Method for validating processing data, method for providing a model trained by machine learning, processing device, computer program and data carrier

Abstract

Computer-implemented method for validating processing data (1), comprising the steps: - Receiving source data (2) based on medical image data acquisition, - receiving or determining the processing data (1) based on an application of a main processing algorithm (10) to the original data (1), - comparing comparison data (11), which are the processing data (1) or are based on the processing data (1), with reference data (12), which are the original data (1) or are based on an application of a reference processing algorithm (13) to the original data (1), by a comparison algorithm (14) in order to determine a comparison result (15), - evaluating a trigger condition (16, 17), the fulfillment of which depends on the comparison result (15), and, - if the trigger condition (16, 17) is met • issuing a notice (18) to a user, and / or • Modification of a predetermined acquisition parameter (19) for subsequent image data acquisition, and / or • Providing new processing data (20), wherein either the reference data (12) are used as new processing data (20) or wherein the new processing data (20) are determined by applying an alternative processing algorithm (21) to the original data (2).

Description

[0001] The invention relates to a method for validating processing data, a method for providing a model trained by machine learning, a processing device, a computer program and a data carrier.

[0002] In medical image data acquisition, acquired raw image data is typically processed by a processing algorithm before evaluation, which can be carried out, for example, by medical professionals or in an automated or semi-automated manner, for example to reconstruct three-dimensional or four-dimensional image data sets from individual projection images and / or to achieve noise reduction, contrast enhancement, background suppression, highlighting of edges and / or vascular structures, suppression of artifacts, for example metal artifacts, or similar.

[0003] Such processing typically makes it significantly easier to identify relevant structures and features compared to the original data. At the same time, however, such processing can potentially lead to distortion of the resulting processed data. For example, strong filtering or similar techniques can remove not only noise but also structures actually depicted in the raw image data, such as fine vessels. On the other hand, features that are not present in the raw image data can also be artificially generated in the processed data, for example, through more complex reconstruction algorithms or an algorithm for highlighting edges or vessel-like structures.These may resemble anatomical features or lesions, medical devices or similar and may therefore lead to an incorrect assessment by medical professionals or an evaluation algorithm.

[0004] Users of medical image data must therefore be aware of this issue and, based on their experience, check the plausibility of the processing results. Alternatively, limiting the processing algorithms or adjusting the algorithm parameters can be done to minimize the risk of corruption. However, this typically reduces the achievable image quality, as, for example, only very weak noise reduction can be achieved.

[0005] The invention is therefore based on the object of further improving the processing of medical image data, in particular avoiding or at least reducing the above-mentioned disadvantages of previous processing.

[0006] The task is solved by a computer-implemented method for validating processing data, which includes the following steps: - Receiving source data based on medical image data acquisition, - receiving or determining the processing data based on an application of a main processing algorithm to the source data, - comparing comparison data, which are the processing data or are based on the processing data, with reference data, which are the source data or are based on an application of a reference processing algorithm to the source data, by a comparison algorithm in order to determine a comparison result, - Evaluating a trigger condition whose fulfillment depends on the comparison result, and, - if the trigger condition is met, • Issue a notice to a user, and / or • Modification of a given acquisition parameter for subsequent image data acquisition, and / or • Providing new processing data, either using the reference data as new processing data or determining the new processing data by applying an alternative processing algorithm to the original data.

[0007] The comparison algorithm, or the trigger condition dependent on its comparison result, allows deviations that may arise during image processing to be automatically detected with good accuracy. For example, the suppression of actually present features and / or the creation of artificial features by the main processing algorithm can be detected.

[0008] Fulfillment of the trigger condition may indicate an impermissible deviation of the comparison data from the reference data. In this case, the processed data can be considered validated processed data after the trigger condition is not fulfilled. If the trigger condition is fulfilled, the user, such as medical personnel, can be informed that the processed or comparison data appears to be corrupted. In this case, the user can, for example, evaluate the raw data to determine whether falsification actually exists and / or select a different main processing algorithm or parameterize a base algorithm implementing the main processing algorithm differently to avoid such falsification.

[0009] To further facilitate the user's work, it may be useful to immediately provide or further process new processing data, either additionally or alternatively, upon detection of an impermissible deviation or upon fulfillment of the trigger condition. The reference processing algorithm or the alternative processing algorithm can, for example, perform processing that modifies the original data less than the main processing algorithm, such as less filtering and / or background suppression.

[0010] Adjusting the parameterization of medical image data acquisition can, in particular, serve to minimize patient exposure to the measurement. In X-ray imaging in particular, the ALARA principle (ALARA: "As Low As Reasonably Achievable") should be adhered to, meaning that the lowest X-ray dose still reasonably possible for an imaging task should be applied to a patient. However, dynamic parameterization of image data acquisition can also be useful for other imaging modalities, such as magnetic resonance imaging, for example, to shorten the required imaging time as much as possible.

[0011] Changing the acquisition parameter in a first direction can lead to higher image quality, particularly a better signal-to-noise ratio, but at the same time can also increase the X-ray dose and / or the measurement time, or generally the patient exposure. Changing the acquisition parameter in the opposite direction, on the other hand, can reduce patient exposure, but may result in lower image quality. The goal of dynamic parameter adjustment is typically to maintain image quality at a required level while simultaneously minimizing patient exposure.

[0012] In this context, a detected excessive deviation of the comparison data from the reference data can indicate that the quality of the source data is inadequate. If, for example, the main processing algorithm performs only very weak noise reduction, such as smoothing with a narrow filter kernel, and even such a minor modification results in anatomical features, medical devices, or similar elements apparently depicted in the source data no longer being present in the processed data, this can indicate, for example, that the achieved signal-to-noise ratio does not allow for a robust evaluation of the source data or that a robust evaluation is not possible due to other interference, such as streak artifacts that can occur during image reconstruction if the X-ray dose is too low.In such cases, it may be expedient to modify a predefined acquisition parameter for the subsequent medical imaging, for example, an applied X-ray dose, an X-ray tube voltage, and / or a measurement time. Such an embodiment of the method can thus also be considered a method for controlling or regulating medical imaging, for example, for dose control in X-ray imaging. The subsequent image data acquisition can, for example, be subsequent image data acquisition during fluoroscopy or, generally, further imaging of the same patient using the same medical imaging device.

[0013] However, a modification of the specified acquisition parameter may also be appropriate, or the trigger condition or another trigger condition may be met alternatively or additionally if a measure of the deviation of the comparison data from the reference data, which is specified, for example, by the comparison result, falls below a limit or the comparison data and the reference data are too similar. For example, if noise reduction performed as the main processing algorithm results in only an insignificant change to the original data, this may indicate that sufficiently good imaging would already be possible with an acquisition parameter that deviates from the specified acquisition parameter, which would, for example, reduce the patient's workload and / or shorten the measurement time.For example, in this case, the X-ray dose or a tube voltage can be reduced, or a tube parameter of an X-ray tube can be modified. In magnetic resonance imaging, for example, sequence parameters can be adjusted to achieve a less dense sampling of k-space and thus shorten the acquisition time.

