP 3 244 368 A1 discloses a noise reduction method that achieves a compromise between minimal radiation exposure and noise reduction in multiple successive 3D CT scans. The input sequences are filtered according to attributes and then weighted and summed, where these attributes can correspond to frequencies.

[0010] The invention is based on the objective of providing an alternative approach to reduce noise under given imaging conditions.

[0011] The problem is solved according to the invention by a computer-implemented method according to claim 1.

[0012] Within the scope of the invention, it was discovered that significantly low noise can be achieved in the intermediate data set by combining image data from several input data sets registered with each other. In the simplest case, the image data from the different image data sets can be combined by averaging. However, other combination methods are also possible, for example, combination using a machine learning method or a bilateral filter, as will be explained in more detail later. A disadvantage of such noise reduction, however, is that not only noise but also real image features can be inadvertently eliminated from the intermediate data set, or at least be significantly less visible there than in the selected input data set.

[0013] However, it was recognized within the scope of the invention that this disadvantage can be significantly reduced or even completely eliminated by performing a frequency-dependent superposition of the selected input data set and the intermediate data set. This is implemented by applying at least one filter function to each of these data sets and summing the resulting filtered data sets, optionally weighted.

[0014] According to the invention, a low-pass filter is applied to the selected input data set and a high-pass filter to the intermediate data set. This ensures that low spatial frequencies are taken from the selected input data set, i.e., the unaltered original image, and high frequency components are taken from the intermediate data set, i.e., the pre-processed image. The often particularly relevant low spatial frequencies are thus reproduced without distortion, while still achieving a low noise level. Compared to applying only a low-pass filter to the selected input data set, the advantage is that, due to the provision of the high spatial frequencies via the intermediate data set, fine structures can be rendered in considerably greater detail.

[0015] The above-described use of exactly one filter function for the selected input data set or the intermediate data set is obviously only a very simple example of the procedure according to the invention. In many applications, it is advantageous to apply several filter functions, in particular the same filter functions, to the selected input data set and the intermediate data set in order to determine an image content for a multitude of frequency bands. This makes it possible to weight different frequency ranges of the various data sets differently as needed. This will be explained in more detail later.

[0016] The filter functions for providing the at least one filtered input data set and the at least one filtered intermediate data set should preferably be phase-preserving to avoid, for example, phase cancellation when different filtered data sets are superimposed. This can be achieved, for example, by transforming the selected input data set or the intermediate data set into the Fourier space, where phase-preserving filtering can be achieved simply by multiplication with an envelope. However, a corresponding phase-preserving filter can also be achieved by a convolution kernel, which can be provided, for example, by transforming a corresponding envelope in the frequency domain into the spatial domain via Fourier transformation.

[0017] The acquisition of the selected input data set and the other input data sets, and / or the registration of the other input data sets with respect to the selected input data set, can be carried out as preparatory steps before the commencement of the computer-implemented method according to the invention. For example, the acquired and optionally already registered input data sets can be provided via an input interface to a provisioning device or routine that implements the computer-implemented method. Alternatively, these steps, or even just the registration of the input data sets with respect to each other, can also be part of the method itself.

[0018] Preferably, a plurality of additional input data sets are used, for example, at least two, at least five, or at least ten additional input data sets. By registering these additional input data sets in relation to the selected input data set, it is ensured, in particular, that a given pixel or voxel at a specific position in the selected input data set and in all other input data sets displays approximately the same area of ​​the object under investigation. This can be achieved, in particular, by registering the additional input data sets elastically relative to the selected input data set.

[0019] The selected input data set and the other input data sets can, in particular, describe 3D image data, for example, reconstructed image data from a computed tomography (CT) scanner, a magnetic resonance imaging (MRI) scanner, a positron emission tomography (PET) scanner, or similar devices. The method according to the invention is thus applied particularly after the reconstruction of 3D image data. However, the method according to the invention can also be applied to 2D image data as input data sets. For example, 2D X-ray images or other 2D images from a medical imaging device can be processed. It is also possible, for example, to process individual images from a CT scanner using the computer-implemented method according to the invention before a reconstruction is performed.

[0020] The output data can be output to a user, further processed, or stored on a data carrier for later display or processing after the completion of the computer-implemented method according to the invention, or as part thereof. Further processing can, for example, include automatic segmentation or automatic feature recognition in the output data set.

[0021] The filtered input data sets can each be assigned to a frequency range of spatial frequencies, whereby spatial frequencies outside the respective assigned frequency range are suppressed or their amplitudes reduced relative to the selected input data set by applying the respective filter function. Additionally or alternatively, the filtered intermediate data sets can each be assigned to a frequency range of spatial frequencies, whereby spatial frequencies outside the respective assigned frequency range are suppressed or their amplitudes reduced relative to the intermediate data set by applying the respective filter function.

[0022] In particular, a filter function extracts the spatial frequencies of the assigned frequency range from the selected input data set or intermediate data set and discards or suppresses the other spatial frequencies. Specifically, the amplitudes of spatial frequencies outside the assigned frequency range are reduced by at least a factor of 1.5, but preferably by at least a factor of 2, 4, 8, or 16. Particularly when filtering is based on a convolution kernel or using a Fourier transform, a substantially complete suppression of frequencies outside the assigned frequency range can be achieved. Within the assigned frequency range, the amplitudes of the spatial frequencies can be essentially preserved, for example, by applying the filter function and reducing them by a maximum of 10%, 20%, or 50%.

[0023] By dividing the selected input dataset and the intermediate dataset into multiple frequency ranges, it is possible to choose in which frequency range data from the selected input dataset or data from the intermediate dataset is used in the output dataset or in the subsequent intermediate dataset, and how the selected input dataset and the intermediate dataset are weighted within this frequency range. In particular, a set of filter functions, whose assigned frequency ranges essentially cover the entire frequency spectrum, can thus be applied to both the selected input dataset and the intermediate dataset. By appropriately selecting the weighting factors for the resulting filtered input and intermediate datasets, a combination of the selected input dataset and the intermediate dataset with a frequency-dependent weighting can be implemented.

[0024] For the highest frequency range, either only a filtered intermediate data set can be determined, and no filtered input data set at all. Alternatively, when calculating the further intermediate data set or the output data set for the highest frequency range, the filtered intermediate data set can be weighted more heavily than the filtered input data set. This ensures that for the highest frequency range, only the image data of the intermediate data set is considered, or at least that the image data of the intermediate data set is weighted more heavily than the image data of the selected input data set. This is advantageous because, in the highest frequency range, the selected input data set is typically heavily influenced by noise, necessitating particularly strong noise reduction in this range. This can be achieved by exclusively or at least more heavily considering the image data of the intermediate data set.