[0014] Compared to known control approaches, for example, based on an approximately determined noise component in the image or a specific image section, the described options for modifying a given acquisition parameter can utilize implicit prior knowledge about the object to be imaged by using a suitable comparison algorithm, which, as will be explained in more detail later, can particularly detect when relevant features, such as anatomical features and / or images of medical devices, are distorted, without requiring an exact model of the imaged area of ​​the patient. This further improves the robustness and accuracy of the dynamic parameterization.

[0015] Instead of an automated adjustment of the specified recording parameter or in addition to this, it is of course possible to give the user a corresponding hint, whereupon he or she can, for example, adjust the corresponding parameter manually.

[0016] If several of the purposes explained above are to be fulfilled with the method described, it may be expedient to use separate process instances for the different purposes, which can be executed, for example, in parallel or sequentially based on the same source data. The process instances can differ from each other, for example, with regard to the main processing algorithms and / or trigger conditions used.

[0017] For example, a relatively mild denoising can be used as the main processing algorithm to detect an X-ray dose that is too low, while a stronger denoising should be used as the main processing algorithm to provide further processing data. This can then be checked to determine whether this could corrupt the original data. The comparison result obtained in this way can then optionally be used in an additional trigger condition to detect the presence of excessively small differences. This can be used to detect unnecessarily high image quality and, in this case, to reduce the X-ray dose during subsequent image acquisition, or similar.

[0018] Preferably, the same comparison algorithm is used in the different process instances, although in principle it would also be possible to use different comparison algorithms.

[0019] In this case, image data acquisition also includes the acquisition of raw data from which an image data set is reconstructed through processing, for example the acquisition of magnetic resonance signals.

[0020] The source data and optionally the processing data can be received, for example, from an external device, an internal data storage device, or a software module. The determination of the source data or the processing data can be performed outside the method according to the invention or, alternatively, as a step of this method.

[0021] The main processing algorithm can be used to enhance the image and / or modify an image impression. Additionally or alternatively, the source data may comprise multiple two-dimensional image datasets, with the main processing algorithm being or comprising a reconstruction of a three-dimensional or four-dimensional image dataset from these two-dimensional image datasets.

[0022] A main processing algorithm used for image enhancement can, for example, be used for noise reduction, contrast enhancement, background suppression, and / or highlighting edges and / or vascular structures and / or artifact reduction, such as suppressing metal artifacts. Additionally or alternatively, the main processing algorithm can implement the creation of a specific image impression desired by the user. For example, different users may prefer different degrees of background suppression and / or noise reduction when evaluating medical image data.

[0023] During reconstruction, the main processing algorithm can reconstruct a computed tomography image from multiple projection images, which can be based, in particular, on X-ray imaging. For example, a cone beam computed tomography image can be reconstructed. The reconstruction can be performed, for example, as a filtered backprojection, as a flow-limited time-resolved filtered backprojection, or as an iterative reconstruction.

[0024] A four-dimensional data set can be understood in particular as a time-resolved three-dimensional data set, such as can be determined, for example, in the context of a time-resolved three-dimensional angiography.

[0025] In particular, the reference processing algorithm and / or the alternative processing algorithm can also be used for image enhancement and / or for the reconstruction of a three-dimensional or four-dimensional image data set.

[0026] The main processing algorithm and, on the one hand, the reference processing algorithm and / or, on the other hand, the alternative processing algorithm can be implemented by different parameterization of a basic processing algorithm.

[0027] For example, these algorithms may differ from one another with regard to a filter parameter used, such as the cutoff frequency of a low-pass filter used for noise suppression, or with regard to a filter kernel used. For example, this may result in different levels of filtering for the different algorithms. While stronger filtering typically leads to better noise suppression, it can also potentially result in features actually present in the original data being removed or at least severely suppressed in the processed data. It may therefore be advantageous to reduce the filtering strength in the reference processing algorithm and / or the alternative processing algorithm compared to the main processing algorithm, for example, by using a higher cutoff frequency of a low-pass filter or a narrower filter kernel.

[0028] The choice of filter parameters can also determine, either additionally or alternatively, whether or to what extent anisotopic filtering should be applied. While anisotropic filtering can potentially achieve a sharper representation of edges or vascular structures with similarly good noise suppression, it can be advantageous to use isotropic filtering or a lower anisotropy in the reference processing algorithm or the alternative processing algorithm than in the main processing algorithm.

[0029] Alternatively and additionally, a degree of edge sharpening and / or background suppression can be adjusted, for example, by parameterizing the basic processing algorithm differently. Strong edge sharpening and / or background suppression can facilitate the analysis of the processed data for a user or a downstream evaluation algorithm or lead to a preferred image impression. However, since such processes can potentially lead to the suppression of relevant features and / or the formation of artifacts that a user or evaluation algorithm could potentially identify as a medical device, anatomical feature, or similar, it can be advantageous to reduce the degree of edge sharpening or background suppression in the reference processing algorithm or the alternative processing algorithm compared to the main processing algorithm.

[0030] If the main processing algorithm is or includes a reconstruction of a three-dimensional or four-dimensional image data set from two-dimensional image data sets, for example, a filter kernel used in a filtered backprojection can be selected or adapted as part of the parameterization of the basic processing algorithm or, for example, parameters of an iterative reconstruction can be adapted.

[0031] Thus, for example, reconstructions with different parameterizations can be carried out within the main processing algorithm and the reference processing algorithm, after which the different reconstruction results can be directly compared in three-dimensional or four-dimensional space.

[0032] However, it is also possible, for example, to first perform a forward projection based on these different reconstructions and then use a comparison of forward projections by the comparison algorithm.

[0033] The parameters mentioned and other parameters with respect to which the algorithms mentioned may differ from one another may in particular be parameters that can be set by the user during image processing or that can be modified by the user by selecting different sets of parameters in order to produce a desired image impression.

[0034] However, it is also possible for the reference processing algorithm and / or the alternative processing algorithm to be based on a different processing approach than the main processing algorithm. For example, at least one of the aforementioned algorithms can be implemented by a model trained by machine learning, and at least one other of the aforementioned algorithms can be manually implemented or parameterized. Models trained by machine learning can often achieve particularly good image quality or a particularly favorable image impression. In individual cases, however, the application of such a trained model can result in significant distortions of the resulting processed data.In the method according to the invention, however, these cases can be robustly detected by comparing them with the original data and / or by using a reference processing algorithm, which is particularly manually implemented. Thus, in the event of such corruption, an indication can be given or a fallback to the new processed data can be made. Even in the case of reconstruction, for example, one of the algorithms can be based on iterative reconstruction and at least one other algorithm can be based on filtered backprojection.

[0035] As already explained above, the source data may comprise multiple two-dimensional image data sets, with the processing data being formed by a three-dimensional or four-dimensional image data set that is reconstructed from these two-dimensional image data sets by the main processing algorithm. In this case, the comparison data may be or comprise two-dimensional image data sets determined by a respective forward projection of this three-dimensional or four-dimensional image data set.