[0025] In addition to or as an alternative to the above explanations regarding the highest frequency range, for a lowest frequency range either only a filtered input data set and no filtered intermediate data set can be determined, or when calculating the further intermediate data set or the output data set, the filtered input data set can be weighted more heavily than the filtered intermediate data set. In other words, in the lowest frequency range, only or at least to a greater extent the data of the selected input data set, i.e., the unaltered original data, can be used. This is advantageous because in the lowest frequency range, a relatively low level of noise is typically already present in the selected input data set, and essential features are also represented there, which, with the help of the frequency-dependent weighting according to the invention, ensures that these features are reproduced optimally.

[0026] The sum or weighted sum of at least one filtered input data set and at least one filtered intermediate data set can be used directly as the output data set, particularly if the registration of the other input data sets for the selected input data set is very good, resulting in no or only negligible registration errors. Such high registration quality can often only be achieved if additional information, such as ECG signals to compensate for cardiac motion, is available. If registration is to be based solely on the image data itself, for example, if only DICOM images are available, larger registration errors often result, which can lead to image errors due to registration artifacts in the determined output data set or in the other intermediate data set.

[0027] As will be explained later, some of these image errors can be avoided or at least reduced by taking appropriate measures during the acquisition of the intermediate data set. However, it may still be necessary or at least advisable to first acquire and process a further intermediate data set in order to reduce the influence of artifacts caused by registration errors.

[0028] In particular, a third intermediate data set can be determined by applying a further filter function to a difference between the selected input data set and the further intermediate data set, wherein the output data set is determined by subtracting the third intermediate data set, in particular after scaling, from the further intermediate data set or from the selected input data set.

[0029] The described approach exploits the fact that, apart from noise reduction and registration artifacts, the intermediate data set essentially corresponds to the selected input data set. The difference between the selected input data set and the further intermediate data set thus primarily represents noise and artifacts resulting from registration errors. Registration artifacts primarily lead to blurring or banding, while noise tends to cause grainy structures. By applying the additional filter function, it is possible to separate the noise and the artifact representation. Depending on the selected filter, it can therefore be achieved that the third intermediate data set represents almost exclusively artifacts due to registration errors or almost exclusively noise.

[0030] For example, a median filter can be used as an additional filter function, which largely eliminates the noise component in the difference between the selected input data set and the additional intermediate data set. Since the third intermediate data set in this case essentially represents artifacts due to registration errors, subtracting this third intermediate data set from the additional intermediate data set can essentially eliminate the contribution of the artifacts from the additional intermediate data set completely. To prevent overcompensation, it can be advantageous to scale the additional intermediate data set before this subtraction, particularly by a factor less than one.

[0031] Alternatively, if another filter function is used that is suitable for largely removing artifacts from the difference between the selected input data set and the subsequent intermediate data set, then ultimately only the noise component of the selected input data set remains in the third intermediate data set. This noise can then be subtracted from the selected input data set to eliminate the noise. Here, too, it can be advantageous to scale the third intermediate data set before this subtraction to avoid overcompensating for the noise.

[0032] The additional input data sets can be registered by an algorithm that performs elastic registration and / or is trained using a machine learning method. Elastic registration can eliminate distortions within the image, such as those caused by a heartbeat or respiration. Various approaches to the elastic registration of medical image data are known in the art and will not be discussed in detail here. As a purely illustrative example, reference is made to the article by M. Brehm et al. cited at the beginning, which uses elastic registration.

[0033] If the algorithm is trained using a machine learning method, it can be, for example, a neural network, a support vector machine, a decision tree, and / or a Bayesian network. Alternatively or additionally, the trained algorithm can also be based on a k-means algorithm, Q-learning, a genetic algorithm, and / or association rules. In particular, a neural network can be a deep neural network, a convolutional neural network, or a convolutional deep neural network. The neural network can also be an adversarial network, a deep adversarial network, and / or a generative adversarial network.

[0034] Methods for training the algorithm are explained below using supervised learning as an example. Alternatively, semi-supervised learning, reinforcement learning, active learning, and / or unsupervised learning would also be possible.

[0035] A key challenge in training an algorithm using supervised machine learning is typically providing sufficiently large amounts of training data. This problem can be at least partially solved for algorithms used to register input data by artificially modifying reference datasets. This involves artificially destroying an initially existing registration, after which the algorithm can be trained to recognize the parameters used to modify the dataset and thus restore the registration. For example, a reference dataset can be randomly rotated, shifted, scaled, and / or locally distorted, or, for instance, according to a known segmentation. The cost function can, for example, attempt to calculate the deviation between the determined registration and the actual registration.To minimize changes across a large amount of training data, the well-known approach of error feedback can be used, for example.

[0036] While training the algorithm can, in principle, be part of the inventive method, it is particularly preferred that the inventive method uses an algorithm that has already been trained before the inventive method is started. The algorithm can thus be trained in a separate process. Therefore, the invention also relates to a method for training an algorithm for registering further input data sets for the selected input data set. Furthermore, the invention relates to a training device configured for carrying out this method, a program that, when executed on a data processing device, performs the steps of this method, and a computer-readable data carrier that stores such a program. In addition, the invention relates to a trained algorithm for registering further input data sets for a selected input data set.Parameter data sets that parameterize a predefined algorithm for providing this algorithm, or a computer-readable data carrier that stores the algorithm or its parameters. The algorithm or its parameters, and thus also the data carrier, can in particular be provided as a product of the training procedure described.

[0037] The image data of a given pixel or voxel in the intermediate dataset can be determined based on the sum or weighted sum of the image data of the respective assigned pixel or voxel in the selected input dataset and the other input dataset, or at least one of the other input datasets. In particular, pixels at the same position in the respective dataset are assigned to each other. For example, a sum or weighted sum of the points at the same x- and y-position, and, in the case of three-dimensional datasets, the z-position, can be calculated.

[0038] The weighting of the summands serves primarily to define the extent to which pixels of a given additional input dataset are taken into account. The weighting can be the same for all pixels or voxels of the respective additional input dataset, for example, if the weighting depends on the time interval between the acquisition of the selected input dataset and the respective additional input dataset. This can be useful, for instance, if non-periodic movements of an object under investigation are expected, or the weighting can oscillate with the expected periodic movement.

[0039] Particularly advantageous is the use of weighting the summands as a supplementary or alternative method to preserve edges in the image. For example, even minor registration errors between individual data sets can lead to a situation where, in at least some of the subsequent input data sets, a pixel or voxel represents a different side of an edge than the corresponding pixel or voxel in the selected input data set. This can, for instance, result in a blurring of the edge when averaging the image data of the pixels.

[0040] Edges can be largely preserved if a contrast difference between the pixels is taken into account. For example, certain pixels or voxels of the other input datasets can be excluded if the image data at that point or voxel deviates too much from the image data of the associated pixel or voxel of the selected input dataset, for example, if a measure of the deviation exceeds a threshold.

[0041] Alternatively, the weighting of the image data for a given pixel or voxel of the respective input dataset could depend on the measure of the difference between this image data and the image data of the corresponding pixel or voxel of the selected input dataset. If weighting factors are influenced by multiple factors, such as a contrast difference and a temporal and / or spatial distance, this is also referred to as bilateral filtering due to the contrast dependence of non-linear filtering. Further details on such bilateral filtering can be found, for example, in the article by C. Tomasi cited at the beginning.

[0042] A pixel or voxel of the intermediate dataset can also depend on image data from pixels or voxels of the selected input dataset, or at least one of the other input datasets, that are located at a different position, particularly directly or indirectly adjacent. In addition to the explained sum or weighted sum of image data from pixels or voxels at the same position, which can be interpreted as filtering across multiple time points or phases, spatial filtering is also possible. This can be particularly advantageous when, as explained above, an edge-preserving or contrast-difference-compensating approach is used, as this allows even minor registration errors to be compensated for during the determination of the intermediate dataset.

[0043] Using an unweighted sum of the image data from the various input datasets and then dividing by the number of input datasets is equivalent to averaging. In the simplest case, the intermediate dataset is thus calculated by averaging the selected input dataset and at least one of the other input datasets.

[0044] The consideration of image data or weighting factors for the image data of the respective assigned pixel or voxel of the respective additional input dataset can depend on a measure of the difference between the image data of that pixel or voxel and the image data of the assigned pixel or voxel of the selected input dataset. In particular, the weighting of voxels of an additional input dataset can decrease as the difference in image data increases. It is also possible that image data from additional input datasets for which the measure of difference exceeds a threshold is not considered. As explained above, the described procedure ensures, in particular, that edges in the selected input dataset are not blurred by filtering across multiple input datasets.

[0045] In an advantageous embodiment of the method, the intermediate data set is determined by an algorithm trained using a machine learning method. As explained above regarding the algorithm for registering the further input data sets for the selected input data set, the training of the algorithm can preferably be carried out in a preparatory step before carrying out the method according to the invention. Thus, the already trained algorithm can be made available to enable the execution of the method according to the invention using this algorithm. The general statements regarding the training of algorithms by machine learning with respect to the algorithm for registration can also be applied to the algorithm for determining the intermediate data set.

[0046] For the sake of simplicity, the training of the algorithm for determining the intermediate data set will also be explained using the example of supervised learning. Training data sets for such supervised learning can be generated synthetically, for example, by adding synthetic noise to several sets of reference data, with the algorithm being trained to remove this noise. Data sets originating from the same reference data set are perfectly registered. They can therefore be used as the chosen input data set and other input data sets during training.

[0047] Alternatively, it would be possible, for example, to shift, rotate, scale, and / or distort some of these datasets to train the algorithm's behavior with respect to registration errors when generating the intermediate dataset. A cost function could, for instance, minimize the difference between the generated intermediate dataset and the reference data before noise is added.

[0048] The algorithm for generating the intermediate data set can ultimately be a filter kernel that is applied jointly to the selected input data set and the subsequent input data sets, with, for example, the recording order of the input data sets serving as an additional dimension. Such problems can be effectively represented, for instance, by a convolutional neural network as the structure of the algorithm to be trained.

[0049] As already explained, the training of the algorithm for determining the intermediate data set can be carried out in a separate process. The invention therefore also relates to a method for training an algorithm for determining an intermediate data set, a training device configured for carrying out this method, a computer program that, when executed on a data processing device, performs the steps of this method, and a computer-readable data carrier that stores such a program.

[0050] The invention further relates to a trained algorithm for providing the intermediate data set, or a parameter data set that parameterizes a given algorithm for providing this algorithm, or a computer-readable data carrier that stores this algorithm or parameter data set. The algorithm, the parameter data set, or the data carrier can, in particular, be provided as a product of the described training method.

[0051] At least one of the filter functions can be applied to both the selected input dataset and the intermediate dataset. The weighting factors used to weight the filtered input dataset and filtered intermediate dataset resulting from the application of the respective filter function add up to one when calculating the further intermediate dataset or the output dataset. This can apply to all filter functions. In particular, the application of a respective filter function, as explained above, can at least approximately restrict the spatial frequencies in the resulting filtered input dataset or filtered intermediate dataset to the assigned frequency range. Thus, the weighting factors, which add up to one, determine the weighting between the selected input dataset and the intermediate dataset for this frequency range.

[0052] Several filter functions can be applied to the selected input data set, chosen such that the sum of the filtered input data sets equals the selected input data set. Additionally or alternatively, several filter functions can be applied to the intermediate data set, chosen such that the sum of the filtered intermediate data sets equals the intermediate data set. In other words, the filter functions can be chosen to decompose the selected input data set or the intermediate data set into different image components, particularly frequency bands, with the different components then summing back to form the original image.Minor deviations can occur, for example, due to the accuracy of the number format used in the calculation, due to approximations in both filtering processes (such as short filter kernel lengths that can lead to oscillations outside the passband), or similar factors. In other words, the sum of the filtered input data sets can only be approximately equal to the selected input data set, and the sum of the filtered intermediate data sets can only be approximately equal to the intermediate data set.

[0053] The selected input dataset and at least one other input dataset can be recorded at different times during the same examination or treatment of the subject, specifically during different phases of movement of an organ of the subject, and / or with the same recording geometry. "Same treatment" or "examination" here refers in particular to procedures between which the patient is not intentionally repositioned. The recording of the additional input datasets for the selected input dataset can thus serve, in particular, to compensate for unintentional movements of the subject or at least approximately periodic movements, such as respiration or a heartbeat.

[0054] The selected input dataset and the subsequent input dataset(s) can, in particular, be different images from a multi-contrast imaging scan, be assigned to different cardiac and / or respiratory phases, and / or differ with respect to at least one acquisition parameter, for example, the X-ray spectrum, especially in the case of dual- or multi-energy computed tomography. It is also possible that different contrast agent concentrations are present in the subject in the various input datasets, or similar factors. In principle, input datasets that differ only due to unintentional subject movements can also be used.