[0036] In cases where the main processing algorithm is used to reconstruct a three- or four-dimensional image data set, it may be useful, as an alternative or in addition to the comparison of different parameterized reconstructions explained above, to check, within the framework of the comparison algorithm, the consistency of the reconstructed image data set with the original two-dimensional image data sets or, in general, with two-dimensional reference data that may result, for example, from filtering or other preprocessing of the two-dimensional image data sets.For example, in the context of a three-dimensional angiography acquired by computed tomography, particularly cone beam tomography, it can be examined whether and to what extent the spread of the contrast agent in the vascular system in the reconstructed four-dimensional image dataset is consistent with the depiction of this spread in the individual acquired projection images. For this purpose, the three-dimensional image dataset corresponding to the time of acquisition of the respective projection image can be selected or interpolated from the four-dimensional image dataset and projected forward according to the acquisition geometry of the respective projection image to provide the comparison data.

[0037] The comparison algorithm can be implemented by applying an evaluation algorithm, on the one hand, to the comparison data to identify segments of the comparison data that each depict a relevant anatomical feature and / or a medical device, and, on the other hand, applying an evaluation algorithm to the reference data to identify segments of the reference data that each depict a relevant anatomical feature and / or medical device. The comparison result can then depend on a comparison of the number and / or positions and / or dimensions of the segments in the comparison data with the number and / or positions and / or dimensions of the segments in the reference data.Alternatively or additionally, a bounding box can be determined in both the reference data and the comparison data for the respective segment, whereby the comparison result depends on a comparison of the positions and / or dimensions of the bounding boxes in the comparison data with the positions and / or dimensions of the bounding boxes in the reference data.

[0038] In other words, it can be checked whether the comparison data show the same relevant areas as the reference data. In the simplest case, only the number of detected segments in the reference data and in the comparison data is compared. However, in order to better detect, for example, the truncation of a vessel that actually extends further, or similar changes that could potentially result from the application of the main processing algorithm, it may be useful, in addition or alternatively, to also take into account the sizes or positions of the various segments. The use of bounding boxes, in particular rectangular or cuboid-shaped ones, can make the comparison of positions or dimensions easier or enable a more robust comparison compared to a comparison of irregularly shaped segments.

[0039] The comparison was carried out in particular between data registered with each other, so that the positions and dimensions of the segments or of the bounding boxes encompassing the respective segment, i.e. in particular of bounding rectangles or bounding boxes, can be directly compared.

[0040] The anatomical features and / or medical devices to be considered are, in particular, predefined. Thus, it is possible to predetermine which segments are to be considered when evaluating the trigger condition. Numerous algorithms for segmenting medical image data or specific anatomical features and / or medical devices in medical image data are known in the field of medical image data processing, which can be used in the method according to the invention as an evaluation algorithm or as part of the evaluation algorithm.

[0041] In an advantageous embodiment, the comparison algorithm is or includes a model trained by machine learning. For example, the model trained by machine learning can implement the evaluation algorithm explained above. Machine learning can also be used to robustly solve complex segmentation or classification tasks in medical image data.

[0042] For example, the model trained by machine learning can be based on supervised learning with error feedback, where the training data sets contain not only the data to be segmented but also a target result. For example, the target result can be or include a manual segmentation or classification of the data to be segmented by at least one expert.

[0043] However, as will be explained in more detail later, a model trained by machine learning can also be trained to immediately recognize whether a relevant change has occurred, or several models can be trained to check the processing data for a specific type of change.

[0044] In general, a model trained by machine learning mimics cognitive functions that humans associate with other people's minds. In particular, training based on training data enables the trained model to adapt to new circumstances and recognize and extrapolate patterns. A "machine learning trained model" can also be referred to as a "trained function."

[0045] In general, model parameters can be adjusted through training. In particular, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used. Representation learning (alternatively called "feature learning") can also be used. In particular, the parameters of the trained model can be adjusted iteratively through multiple training steps. In particular, a given cost function can be minimized during training. Error feedback can be used, particularly when training a neural network.

[0046] A model trained by machine learning can, for example, be based on a neural network, a support vector machine (SVM), a decision tree, and / or a Bayesian network, and / or on k-means clustering, Q-learning, genetic algorithms, and / or matching rules. In particular, a neural network can be a deep neural network, a convolutional neural network (CNN), or a deep CNN. Furthermore, the neural network can be an adversarial network, a deep adversarial network, and / or a generative adversarial network (GAN).

[0047] Preferably, the training of the model trained by machine learning takes place outside of the claimed computer-implemented method. For example, the computer-implemented method can be carried out as part of the application of medical imaging and / or the evaluation of image data from such imaging, for example, in a clinic or by a physician's office. The training of the model can be carried out spatially and temporally independently of this and by other persons, for example, by a manufacturer of an imaging device used as part of its production or development.

[0048] The comparison algorithm can comprise at least a first and a second model trained by machine learning, which each process the comparison data and the reference data as input data. The first model trained by machine learning can determine a first intermediate result relating to the presence and / or extent of a deviation of a first deviation type between the comparison data and the reference data. The second model trained by machine learning can determine a second intermediate result relating to the presence and / or extent of a deviation of a second deviation type, different from the first deviation type, between the comparison data and the reference data. The comparison result can then depend on the first and second intermediate results.

[0049] More than two deviation types can also be detected by separate models trained to detect the respective deviation type. Thus, the comparison algorithm can additionally comprise at least one further model trained by machine learning, which determines a respective further intermediate result that describes whether a deviation of the respective further deviation type is detected between the comparison data and the reference data, whereby the fulfillment of the trigger condition additionally depends on the respective further intermediate result.

[0050] For example, the trigger condition can already be met if one of the trained models used detects a deviation between the first and second input data.

[0051] Detecting different deviation types using separately trained models can enable the use of a simpler underlying algorithm for each model. For example, if the model is based on a neural network, it can have fewer nodes, layers, or links than would be required for a model that can detect all relevant deviation types on its own. Training separate models to detect different deviation types can also enable faster training with a smaller number of required training data sets. Alternatively, however, it would also be entirely possible to use a single model trained by machine learning to detect all relevant deviation types.

[0052] As explained above, the training of the models trained by machine learning can preferably take place outside of the claimed computer-implemented method. The first and / or the second model trained by machine learning can be trained, for example, using supervised learning. For this purpose, training data sets can be used, each comprising pairs of original or reference data and comparison data. In addition, the presence and / or extent of the deviation of the respective deviation type can be stored in the training data set as target information for at least one deviation type. The pairs can originate from previous medical imaging processes and / or from a simulation, and the target information can be specified, for example, based on a manual assessment of the respective pair by medical professionals.Based on these training data sets, supervised learning can be carried out in a conventional manner, for example by minimizing a cost function through error feedback, in particular by a gradient descent method.

[0053] A discriminator can be used as the model trained by machine learning or as at least one of the models trained by machine learning, the parameterization of which is based on the training of a generative adversarial network comprising the discriminator. A generative adversarial network (GAN) comprises a generator and a discriminator, with the generator generating synthetic data and the discriminator distinguishing between synthetic and real data. By training the generator and / or the discriminator, the generator is configured to generate synthetic data that is incorrectly classified as real by the discriminator. The discriminator is configured to distinguish between real data and synthetic data generated by the generator.