[0055] In addition to the computer-implemented method according to the invention, the invention relates to a provisioning device configured for carrying out the computer-implemented method according to the invention. In particular, the provisioning device can comprise an input interface for receiving the selected input data set and at least one further input data set, as well as an output interface for outputting the output data set. It can be configured as a data processing device configured for carrying out the data processing described in connection with the method. The data processing can be carried out, for example, by a suitably programmed processor, such as a microprocessor, microcontroller, FPGA, DSP, or similar device, or distributed across a multitude of processors, for example, via a cloud.

[0056] The features described for the computer-implemented method according to the invention can be transferred to the provisioning device according to the invention with the aforementioned advantages, and vice versa.

[0057] Furthermore, the invention relates to a medical imaging device, in particular a computed tomography device, which includes a provisioning device according to the invention. This makes it possible, in particular, for the selected input data set and at least one further input data set to be processed by the computer-implemented method according to the invention immediately upon or subsequently upon their acquisition, for example, during a first visualization for an operator. Alternatively, instead of being integrated into the medical imaging device, the provisioning device can also be a separate device that receives input data and / or outputs data, for example, via a network or a database.

[0058] The invention also relates to a computer program for a data processing device with program instructions which, when executed on a data processing device, perform a computer-implemented method according to the invention.

[0059] Furthermore, the invention relates to a computer-readable data carrier comprising a computer program according to the invention.

[0060] When comparing image noise in multiphase images with image noise in standard images (one phase), with otherwise constant recording parameters, especially constant dose (in the individual phases), e.g. in phantom scans, a reduction in noise can be achieved in the multiphase images compared to the standard image when using the described algorithm.

[0061] When comparing noise reduction with soft and sharp input images using the described method, non-linearity can occur due to the frequency dependence of the phase averaging or the formation of the weighted sum.

[0062] The use of registered input data sets, or registration, can lead to smaller differences in noise reduction when comparing the use of identical input data sets with significantly different input data sets than would be the case without this registration.

[0063] Further advantages and details of the invention will become apparent from the following exemplary embodiments and the accompanying drawings. These schematically illustrate: Fig. 1 a flowchart of an embodiment of a computer-implemented method according to the invention, Fig. 2 the within the framework of the in Fig. 1 Figure 3 shows the algorithms and data structures used in the method shown, Figure 3 shows the interaction of an embodiment of a provisioning device according to the invention with a medical imaging device, Figures 4 and 5 show exemplary training processes for training algorithms usable in the computer-implemented method according to the invention by machine learning methods, and Figures 6 and 7 show exemplary structures of trained algorithms that can be used in the method according to the invention.

[0064] Fig. 1 Figure 1 shows a flowchart of a computer-implemented procedure for providing an output dataset with improved image quality compared to a selected input dataset. The procedure is further described below with additional reference to... Fig. 2 This section explains the algorithms and data structures used in the procedure, presenting them schematically. For clarity, the datasets are shown two-dimensionally, although the described procedure can be used with both two-dimensional and three-dimensional datasets. As described in Fig. 1 As represented by dashed lines, the procedure described below can be divided into three sections. Steps S1 and S2 can be performed as preparatory steps or optionally as part of the procedure. Steps S3 to S5 constitute the core of the described procedure, and steps S6 and S7 can be performed additionally, optionally, to minimize the influence of registration errors on the determined output data set.

[0065] In step S1, a selected input data set 33 and a multitude of further input data sets 34 are initially acquired. The input data sets 33 and 34 can, for example, be reconstructed three-dimensional data sets whose source data were acquired by a computed tomography scanner. However, any other two-dimensional or three-dimensional image data can also be processed. Here, the selected input data set 33 and the further input data sets 34 are acquired at different times during an examination of the same object, so that due to unintentional movements of the object or unavoidable movements, such as a heartbeat or respiration, the various input data sets 33 and 34 are not yet registered initially.Furthermore, the various input datasets 33, 34 may differ with respect to contrast agent concentration in the subject, imaging parameters used, and / or other factors. The different acquisitions at different times or under different conditions can also be referred to as phases of multiphase imaging.

[0066] In the further prepared step S2, an algorithm 35 registers the additional input data sets 34 for the selected input data set 33. The described procedure for improving the selected input data set 33 can, of course, be repeated for several of the input data sets 33, 34 in order to also improve the image quality of the additional input data sets 34.

[0067] Algorithm 35 can be trained using a machine learning method. An example of a possible training method will be given later with reference to Fig. 4 This will be explained. In particular, a large number of parameters 36 of the algorithm 35 can be determined during training. Alternatively, approaches known in the prior art for the elastic registration of image datasets can be used, which will not be explained in detail here.

[0068] In the resulting registered further input data sets 37, points or voxels 41, 42 at the same position in the selected input data set 33 and the registered further input data sets 37 essentially represent the same area of ​​the object under investigation.

[0069] In step S3, an intermediate data set 39 is determined by an algorithm 38, which is fed the selected input data set 33 and the registered further input data sets 37 as input data. Here, for each pixel or voxel 43 of the intermediate data set, whose position in the case of three-dimensional input data sets is specified by the coordinate tuple x, y, z, a value of the image data Z of the intermediate data set is determined as a function of the image data of the pixels or voxels 41, 42 at the respective same coordinate x, y, z of the selected input data set E<g< or 33 and the N different further input data sets E<w1< to E<wN<: Z x , y , z = g E x , y , z g , E x , y , z w 1 , … , E x , y , z wN

[0070] The function g is, in particular, a weighted sum of the aforementioned values. In the simplest case, the mean of these values, i.e., the mean of points or voxels 41 and 42, can be calculated to determine the image data of pixel or voxel 43.

[0071] The weights for the image data of pixel or voxel 42 of the other input data sets 37 can depend, in particular, on a measure of the difference between the image data of the respective pixel or voxel 42 and the image data of the associated pixel or voxel 41 of the selected input data set 33. Thus, the algorithm 38 implements, in particular, a non-linear, specifically bilateral, filter. In addition to the quantities mentioned, the function g can also depend on image data of pixels or voxels that are adjacent to pixel or voxel 41, 42. Thus, spatial filtering can also be performed.

[0072] Algorithm 38 can, for example, be specified by choosing a suitable filter kernel. However, it can be advantageous to train Algorithm 38 using a machine learning method, as will be discussed later with reference to… Fig. 5 This will be explained. In particular, it can be used to exploit the fact that filter functions can be very well implemented using a convolutional neural network.