[0054] As explained above, the training of the model trained by machine learning can preferably take place outside of the claimed computer-implemented method. The GAN generator used during training can be parameterized during training such that it generates features with which the discriminator, and thus the model trained by machine learning used in the comparison algorithm, is to detect differences, as realistically as possible. In the sense of game theory, a GAN can be interpreted as a zero-sum game. The training of the generator and / or the discriminator can, in particular, be based on the minimization of a cost function. The minimization of the cost function can be achieved using a gradient descent method, in particular through error feedback. The generator and the discriminator can, for example, each be implemented as a convolutional neural network (CNN).

[0055] For example, during GAN training, the generator can be trained to generate a synthetic image parameterized by input data, which, for example, should represent at least one anatomical feature and / or at least one medical device, the recognizable representation of which in the processing data is to be checked by the comparison algorithm. By jointly training the discriminator and the generator, the discriminator is thus trained to robustly detect both the absence and incorrect representation of these features and is thus well suited to detecting the suppression or falsification of corresponding features by the main processing algorithm.

[0056] Alternatively, the main processing algorithm can also be used directly as a generator during such training. The main processing algorithm can be predefined or trained during the GAN training process, for example, to train a filter algorithm using machine learning.

[0057] The comparison result may describe the presence and / or extent of a deviation of at least one of the following deviation types: - an absence, interruption and / or a shortened representation of at least one vessel, and / or - the absence of an image of a medical device, - and / or the absence of an anatomical feature in the comparison data compared to the reference data and / or in the reference data compared to the comparison data.

[0058] The absence of a vessel, medical device, or an anatomical feature in the comparison data compared to the reference data indicates suppression of the respective representation, for example, due to excessive filtering. Excessive filtering or other changes to the representation, such as overly aggressive suppression of an image background, can also lead to an apparent shortening or interruption of an imaged vessel due to the application of the main processing algorithm.

[0059] The opposite case, in which features appear to be missing in the reference data or vessels appear to be shortened or interrupted compared to the comparison data, indicates that the application of the main processing algorithm has added a feature that was not actually present in the reference or original data.

[0060] The deviations described can be detected, for example, through the segmentation already explained above. In addition or alternatively, corresponding features, and thus also the aforementioned differences or changes, can also be detected by a suitable model trained through machine learning. For example, in training data sets that each contain pairs of reference data and comparison data, medical professionals can manually annotate whether deviations exist, or which of the aforementioned deviations exist. Supervised learning can then be performed based on several of these training data sets in a known manner.

[0061] Since annotating training data sets can be relatively complex, a GAN can also be used during training, as already explained above, to train a discriminator capable of detecting these deviations. Through the interaction of generator and discriminator in the GAN, training success can be achieved with significantly less training data, or, given a given amount of training data, training success can be further improved. Furthermore, annotation of training data can be omitted, and unsupervised learning can be performed.

[0062] In an advantageous embodiment of the method, additional previous source data can be received which are based on a previous medical image data acquisition which took place before the medical image data acquisition, wherein the or an acquisition parameter for the or a subsequent image data acquisition is specified as a function of a further comparison result which is determined by the comparison algorithm by comparing current input data which correspond to or are based on the original data with previous input data which correspond to or are based on the previous original data.

[0063] Using the procedure described, changes in image quality in successive image data acquisitions, for example during fluoroscopy or in consecutively acquired projection images as part of a computed tomography scan, can be detected and taken into account when specifying the acquisition parameter, such as an X-ray dose, an X-ray tube parameter, or a measurement time. For example, as part of dose control, the X-ray dose can be reduced in successive image data acquisitions until it is recognized from the further comparison result that at least one relevant feature is no longer being imaged, or something similar. Adjusting the acquisition parameter can also be useful, for example, to at least largely compensate for a change in X-ray attenuation caused by the patient when the acquisition geometry changes between the various image data acquisitions.

[0064] In this case, the recording parameter can, as explained above, additionally depend on the comparison result or can depend exclusively on the further comparison result, for example if the evaluation of the trigger condition explained above serves exclusively to provide information and / or to provide new processing data.

[0065] The comparison information can, at least if the trigger condition is met, relate to at least one segment of the comparison data in which the comparison data deviates from the reference data, wherein the indication comprises segment information relating to this segment. For example, the comparison data or the processing data, on the one hand, and the reference data or the original data, on the other hand, can be displayed to a user as an indication, wherein the at least one segment for which the deviation was detected is highlighted in order to clarify to the user the type and severity of the deviation or falsification.

[0066] Alternatively or additionally, the hint may include a suggestion for selecting a suitable main processing algorithm or a suggested value for at least one parameter of the or a basic algorithm underlying the main processing algorithm. This may allow the user to continue to select the main processing algorithm or the parameter independently, but receive a hint for a likely particularly advantageous choice.

[0067] The invention also relates to a computer-implemented method for providing a model trained by machine learning for use as a comparison algorithm or as a sub-algorithm of the comparison algorithm in the computer-implemented method according to the invention for validating processing data, wherein the computer-implemented method for providing a model trained by machine learning comprises the following steps: - receiving input training data comprising a plurality of training data sets, each of which in turn comprises training data based on a medical image data acquisition and / or a simulation of a medical image data acquisition, - Training a model based on the input training data to determine the model trained by machine learning, - Deploying the model trained by machine learning.

[0068] Various options for conducting such training have already been explained above and will therefore not be repeated. The training data can be used, in particular, as source or reference data during the training. If processing or comparison data is required during the training, this can be determined based on the respective training data or provided directly as part of the respective training dataset.

[0069] In general, unsupervised learning can be used, for example, by using a GAN detector as a trained model, or supervised learning can be used, for example, to learn from deviations previously classified manually or by other methods. The latter can be used, in particular, to separately train the first and second trained models explained above. The design options already explained above can be transferred, with the aforementioned advantages, to the computer-implemented method for providing a model trained by machine learning.

[0070] Furthermore, the invention relates to a processing device configured to carry out the computer-implemented method according to the invention for validating processing data and / or the computer-implemented method according to the invention for providing a model trained by machine learning. The processing device can be configured, for example, as suitably programmed data processing devices, or alternatively, the aforementioned functionality can be implemented at least partially in a hard-wired manner. The processing device can be integrated into a medical imaging device, in particular an X-ray device or a magnetic resonance imaging scanner, or configured separately from it. It can be implemented, for example, as a workstation computer, server, or cloud solution.

[0071] The invention also relates to a computer program with instructions which, when executed on a data processing device, are configured to carry out the computer-implemented method according to the invention for validating processing data and / or the computer-implemented method according to the invention for providing a model trained by machine learning.

[0072] The invention also relates to a data carrier comprising the computer program according to the invention.

[0073] Further advantages and details of the invention will become apparent from the following exemplary embodiments and the accompanying drawings. These schematically show: Fig. 1 a flowchart of an embodiment of the computer-implemented method according to the invention for validating processing data, Fig. 2 an embodiment of a processing device according to the invention, Fig. 3-5 partial flow diagrams of further embodiments of the computer-implemented method according to the invention for validating processing data, Fig. 6 is a flowchart of an embodiment of the computer-implemented method according to the invention for providing a model trained by machine learning, and Fig. 7&8 possible structures of the model trained by machine learning.