[0073] In step S4, the filter functions 44, 45, and 46 are each applied to the selected input data set 33 to generate a respective filtered input data set 47, 48, and 49, respectively, and also to the intermediate data set 39 to generate a respective filtered intermediate data set 50, 51, and 52. The filter functions 44 to 46, of which only three are shown as examples, are each assigned to a specific frequency range and essentially completely suppress spatial frequencies outside the assigned frequency range in the respective filtered input and intermediate data sets 47 to 52.

[0074] For example, filtering can be performed in Fourier space, or a window function in Fourier space can be transformed into real space to provide a convolution kernel. The filter functions are chosen such that the sum of the filtered input data sets 47 to 49 corresponds to the selected input data set 33, and the sum of the filtered intermediate data sets 50 to 52 corresponds to the intermediate data set 39. In other words, both the selected input data set 33 and the intermediate data set 39 are decomposed into different frequency components by the filter functions 44 to 46.

[0075] In step S5, a further intermediate data set 60 is determined by weighting the filtered input data sets 47 to 49 and the filtered intermediate data sets 50 to 52 with weighting factors 53 to 58 and thereby summing an adder 59.

[0076] The described procedure ensures that the selected input data set 33 and the intermediate data set 39 are weighted frequency-dependently to generate the further intermediate data set 60. Image noise in intermediate data set 39 is largely suppressed by taking the further input data sets 37 into account. In frequency ranges where intermediate data set 39 is heavily weighted, strong noise reduction can therefore be achieved. This is particularly relevant for high frequencies.

[0077] At the same time, by giving greater or exclusive consideration to the selected input data set 33 in certain frequency ranges, for example for low frequencies, it is ensured that features characteristic of these frequency ranges cannot be masked or attenuated by the noise reduction processing. Overall, very good noise reduction can thus be achieved without blurring or masking relevant features.

[0078] The weighting factors 53 to 58 are preferably chosen such that, for a given frequency range or filter, the respective weighting factors 53, 55, 57 for the filtered input data set 47, 48, 49 and the respective weighting factors 54, 56, 58 for the filtered intermediate data set 50, 51, 52 add up to one. Thus, the calculation of the further intermediate data set Z' or 60 can be expressed as follows: Z ′ = w 1 ⋮ w N ⋅ F 1 E g ⋮ F N E g + 1 − w 1 ⋮ 1 − w N ⋅ F 1 Z ⋮ F N Z

[0079] Here, F1 to Fn are the filter functions 44, 45, and 46, which are applied on the one hand to the selected input data set E<g< or 33 and on the other hand to the intermediate data set Z or 39. For each frequency band assigned to the individual filter functions F1 to Fn, a weighting factor w1 to wN is specified, indicating how strongly the selected input data set E<g< should be weighted in the respective frequency band compared to the intermediate data set Z.

[0080] Assuming that the registration of the registered additional input data records 37 to the selected input data record 33 is essentially error-free or leads only to negligible artifacts, the additional intermediate data record 60 can be used directly as the output data record 65. However, if, as in the example shown, registration is based solely on the image data of the input data records 33 and 34, this is typically not guaranteed. Therefore, it may be advisable to perform the optional further processing of the additional intermediate data record 60 in steps S6 and S7, as explained below.

[0081] In step S6, a difference 61 is first calculated between the further intermediate data set 60 and the selected input data set 33. Since the further intermediate data set 60 corresponds at least approximately to the selected input data set 33, apart from the noise and artifacts due to faulty registration, the difference 61 approximately represents the sum of the noise component of the selected input data set 33 and the artifacts caused during processing.

[0082] By applying a further filter function 62, for example a median filter, which at least largely filters out the noise component from the difference 61, a third intermediate data set 63 is generated, which thus largely reflects the artifacts generated during processing.

[0083] In step S7, this can be subtracted from the further intermediate data set 60 by the subtractor 64 to determine the output data set 65. To avoid overcompensating for artifacts, the third intermediate data set 63 can be scaled before the subtraction. The determination of the output data set A or 65 from the further intermediate data set Z' or 60 can be expressed as follows: A = Z ′ − a ⋅ AF Z ′ − E g

[0084] Here, the function AF is the additional filter function and a is a scaling factor.

[0085] As an additional filter function, AF or 62, for example a median filter can be used, which is well suited to suppressing a noise component.

[0086] As an alternative to the described procedure, it would also be possible in principle to use a further filter function 62 that is configured to filter artifacts from the difference 61 and thus ultimately leave only the noise component of the further input data set 33 as the third intermediate data set 63. In this case, the third intermediate data set can be conveniently subtracted from the selected input data set 33 in order to remove the noise component directly.

[0087] Fig. 3 Figure 66 shows a medical imaging device, in this example a computed tomography device, which can acquire the selected input data set and the other input data sets relating to the object of examination 67. A control unit 68, which controls the operation of the imaging device 66, can provide corresponding measurement data via a network 69 to an evaluation unit 70, for example a workstation computer.

[0088] To improve the image quality of the selected input data set, a provisioning device 71 is used, to which the input data sets are provided and processed as explained above. The provisioning device 71 can be provided by appropriate programming of a data processing device 87. The resulting output data set can be provided directly to the evaluation device 70 or stored in a database 72. The input data can also be taken from the database 72 and do not necessarily have to be provided directly by the medical imaging device 66.

[0089] In the example shown, the data provisioning unit 71 is designed as a separate unit. However, it can also be advantageous to integrate the data provisioning unit 71 directly into the medical imaging unit 66 in order to be able to provide output data sets with improved image quality immediately. For example, the control unit 68 can also serve as a data provisioning unit.

[0090] As explained above, algorithms 35 and 38 for registering further input data sets for the selected input data set and for determining the intermediate data set, respectively, can each be trained using a machine learning method. Such training preferably takes place on a separate training facility. Therefore, in the example, a training facility 73 for training algorithm 35 and a training facility 74 for training algorithm 38 are shown. While in the example shown the algorithms or the parameters that parameterize them are provided via network 69, it is also possible to install them elsewhere on the provisioning facility 71, so that the permanent network connection shown between the provisioning facility 71 and the training facilities 73 and 74 is not necessary.

[0091] Fig. 4 Figure 1 shows a possible configuration of a training procedure for training algorithm 35 for registering the input data sets. Here, several reference data sets 75 are initially provided, which are shifted, rotated, scaled, and / or distorted by a deformation algorithm 76 depending on predefined parameter sets 77. Each pair consisting of one of the several reference data sets 75 and one of the parameter sets 77 thus generates a modified reference data set 78.

[0092] Algorithm 35 then attempts, with initial parameterization, to register the modified reference data sets 78 individually or in groups with each of the reference data sets 75. A cost function 80, which is to be minimized, can compare the registration results 79 with the reference data sets 75 to determine a measure of their difference. The parameters 36 are varied such that the cost function, and thus the measure of the deviation, is minimized. For example, the well-known approach of error feedback can be used for this purpose.