[0074] As already explained at the beginning, when processing original data 2 from a medical image, for example during filtering, reconstruction, contrast enhancement, or similar, individual features can be suppressed or even artificially created. To prevent this from leading, for example, to incorrect evaluation of the image data, it is advisable to validate the processing results.

[0075] Fig. 1 shows a flowchart of a computer-implemented method for performing such a validation of processing data 1. Since the Fig. 1 is relatively complex, a brief overview of the central parts of the process is given for better understanding before the individual steps of the concrete example are explained in detail.

[0076] In steps S1 and S2, firstly, original data 2, which is based on a medical image data acquisition, and processing data 1, which is based on an application of a main processing algorithm 10 to the original data 1, are obtained.

[0077] In step S4, comparison data 11, which in the example corresponds to the processing data 1, are then compared with reference data 12, which in the example is based on an application of a reference processing algorithm 13 to the original data 1. In modifications of the exemplary embodiment shown, it would also be possible for the processing data 1 to be further processed before the comparison and / or for the original data 1 to be used directly as reference data 12. An example of this alternative embodiment will be described later with reference to Fig. 3 will be explained.

[0078] The comparison result is evaluated in steps S5 and S8 using a respective trigger condition 16, 17. Depending on whether one or which of the trigger conditions 16, 17 is met, various actions can be performed in steps S6, S7, S9, and S10. In the example, in step S6, a hint 18 is given to a user, in step S7, new processing data 20 is provided by applying an alternative processing algorithm 21 to the original data 2, in step S9, a predetermined acquisition parameter 19 is modified for subsequent image data acquisition, and in step S10, the validated processing data 1 is provided. In a modification of the determining procedure, it would also be possible for the already known reference data 12 to be provided directly as new processing data 20 in step S7.

[0079] The following are the individual steps of the Fig. 1 is explained in more detail. The following explanation assumes, for example, that the source data 2 is a single two-dimensional image data set, for example, an X-ray image, whereby it is also assumed purely for example that the processing data 1 to be validated is based on a filtering of this source data 2. Alternatively or in addition to filtering, the main processing algorithm 10 could also perform other processing steps to enhance the image and / or to change an image impression. Examples of this have already been discussed in the general part of the description.

[0080] The procedure is explained with additional reference to Fig. 2, which illustrates an application situation for the explained method and an exemplary implementation of a processing device 5 implementing the method. In the example, the processing device 5 is integrated into a medical imaging device 3, for example, an X-ray device, in particular a computer tomography scanner.

[0081] In the example, the processing device 5 is implemented by a processor 7 of a freely programmable data processing device 51 executing a computer program 9 stored in the memory 8, whose instructions implement the method steps. Various other implementation options for suitable processing devices 5 have already been discussed in the general part of the description.

[0082] In step S1, source data 2, which in the example are two-dimensional X-ray images of patient 4, are first acquired by the medical imaging device 3 and received by the processing device 5. In alternative embodiments, it would also be possible to read out source data previously acquired outside the method, for example, from a database.

[0083] In step S2, the original data 2 is processed by the main processing algorithm 10 to obtain the processing data 1. Alternatively, it would also be possible to perform the processing, for example, in a separate device, for example in the workstation computer 6, and thus to receive the processing data 1 from the separate device for validation by the processing device 5.

[0084] In the example, the main processing algorithm 10, just like the reference processing algorithm 13 and the alternative processing algorithm 21, which will be explained later, is implemented by specifying at least one parameter 25 of a basic processing algorithm 24, wherein this parameter has different values ​​for the different algorithms mentioned. Parameter 25 can, for example, specify a cutoff frequency of a filter or the size and / or shape of a filter kernel for filtering.

[0085] In step S3, the reference processing algorithm 13 is applied to the original data 2 to obtain reference data 12. As explained above, the reference processing algorithm 13 in the example corresponds to the main processing algorithm 10, except for the choice of parameter 25. The parameter 25 of the reference processing algorithm 13 is chosen such that falsification of relevant features in the reference data 12 can be excluded. For example, by appropriately choosing the parameter 25, only very weak filtering can be performed, so that relevant features that could potentially be suppressed by the stronger filtering of the main processing algorithm 10 are robustly retained. In alternative embodiments, it would also be possible to use the original data 2 directly as reference data 12.

[0086] In step S4, comparison data 11, which in the example correspond to the processing data 1, but in alternative embodiments could also result from further processing, are compared with the reference data 12 by the comparison algorithm 14 in order to determine a comparison result 15.

[0087] The comparison algorithm 14 is implemented in the example by a model 33 trained by machine learning. As will be explained in more detail later with reference to the Fig. 6 and Fig. 7, the model 33 trained by machine learning in the example is a discriminator 40 whose parameterization is based on a training of a generative adversarial network 41 comprising the discriminator 40.

[0088] In the simplest case, the comparison information 15 can indicate a degree of deviation, for example standardized between 0 and 1, or a probability that the comparison data 11 deviate from the reference data 12 in an inadmissible manner, for example to such an extent that essential features are suppressed or falsified.

[0089] In step S5, a trigger condition 16 is then evaluated, which in the example is fulfilled when the comparison information 15 exceeds a limit value and it can therefore be assumed that the main processing algorithm 10 has falsified the original data 2 in an inadmissible manner.

[0090] In this case, on the one hand, in step S6, for example via the workstation computer 6, a hint 18 is issued to a user. As already explained in the general part, the hint 18 can comprise segment information 46, which can, for example, describe in which segment of the processing data 1 the corresponding deviation occurs. In the example shown, for example, the model 33 trained by machine learning could be designed and trained to additionally provide the segment information 46. Corresponding segment information 46 can be provided particularly simply and robustly if segmentation takes place within the scope of the comparison algorithm 14, as will be explained later with reference to Fig. 4 will be explained.

[0091] In addition, in step S7, new processing data 20 are immediately provided by processing the original data 2 by the alternative processing algorithm 21. By appropriately selecting the parameter 25, the alternative processing algorithm 21 can, for example, perform filtering with a strength that lies between the filter strength of the reference processing algorithm 13 and the main processing algorithm 10, so that it can initially be assumed with a high degree of certainty that the corruption of the processing data 1 that led to the fulfillment of the trigger condition 16 in step S5 is not present in the new processing data 20, or at least is significantly less present. Optionally, it would be possible, for example, to perform the validation explained above again for the new processing data 20.If this validation also fails, for example, the original data 2 could be immediately provided as new processing data 20 or the parameter 25 could be modified again.

[0092] If trigger condition 16 is not met, a check is performed in step S8 to determine whether further trigger condition 17 is met. In the example, this further trigger condition 17 is to be met if comparison result 15 indicates an extremely small deviation between comparison data 11 and reference data 12, for example, if comparison result 15 falls below a predefined threshold. This indicates that main processing algorithm 10, or the denoising implemented thereby in the example, has only caused a negligibly small change to the original data, which in turn indicates that the image quality of original data 2 is unnecessarily high for the given imaging task.