[0093] Fig. 5 This section presents an approach for training algorithm 38 to determine the intermediate data set 39. The intermediate data set 39 is designed to reduce noise by considering image data from multiple input data sets that depict the same object and are registered relative to each other. A large number of training data sets can be generated by adding noise images 83, generated by a noise generator 82, to reference data sets 81. This results in a multitude of noisy reference data sets 84 that are at least partially registered relative to each other and depict the same object. At least some of the noise data sets 84 that are registered relative to each other and depict the same object are selected as input data sets, and further input data sets are fed to algorithm 38 to generate a respective intermediate data set 85.The denoised intermediate data set 85 is compared with the respective reference data set 81, and a cost function 86, which describes a measure of the difference between the intermediate data set 85 and the associated reference data set 81, is minimized by varying the parameters 40 of the algorithm 38. This can, in turn, be done, for example, by error feedback.

[0094] With regard to the Figuren 6 and 7 The following section explains the structures of algorithms that can be trained using machine learning, using the example of a neural network and a convolutional neural network. For example, corresponding structures can be used for algorithms 35 and 38. The structures are described using simple examples that can be extended for real-world applications.

[0095] Fig. 6 Figure 1 shows an embodiment of an artificial neural network 1. English terms for the artificial neural network 1 are "artificial neural network", "neural network", "artificial neural net" or "neural net".

[0096] Artificial neural network 1 comprises nodes 6 to 18 and edges 19 to 21, where each edge 19 to 21 is a directed connection from a first node 6 to 18 to a second node 6 to 18. Generally, the first node 6 to 18 and the second node 6 to 18 are distinct nodes 6 to 18; however, it is also conceivable that the first node 6 to 18 and the second node 6 to 18 are identical. For example, in Fig. 6 Edge 19 is a directed connection from node 6 to node 9, and edge 21 is a directed connection from node 16 to node 18. An edge 19 to 21 from a first node 6 to 18 to a second node 6 to 18 is called an incoming edge for the second node 6 to 18 and an outgoing edge for the first node 6 to 18.

[0097] In this embodiment, the nodes 6 to 18 of the artificial neural network 1 can be arranged in layers 2 to 5, wherein layers 2 to 5 can have an intrinsic order introduced by the edges 19 to 21 between nodes 6 to 18. In particular, edges 19 to 21 can only be provided between adjacent layers of nodes. In the illustrated embodiment, there is an input layer 2 that contains only nodes 6, 7, and 8, each without an incoming edge. The output layer 5 comprises only nodes 17 and 18, each without outgoing edges, with hidden layers 3 and 4 further situated between the input layer 2 and the output layer 5. In the general case, the number of hidden layers 3 and 4 can be chosen arbitrarily.The number of nodes 6, 7, 8 in input layer 2 usually corresponds to the number of input values ​​into neural network 1, and the number of nodes 17, 18 in output layer 5 usually corresponds to the number of output values ​​of neural network 1.

[0098] In particular, nodes 6 to 18 of neural network 1 can be assigned a (real) number. Here, x(n) < i denotes the value of the i-th node 6 to 18 of the n-th layer 2 to 5. The values ​​of nodes 6, 7, 8 of input layer 2 are equivalent to the input values ​​of neural network 1, while the values ​​of nodes 17, 18 of output layer 113 are equivalent to the output values ​​of neural network 1. Furthermore, each edge 19, 20, 21 can be assigned a weight in the form of a real number. In particular, the weight is a real number in the interval [-1, 1] or in the interval [0, 1,]. Here, w (m,n)< i,j denotes the weight of the edge between the i-th node 6 to 18 of the m-th layer 2 to 5 and the j-th node 6 to 18 of the n-th layer 2 to 5. Furthermore, the abbreviation w i , j n for the weight w i , j n , n + 1 defined.

[0099] To calculate the output values ​​of neural network 1, the input values ​​are propagated through neural network 1. In particular, the values ​​of nodes 6 to 18 of the (n+1)th layer 2 to 5 can be calculated based on the values ​​of nodes 6 to 18 of the nth layer 2 to 5 by x j n + 1 = f ∑ i x i n ⋅ w i , j n .

[0100] Here, f is a transfer function, which can also be called an activation function. Well-known transfer functions include step functions, sigmoid functions (for example, the logistic function, the generalized logistic function, the hyperbolic tangent, the arctangent, the error function, the smoothstep function), and rectifier functions. The transfer function is primarily used for normalization purposes.

[0101] Specifically, the values ​​are propagated layer by layer through neural network 1, with values ​​of input layer 2 being given by the input data of neural network 1. Values ​​of the first hidden layer 3 can be calculated based on the values ​​of input layer 2 of neural network 1, values ​​of the second hidden layer 4 can be calculated based on the values ​​in the first hidden layer 3, and so on.

[0102] To determine the values w i , j n To determine the values ​​for edges 19 to 21, neural network 1 must be trained using training data. Specifically, training data includes training input data and training output data, referred to below as ti. For a training step, neural network 1 is applied to the training input data to determine computed output data. Specifically, the training output data and the computed output data comprise a number of values, where the number is determined by the number of nodes 17 and 18 in output layer 5.

[0103] In particular, a comparison between the calculated output data and the training output data is used to recursively adjust the weights within neural network 1 (backpropagation algorithm). Specifically, the weights can be adjusted accordingly. w ′ i , j n = w i , j n − γ ⋅ δ j n ⋅ x i n to be changed, where y is a learning rate and the numbers δ j n can be calculated recursively as δ j n = ∑ k δ k n + 1 ⋅ w j , k n + 1 ⋅ f ′ ∑ i x i n ⋅ w i , j n based on δ j n + 1 , if the (n+1)th layer is not the starting layer 5, and δ j n = x k n + 1 − t j n + 1 ⋅ f ′ ∑ i x i n ⋅ w i , j n if the (n+1)th layer is the output layer 5, where f' is the first derivative of the activation function and y j n + 1 the comparison training value for the j-th node 17, 18 of the initial layer 5 is.

[0104] The following will be discussed with regard to Fig. 7 An example of a Convolutional Neural Network (CNN) is also given. It's important to note that the term "layer" is used slightly differently there than in classical neural networks. For a classical neural network, the term "layer" refers only to the set of nodes that form a layer, i.e., a specific generation of nodes. For a Convolutional Neural Network, the term "layer" is often used to describe an object that actively modifies data; in other words, a set of nodes of the same generation and either the set of incoming or outgoing edges.