[0093] Thus, if the further triggering condition 17 is fulfilled in step S9, for example, a reduction of an X-ray dose and / or a resting voltage or generally a modification of at least one predetermined acquisition parameter 19 for a subsequent image data acquisition can take place, since in this case a somewhat lower signal-to-noise ratio can be permitted and thus a burden for the patient 4 can be reduced.

[0094] Irrespective of whether the further trigger condition 17 is fulfilled, in step S 10, i.e. always when the trigger condition 16 is not fulfilled, the now validated processing data 2 are provided in order to display them and / or continue processing them, for example, on the workstation computer 6.

[0095] As already explained with reference to step S9, the comparison of different data by the comparison function 14 can also be useful for dose control of an X-ray dose or generally for the parameterization of subsequent image data acquisition. Therefore, in the Fig. In the example shown in Figure 1, steps S11 to S13 for adapting the detection parameter 19 are carried out in parallel to the steps explained above.

[0096] For this purpose, in step S11, previous original data 42 is first received, which is based on a previous medical image data acquisition that occurred prior to the medical image data acquisition. For example, during a fluoroscopy, the previously acquired image data can be read from memory 8.

[0097] In step S12, the comparison algorithm 14 is then used again to compare current input data 44, which corresponds to or is based on the original data 2, with previous input data 45, which corresponds to or is based on the previous original data 42. The resulting further comparison result 43 thus indicates whether, for example due to a dose adjustment between the image data acquisitions and / or other influences, there is reason to fear that features are missing or corrupted in the current original data due to an excessively low X-ray dose and / or whether artifacts occur. In step S13, the acquisition parameter 19 is then adjusted depending on this further comparison result 43, for example, to increase a dose. This can also be done depending on the value of the acquisition parameter 19 potentially specified in step S9.

[0098] For the sake of completeness, it should be noted that the information relating to Fig. The procedure explained in Figure 1 can also be used in cases where the main processing algorithm 10 performs a reconstruction of three-dimensional or four-dimensional image data sets, for example, in the context of a computed tomography (CT) or a time-resolved CT angiography. In this case, the reference processing algorithm 13 can also perform a corresponding reconstruction, however, for example, a different filter kernel is used in the reconstruction than in the main processing algorithm 10, or instead of an iterative reconstruction in the main processing algorithm 10, a reconstruction by backprojection is used in the reference algorithm 13, or similar. In this case, the comparison algorithm 14 compares in the Fig. 1 directly displays the three-dimensional and four-dimensional image data sets.

[0099] A direct comparison of three-dimensional or four-dimensional image data sets using the comparison algorithm 14 is potentially very computationally intensive. Furthermore, potentially very large amounts of input data must be processed, so that, for example, an implementation of the comparison algorithm 14 as a neural network requires a very large number of input nodes and thus typically a very large number of layers and nodes per layer. This, in turn, leads to a very large number of free parameters, making training a corresponding algorithm complex and typically requiring a very large number of training data sets.

[0100] In Fig. 3 will be in Fig. The exemplary embodiment shown in Figure 1 is therefore slightly modified for validating processing data 1 formed by a three- or four-dimensional image data set 23, respectively, by replacing steps S2 and S3 with steps S2' and S3' explained below. Here, the original data 2, as schematically illustrated in step S1, comprises several two-dimensional image data sets 22.

[0101] In step S2', the three- or four-dimensional image data set 23 is first reconstructed from the two-dimensional image data sets 22. For example, when reconstructing a four-dimensional data set of a computed tomography-based angiography, often only an approximate reconstruction is carried out, which is why the validation of the processing data 1 is particularly relevant in this case.

[0102] In step S3', a forward projection 27 then takes place to provide two-dimensional image data sets 26 as comparison data 11. If the processing data 1 is a three-dimensional image data set 23, such a forward projection 27 can take place immediately. However, if it is a four-dimensional image data set 23, a three-dimensional image data set can first be determined for the respective point in time at which the respective two-dimensional image data set 22 of the original data 2 was acquired, i.e., for example, selected or interpolated from the four-dimensional image data set, and this three-dimensional image data set can be forward projected.

[0103] In step S4, the comparison data 11 determined in step S3' is compared with reference data 12, which in the example shown are formed directly from the original data 2. Alternatively, as already mentioned Fig. 1, a reference processing algorithm can be used, for example, to filter the image data sets 22. In particular, the individual image data sets 26 of the comparison data can be compared with the individual image data sets 22 of the reference data 12.

[0104] The procedure can then, as already Fig. 1 explained.

[0105] Fig. 4 shows an alternative embodiment of the comparison algorithm 14 in a step S4', which in Fig. 1 or Fig. 3 can replace step S4. Step S4' comprises several substeps.

[0106] In the first sub-step S4.1', a first model 34 trained by machine learning processes as input data 36, ​​37, on the one hand, the comparison data 11 and, on the other hand, the reference data 12, in order to determine a first intermediate result 38 which relates to the presence and / or extent of a deviation of a first deviation type between the comparison data 11 and the reference data 12.

[0107] In the second sub-step S4.2', the same input data 36, ​​37 are processed by a second model 35 trained by machine learning in order to determine a second intermediate result 38 relating to the presence and / or extent of a deviation of a second deviation type between the comparison data 11 and the reference data 12.

[0108] The first and second intermediate results 38, 39 can, for example, each indicate a probability, normalized between 0 and 1, that a deviation of a certain deviation type exists.

[0109] The intermediate results 38, 39 are then combined into the comparison result 15 in the third sub-step 4.3'. In the simplest case, the comparison result 15 can be a list of the various intermediate results 38 and 39, in which case the trigger condition 16 can, for example, comprise several subconditions that can, for example, implement a limit value comparison for each intermediate result 38, 39. The trigger condition 18 can then, for example, already be met if one of the subconditions is met. However, it can also be expedient to combine the various intermediate results 38, 39 into a common value, for example by using the largest of the intermediate results 38, 39 as the comparison result 15 or by using the sum or a product of the intermediate results 38 and 39 as the comparison result 15. In this case, the trigger condition 18 can, for example, be implemented by a simple limit value comparison.

[0110] A suitable approach for training the first and second machine learning trained models 34, 35 will be discussed later with reference to the Fig. 6 and Fig. 8 will be explained.

[0111] Fig. 4 shows a further alternative embodiment of the comparison algorithm 14 in a step S4'', which in Fig. 1 or Fig. 3 can replace step S4. Step S4'' comprises several substeps.

[0112] In sub-step S4.1'', an evaluation algorithm 28 is applied to the comparison data 11 to identify segments 29 of the comparison data 11, each of which depicts a relevant anatomical feature and / or a medical device. Since a variety of suitable segmentation or evaluation algorithms are known, the segmentation will not be described in detail. Purely as an example, a model trained by machine learning could be used to segment or classify the individual segments. However, known analytical algorithms can also be used, which may be based, for example, on threshold comparisons, region growing, or similar.