[0105] Fig. 7Figure 1 shows an embodiment of a convolutional neural network 22. In the illustrated embodiment, the convolutional neural network 22 comprises an input layer 23, a convolutional layer 24, a pooling layer 25, a fully connected layer 26, and an output layer 27. In alternative embodiments, the convolutional neural network 22 can contain multiple convolutional layers 24, multiple pooling layers 25, and multiple fully connected layers 26, as well as other types of layers. The order of the layers can be chosen arbitrarily, with fully connected layers 26 typically forming the last layers before the output layer 27.

[0106] In particular, within a Convolutional Neural Network 22, the nodes 28 to 32 of one of the layers 23 to 27 can be arranged as a d-dimensional matrix or as a d-dimensional image. Specifically, in the two-dimensional case, the value of a node 28 to 32 with indices i, j in the nth layer 23 to 27 can be denoted as x(n) < [i,j]. It should be noted that the arrangement of the nodes 28 to 31 of a layer 23 to 27 has no effect whatsoever on the computations within the Convolutional Neural Network 22 itself, since this effect is determined solely by the structure and the weights of the edges.

[0107] A convolution layer 24 is particularly distinguished by the fact that the structure and weights of the incoming edges form a convolution operation based on a specific number of kernels. In particular, the structure and weights of the incoming edges can be chosen such that the values x k n node 29 of the folding layer 24 as a folding x k n = K k * x n − 1 based on the values ​​x (n-1)< of the nodes 28 of the preceding layer 23, where the convolution * in the two-dimensional case can be defined as x k n i j = K k * x n − 1 i j = ∑ i ′ ∑ j ′ K k i ′ , j ′ ⋅ x n − 1 i − i ′ , j − j ′ .

[0108] In this example, the k-th kernel Kk is a d-dimensional matrix, in this embodiment a two-dimensional matrix that is typically small compared to the number of nodes 28 to 32, for example, a 3x3 matrix or a 5x5 matrix. In particular, this implies that the weights of the incoming edges are not independent but are chosen to generate the convolution equation above. In the example for a kernel forming a 3x3 matrix, there are only nine independent weights (where each entry of the kernel matrix corresponds to an independent weight), regardless of the number of nodes 28 to 32 in the corresponding layer 23 to 27. Specifically, for a convolution layer 24, the number of nodes 29 in the convolution layer 24 is equivalent to the number of nodes 28 in the preceding layer 23 multiplied by the number of convolution kernels.

[0109] If the nodes 28 of the preceding layer 23 are arranged as a d-dimensional matrix, the use of multiple kernels can be understood as adding another dimension, also called the depth dimension, so that the nodes 29 of the convolution layer 24 are arranged as a (d+1)-dimensional matrix. If the nodes 28 of the preceding layer 23 are already arranged as a (d+1)-dimensional matrix with one depth dimension, the use of multiple convolution kernels can be understood as an expansion along the depth dimension, so that the nodes 29 of the convolution layer 221 are likewise arranged as a (d+1)-dimensional matrix, the size of the (d+1)-dimensional matrix in the depth dimension being larger than in the preceding layer 23 by the factor formed by the number of kernels.

[0110] The advantage of using convolutional layers 24 is that the spatially local correlation of the input data can be exploited by creating a local connection pattern between nodes of neighboring layers, in particular by ensuring that each node has connections only to a small area of ​​the nodes of the preceding layer.

[0111] In the illustrated embodiment, the input layer 23 comprises 36 nodes 28 arranged as a two-dimensional 6x6 matrix. The convolution layer 24 comprises 72 nodes 29 arranged as two two-dimensional 6x6 matrices, each of which is the result of convolution of the values ​​of the input layer 23 with a convolution kernel. Similarly, the nodes 29 of the convolution layer 24 can be understood as arranged in a three-dimensional 6x6x2 matrix, the latter dimension being the depth dimension.

[0112] A pooling layer 25 is characterized by the fact that the structure and weights of the incoming edges, as well as the activation function of their nodes 30, define a pooling operation based on a non-linear pooling function f. For example, in the two-dimensional case, the values ​​x(n)< of the nodes 30 of the pooling layer 25 can be determined based on the values ​​x(n+1)< of the nodes 29 of the preceding layer 24 as x n i j = f x n − 1 id 1 , jd 2 , … , x n − 1 id 1 + d 1 − 1 , jd 2 + d 2 − 1 The number of nodes 29, 30 can be calculated by using a pooling layer 25. This is achieved by replacing a number of d1*d2 neighboring nodes 29 in the preceding layer 24 with a single node 30, which is calculated as a function of the values ​​of the aforementioned number of neighboring nodes 29. Specifically, the pooling function f can be a maximus function, an averaging function, or the L2 norm. In particular, for a pooling layer 25, the weights of the incoming edges can be fixed and not modified by training.

[0113] The advantage of using a pooling layer 25 is that the number of nodes 29, 30 and the number of parameters are reduced. This leads to a reduction in the required computational effort within the Convolutional Neural Network 22 and thus to the suppression of overfitting.

[0114] In the illustrated embodiment, pooling layer 25 is a max-pooling layer in which four adjacent nodes are replaced by a single node whose value is the maximum of the values ​​of the four adjacent nodes. Max-pooling is applied to each d-dimensional matrix of the preceding layer; in this embodiment, max-pooling is applied to each of the two two-dimensional matrices, thus reducing the number of nodes from 72 to 18.

[0115] A fully connected layer 26 is characterized by the presence of a plurality, in particular all, edges between the nodes 30 of the preceding layer 25 and the nodes 31 of the fully connected layer 26, whereby the weight of each edge can be individually adjusted. In this embodiment, the nodes 30 of the layer 25 preceding the fully connected layer 26 are shown both as two-dimensional matrices and as non-connected nodes (represented as a row of nodes, the number of which has been reduced for clarity). In this embodiment, the number of nodes 31 in the fully connected layer 26 is equal to the number of nodes 30 in the preceding layer 25. In alternative embodiments, the number of nodes 30 and 31 can be different.

[0116] Furthermore, in this embodiment, the values ​​of the nodes 32 of the output layer 27 are determined by applying the softmax function to the values ​​of the nodes 31 of the preceding layer 26. By applying the softmax function, the sum of the values ​​of all nodes 32 of the output layer 27 is one, and all values ​​of all nodes 32 of the output layer are real numbers between 0 and 1. If the convolutional neural network 22 is used to classify input data, the values ​​of the output layer 27 can, in particular, be interpreted as the probability that the input data falls into one of the different classes.