[0113] In sub-step S4.2'', the evaluation algorithm 28 is applied to the reference data 12 in order to recognize segments 30 of the reference data 12, each of which depicts a relevant anatomical feature and / or a medical device.

[0114] In step S4.3'', the number and / or positions and / or dimensions of the segments 29 in the comparison data 11 can then be compared with the number and / or positions and / or dimensions of the segments 30 in the reference data 12 to determine the comparison result 15. For example, it can be checked whether all of the segments 29 detected in the comparison data 11 were also detected in the reference data 12 and vice versa.

[0115] As already explained in the general part, in addition to or as an alternative to the comparison of the segments 29,30, these limiting bounding boxes 31,32 can also be compared in order to determine the comparison result 15.

[0116] Fig. 6 shows a flowchart of a method for providing a respective machine learning trained model 33, 34, 35, which, as already described with reference to Fig. 1 or 4, can be used as a comparison algorithm 14 or as a sub-algorithm of the comparison algorithm 14. Here, first with additional reference to Fig. 7 explains the training of model 33 as an example.

[0117] In step S14, input training data 47 is received or provided, which comprises a plurality of training data sets 48, which in turn each comprise training data 49 based on a medical image data acquisition and / or a simulation of a medical image data acquisition.

[0118] In step S15, a model is then trained on the basis of the input training data 47 to determine the model 33 trained by machine learning. In the example, a discriminator 40 is used as the model 33 trained by machine learning, the parameterization of which is based on the training of a generative adversarial network 41 comprising the discriminator 40. In this case, the training data sets should additionally include training parameters 61, which can be converted into synthetic output data G(x) by the generator G of the generative adversarial network 41. These can, for example, describe the image geometry for the respective training data 49 and patient parameters, such as the patient's gender, height, and weight.

[0119] Fig. shows an example data flow diagram of such a generative adversarial network 41. The generative adversarial network 41 is used to generate synthetic output data G(x) based on input data x, whereby the output data G(x) should be indistinguishable from real output data. In the example, the training parameters 61 are used as the input data x in training. The training data, for example, X-ray projection images, are used as the real output data y. The synthetic output data G(x) has the same structure as the real output data y, but its content is not derived from real data.

[0120] The Generative Adversarial Network 41 consists of a generator function G and a classification function C, which are trained together. The classification function forms the discriminator 40 and, after its training, can be used as the machine learning-trained model 33 in the method according to Fig. 1 can be used.

[0121] The task of the generator function G is to provide realistic synthetic output data G(x) based on input data x, and the task of the classification function C is to distinguish between real output data y and synthetic output data G(x). Specifically, the output of the classification function C is a real number between 0 and 1 corresponding to the probability that the input value is real data, so an ideal classification function would calculate an output value of C(y) ≈ 1 for real data y and C(G(x)) ≈ 0 for synthetic data G(x).

[0122] During the training process, the parameters of the generator function G are adjusted so that the synthetic output data has the same properties as the real output data, so that the classification function C can no longer distinguish between real and synthetic data. At the same time, the parameters of the classification function C are adjusted so that it best distinguishes between real and synthetic data. The training is based on pairs of input data x and the corresponding real output data y. In a single training step, the generator function G is applied to the input data x to generate synthetic output data G(x). Furthermore, the classification function C is applied to the real output data y to generate an initial classification result C(y).In addition, the classification function C is applied to the synthetic output data G(x) to generate a second classification result C(G(x)).

[0123] The adaptation of the parameters 50 of the generator G and the classification function C is based on the minimization of a cost function using the backpropagation algorithm. In this embodiment, the cost function K is C for the classification function C Kc∝−BCE(C(y),1)−BCE(C(G(x)),0), where BCE denotes the binary cross entropy, which is BCE(z,z')=z'−log(z)+(1−z')−log(1−z) By using this cost function, the cost function K to be minimized increases C both in the false classification of real output data as synthetic (characterized by C(y) ≈ 0) and in the false classification of synthetic output data as real (characterized by C(G(x)) ≈ 1).

[0124] The cost function K G for the generator function G is in the example Kc∝−BCE(C(x),1)=−log(C(C(x)). Using this cost function, correctly classified synthetic output data (given as C(G(x)) ≈ 0) leads to an increase in the cost function KG to be minimized.

[0125] The model 33 trained by machine learning, i.e. the discriminator 40 with the parameters 50 determined during training, can then be provided in a step S 16 and, for example, in the Fig. 1 shown procedures can be used.

[0126] As already mentioned with reference to Fig. 4, it may be appropriate to detect different types of changes using separate machine learning-trained models 34, 35. Suitable training will involve sequences with additional reference to Fig. 8. In the example, supervised training is to be carried out. Here, the training data 49 is used as source data during training. In addition, the training data sets 48 provided in step S14 each comprise processing or comparison data 52, which are each determined from the training data 49 of the respective training data set 48 by processing, for example, noise reduction, and a respective target result 53, wherein the target result 53 indicates whether or to what extent the respective processing or comparison data 52 deviates from the respective training data with regard to the type of change for which the model is to be trained. The target result 53 can, for example, be specified by a manual assessment of the respective training data set by medical personnel.

[0127] The model used in the example is a Fig. The convolutional neural network shown in Figure 5 is used. Layers L.1 to L.10 are present twice, namely once in the illustrated subnetwork 58 and again in subnetwork 59, with layer L.11 processing the output data of the respective layer L.10 of both subnetwork 58 and subnetwork 59 as input data. Subnetwork 58, or its layer L.1, processes the training data 49 or, after training, the reference data 12. Subnetwork 59, or its layer L.1, processes the processing or comparison data 52 or, after training, the comparison data 11.

[0128] In the example, the neural network consists of convolutional layers, pooling layers, and fully connected layers. The input layer L.1 contains a node for each pixel of the respective input data, with each pixel having a channel (the respective intensity value). The input layer is followed by four convolutional layers L.2, L.4, L.6, L.8, each of which is followed by a pooling layer L.3, L.5, L.7, L.9. Each of the convolutional layers uses a 5x5 kernel (indicated by "5x5 kernel") with a padding of 2 (indicated by "P: 2") and an increasing number of filters / convolution kernels (indicated by "F: 2", "F: 4", or "F: 8"). In addition, there are four pooling layers L.3, L.5, L.7, L.9, where the first three layers L.3, L.5, L.7 perform averaging over arrays of size 4x4 and the last pooling layer L.9 performs maximum selection over arrays of size 2x2. Fig.Figure 8 shows an additional layer L.10, which flattens the input images (i.e., the eight 4x4 images are combined into a 128-entry vector). However, this layer is not relevant for the actual computation.

[0129] The final layers of the network are three fully connected layers L.11, L.12, L.13, where the first fully connected layer has 256 input nodes and 60 output nodes, the second fully connected layer L.12 has 60 input nodes and 10 output nodes, and the third fully connected layer L.13 has 10 input nodes and one output node. The output node forms the output layer of the entire machine learning model. The value of the output layer node corresponds to the probability that the processing or comparison data 52 fed to the subnetwork 59 deviates from the training data 49 with respect to the respective trained change type.