[0117] A Convolutional Neural Network 22 can also have a ReLU layer, where ReLU is an acronym for "rectified linear units". Specifically, the number of nodes and the structure of the nodes within a ReLU layer are equivalent to the number of nodes and the structures of the nodes in the preceding layer. The value of each node in the ReLU layer can be calculated, in particular, by applying a rectifier function to the value of the corresponding node in the preceding layer. Examples of rectifier functions are f(x) = max(0,x), the hyperbolic tangent, or the sigmoid function.

[0118] Convolutional Neural Networks 22 can be trained in particular based on the backpropagation algorithm.

[0119] To avoid overfitting, regularization procedures can be used, for example, dropout of individual nodes 28 to 32, stochastic pooling, use of artificial data, weight decay based on the L1 or L2 norm, or maximum norm restrictions.

[0120] Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited by the disclosed examples and other variations can be derived by the person skilled in the art within the scope of the attached claims without departing from the scope of protection of the invention.

Claims

1. Computer-implemented method for providing an output dataset (65) with a reduced noise compared with a selected input dataset (33), wherein the selected input dataset (33) and at least one further input dataset (34, 37) relating to the same examination object (67) are determined by a medical imaging method, wherein the selected input dataset (33) and the at least one further input dataset (34, 37) are recorded at different points in time during the same examination or treatment of the examination object (67) and / or with recording parameters which differ from one another, wherein the further input dataset (34, 37) is or will be registered to the selected input dataset, which comprises the following steps: - determining an intermediate dataset (39), wherein the image data of the image points or voxels (43) of the intermediate dataset (39) in each case depend on the image data of a respective assigned image point or voxel (41, 42) of the selected input dataset (33) and the further input dataset (37), wherein on account of the combination of the image data of the several input datasets registered with one another a reduced noise compared with the selected input dataset (33) is achieved in the intermediate dataset (39), - applying at least one filter function (44-46) to the selected input dataset (33), in order to generate a respective filtered input dataset (47-49), - applying at least one filter function (44-46) to the intermediate dataset (39) in order to generate a respective filtered intermediate dataset (50-52), and - determining the output dataset (65) or a further intermediate dataset (60), as a function of which the output dataset (65) is determined, as a total or weighted total of the at least one filtered input dataset (47-49) and the at least one filtered intermediate dataset (50-52).

2. Computer-implemented method according to claim 1, characterised in that the filtered input datasets (47-49) are in each case assigned to a frequency range of spatial frequencies, wherein spatial frequencies outside of the respective assigned frequency range are suppressed compared with the selected input dataset (33) by applying the respective filter function (44-46) or their amplitudes are reduced, and / or the filtered intermediate datasets (50-52) are in each case assigned to a frequency range of spatial frequencies, wherein spatial frequencies outside of the respective assigned frequency range are suppressed compared with the intermediate dataset (39) by applying the respective filter function (44-46) or their amplitudes are reduced.

3. Computer-implemented method according to claim 2, characterised in that - on the one hand, for a highest frequency range, either exclusively a filtered intermediate dataset (50-52) and no filtered input dataset (47-49) is determined or with the calculation of the further intermediate dataset (60) or the output dataset (65), the filtered intermediate dataset (50-52) is weighted more heavily than the filtered input dataset (47-49), and / or - on the other hand, for a lowest frequency range, either exclusively a filtered input dataset (47-49) and no filtered intermediate dataset (50-52) is determined or with the calculation of the further intermediate dataset (60) or the output dataset (65), the filtered input dataset (47-49) is weighted more heavily than the filtered intermediate dataset (50-52).

4. Computer-implemented method according to one of the preceding claims, characterised in that a third intermediate dataset (63) is determined by a further filter function (62) being applied to a difference (61) between the selected input dataset (33) and the further intermediate dataset (60), wherein the output dataset (65) is determined by the third intermediate dataset (63), in particular in accordance with a scaling, being subtracted from the further intermediate dataset (60) or from the selected input dataset (33).

5. Computer-implemented method according to one of the preceding claims, characterised in that a number of further input datasets (34, 37) are provided, which will be or are registered with the selected input dataset (33) by an algorithm (35), said input dataset carrying out an elastic registration and / or being trained by a machine learning method.

6. Computer-implemented method according to one of the preceding claims, characterised in that the image data of a respective image point or voxel (43) of the intermediate dataset (39) is determined as a function of the total or the weighted total of the image data of the respective assigned image point or voxel (41, 42) of the selected input dataset (33) and of the further input dataset (37) or of at least one of the further input datasets (37).

7. Computer-implemented method according to claim 6, characterised in that the consideration of the image data or the weighting factor for the image data of the respective assigned image point or voxel (42) of the respective further input dataset (37) is dependent on a measure of a difference between the image data of this image point or voxel (42) and the image data of the assigned image point or voxel (41) of the selected input dataset (33).

8. Computer-implemented method according to one of the preceding claims, characterised in that the intermediate dataset (39) is determined by an algorithm (38) which is trained by a machine learning method.

9. Computer-implemented method according to one of the preceding claims, characterised in that at least one of the filter functions (44-46) is applied in each case both to the selected input dataset (33) and also to the intermediate dataset (39), wherein the weighting factors (53-58) add up to form one, with which weighting factors (53-58) the filtered input dataset (47-49) resulting from the application of the respective filter function (44-46) and filtered intermediate dataset (50-52) are weighted upon calculation of the further intermediate dataset (60) or the output dataset (65).

10. Computer-implemented method according to one of the preceding claims, characterised in that a number of filter functions (44-46) are applied to the selected input dataset (33), which are selected such that the total of the filtered input datasets (47-49) corresponds to the selected input dataset (33), and / or a number of filter functions (44-46) is applied to the intermediate dataset (39), said filter functions (44-46) being selected so that the total of the filtered intermediate datasets (50-52) corresponds to the intermediate dataset (39).

11. Computer-implemented method according to one of the preceding claims, characterised in that the selected input dataset (33) and the at least one further input dataset (34, 37) are or will be recorded at different points in time during the same examination or treatment of the examination object (67), namely in different movement phases of an organ of the examination object and / or with the same recording geometry.

12. Provisioning facility (71), which is designed to carry out the computer-implemented method according to one of the preceding claims.

13. Medical imaging facility (66), in particular computed tomography facility, comprising a provisioning facilty (71) according to claim 12.

14. Computer program for a data processing facility (87) with program instructions, which, upon execution on a data processing facility (87), carry out a computer-implemented method according to one of claims 1 to 11.

15. Computer-readable data carrier, comprising a computer program according to claim 14.