[0130] A database containing 500 medical images was used to train the neural network. The processing or comparison data 52 was generated by a processing algorithm, as explained above. The desired result 53 was specified by radiologists as experts. In the example, the experts determined for each of the 500 data sets whether or to what extent the image of a vessel was correctly reproduced in the processing or comparison data 52, normalized to the interval 0 to 1 (ground truth). The database was divided into training data (320 data sets), validation data (80 data sets), and test data (100 data sets).

[0131] For training, the backpropagation algorithm was used based on a cost function that sums the squared deviation between the actual result 60 and the target result 61 over the training data sets and then divides it by the number of training data sets.

[0132] Based on the validation set of 80 datasets and their associated annotations, the best-performing model was selected from several machine learning models (with different hyperparameters, e.g., number of layers, size and number of kernels, padding, etc.). Specificity and sensitivity were determined based on the test set of 100 datasets and their associated annotations.

[0133] Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited to the disclosed examples and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention.

[0134] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

Claims

[1] Computer-implemented method for validating processing data (1), comprising the steps: - Receiving source data (2) based on medical image data acquisition, - receiving or determining the processing data (1) based on an application of a main processing algorithm (10) to the original data (1), - comparing comparison data (11), which are the processing data (1) or are based on the processing data (1), with reference data (12), which are the original data (1) or are based on an application of a reference processing algorithm (13) to the original data (1), by a comparison algorithm (14) in order to determine a comparison result (15), - evaluating a trigger condition (16, 17), the fulfillment of which depends on the comparison result (15), and, - if the trigger condition (16, 17) is met • issuing a notice (18) to a user, and / or • Modification of a predetermined acquisition parameter (19) for subsequent image data acquisition, and / or • Providing new processing data (20), wherein either the reference data (12) are used as new processing data (20) or wherein the new processing data (20) are determined by applying an alternative processing algorithm (21) to the original data (2). [2] Computer-implemented method according to claim 1, characterized by that, on the one hand, the main processing algorithm (10) serves to improve the image and / or to change an image impression and / or that, on the other hand, the original data (2) comprise a plurality of two-dimensional image data sets (22), wherein the main processing algorithm (10) is or comprises a reconstruction of a three-dimensional or four-dimensional image data set (23) from these two-dimensional image data sets (22). [3] Computer-implemented method according to claim 1 or 2, characterized by that the main processing algorithm (10) and, on the one hand, the reference processing algorithm (13) and / or, on the other hand, the alternative processing algorithm (21) are implemented by mutually different parameterization of a basic processing algorithm (24). [4] Computer-implemented method according to one of the preceding claims, characterized byin that the original data (2) comprise a plurality of two-dimensional image data sets (22), wherein the processing data (1) are formed by a three-dimensional or four-dimensional image data set (23) which is reconstructed by the main processing algorithm (10) from these two-dimensional image data sets (22), wherein the comparison data (11) are or comprise two-dimensional image data sets (26) which are determined by a respective forward projection (27) of this three-dimensional or four-dimensional image data set (23). [5] Computer-implemented method according to one of the preceding claims, characterized bythat an evaluation algorithm (28) is applied, on the one hand, to the comparison data (11) in order to recognize segments (29) of the comparison data (11) which each depict a relevant anatomical feature and / or a medical device, and, on the other hand, is applied to the reference data (12) in order to recognize segments (30) of the reference data (12) which each depict a relevant anatomical feature and / or medical device, - wherein the comparison result (15) depends on a comparison of the number and / or positions and / or dimensions of the segments (29) in the comparison data (11) with the number and / or positions and / or dimensions of the segments (30) in the reference data (12), and / or - wherein a bounding box (31, 32) is determined for the respective segment (29, 30) both in the reference data (12) and in the comparison data (11), wherein the comparison result (15) depends on a comparison of the positions and / or the dimensions of the bounding boxes (31) in the comparison data (11) with the positions and / or the dimensions of the bounding boxes (32) in the reference data (12). [6] Computer-implemented method according to one of the preceding claims, characterized by that the comparison algorithm (14) is or comprises a model (33, 34, 35) trained by machine learning. [7] Computer-implemented method according to one of the preceding claims, characterized by that the comparison algorithm (14) comprises at least a first and a second model (34, 35) trained by machine learning, which each process the comparison data (11) on the one hand and the reference data (12) on the other hand as input data (36, 37), wherein the first machine learning-trained model (34) determines a first intermediate result (38) relating to the presence and / or extent of a deviation of a first deviation type between the comparison data (11) and the reference data (12), wherein the second model (35) trained by machine learning determines a second intermediate result (39) relating to the presence and / or extent of a deviation of a second deviation type different from the first deviation type between the comparison data (11) and the reference data (12), wherein the comparison result (15) depends on the first and the second intermediate result (38, 39). [8] Computer-implemented method according to claim 6 or 7, characterized bythat a discriminator (40) is used as the model (33, 34, 35) trained by machine learning or as at least one of the models (33, 34, 35) trained by machine learning, the parameterization of which is based on a training of a generative adversarial network (41) comprising the discriminator (40). [9] Computer-implemented method according to one of the preceding claims, characterized by that the comparison result (15) describes the presence and / or extent of a deviation of at least one of the following deviation types: - an absence, interruption and / or a shortened representation of at least one vessel, and / or - the absence of an image of a medical device, - and / or the absence of an image of an anatomical feature, on the one hand in the comparison data (11) compared to the reference data (12) and / or on the other hand in the reference data (12) compared to the comparison data (11). [10] Computer-implemented method according to one of the preceding claims, characterized by that preceding original data (42) are received which are based on a previous medical image data acquisition which took place before the medical image data acquisition, wherein the or an acquisition parameter (19) for the or a subsequent image data acquisition is predetermined as a function of a further comparison result (43) which is determined by the comparison algorithm (14) by comparing current input data (44) which correspond to or are based on the original data (2) with preceding input data (45) which correspond to or are based on the previous original data (42). [11] Computer-implemented method according to one of the preceding claims, characterized bythat the comparison information (15), at least in the case that the triggering condition (16) is fulfilled, relates to at least one segment of the comparison data (11) in which the comparison data (11) deviate from the reference data (12), wherein the indication (18) comprises segment information (46) relating to this segment. [12] Computer-implemented method for providing a model (33, 34, 35) trained by machine learning for use as a comparison algorithm (14) or as a sub-algorithm of the comparison algorithm (14) in the computer-implemented method according to one of the preceding claims, comprising the steps: - receiving input training data (47) comprising a plurality of training data sets (48), each of which in turn comprises training data (49) based on a medical image data acquisition and / or a simulation of a medical image data acquisition, - training a model based on the input training data (47) to determine the machine learning trained model (33, 34, 35), - Provision of the model trained by machine learning (33, 34, 35). [13] Processing device, characterized by that it is designed to carry out the computer-implemented method according to one of the preceding claims. [14] Computer program with instructions which are designed to carry out the computer-implemented method according to one of claims 1 to 13 when executed on a data processing device (51). [15] Data carrier comprising a computer program (9) according to claim 14.

Images