Providing a result image data set

Abstract

Computer-implemented method for providing an output image data set (OD), comprising - Receiving (REC-ID) an input image data set (ID) of a first examination volume (VOL), wherein the input image data set (ID) is an X-ray image data set of the first examination volume (VOL), - Receiving or determining (REC-DET-IP) an input image parameter (IP), where the input image parameter (IP) relates to a property of the input image record (ID), - Receiving or determining (REC-DET-OP) a result image parameter (OP), - Determining (DET-OD) a result image data set (OD) of the first investigation volume (VOL) by applying a trained generator function (GF) to input data, where the input data is based on the input image data set (ID) and the result image parameter (OP), where the input data continues to be based on the input image parameter (IP), where the result image parameter relates to one or more of the following properties of the result image data set (OD): -- X-ray dose of the result image dataset (OD), -- Noise level of the result image data set (OD), - X-ray source (XSYS.SRC) and / or X-ray detector (XSYS.DTC) corresponding to the result image data set (OD), and wherein a parameter of the trained generator function (GF) is based on a generative adversarial algorithm, - Provisioning (PROV-OD) of the result image data set (OD).

Description

[0001] In medical practice, X-ray devices are often used to monitor (especially minimally invasive) surgical procedures; in some cases, certain procedures are only possible through X-ray monitoring, for example, the implantation of an aortic valve using a catheter.

[0002] The advantages of such an X-ray-guided procedure must be weighed against the radiation exposure from the absorbed X-ray dose. Since reducing the X-ray dose typically also leads to a reduction in image quality or an increase in the signal-to-noise ratio, a compromise often has to be found between good image quality and a low X-ray dose.

[0003] If the signal-to-noise ratio is too high, this can lead to low image quality, especially in digital subtraction angiography (DSA). In particular, the necessary registration of the mask image dataset and the fill image dataset may not be possible, or the noise may cause artifacts during registration.

[0004] It is known that image quality can be improved through various noise reduction methods. However, these methods can alter the image appearance and / or lead to artifacts. If noise reduction is applied too extensively, it can, for example, result in a cartoonish image appearance.

[0005] Furthermore, it is known that the signal-to-noise ratio can be optimized by using optimized protocols (i.e., by optimizing the parameters of the X-ray device). However, this choice of protocols can also alter the image appearance (for example, the values ​​of the image pixels for the same imaged structure can differ depending on the image acquisition parameters). This poses a particular problem when the image data is to be further processed by trained algorithms, especially if the algorithms used for training were acquired using only a limited number of protocols or a limited number of X-ray devices.

[0006] Methods for generating synthetic CT image datasets based on MR image datasets are known from the publication NIE, D. et al.: Medical Image Synthesis with Context-Aware Generative Adversarial Networks. arXiv, 2016, URL: https: / / arxiv.org / abs / 1612.05362.

[0007] The determination of synthetic two-dimensional mammography datasets based on three-dimensional mammography datasets using generative adversial networks is known from publication US 2019 / 0 090 834 A1.

[0008] Generative adversarial networks and conditional statements and branches are described in the documents "Generative Adversarial Networks". In: Wikipedia. URL:https: / / de.wikipedia.org / w / index.php?title=Generative_Adversarial_Networks&oldid=186990065, accessed March 27, 2019, and "Conditional Instruction and Branching". In: Wikipedia. URL:https: / / de.wikipedia.org / w / index.php?title=Bedingte_Anweisung_und_Verzweigung&oldid=185434103, accessed February 6, 2019.

[0009] The automated selection of examination modalities using a symptom- and / or diagnosis-based database is known from publication DE 101 56 215 A1.

[0010] From publication US 2019 / 0108904 A1, it is known to perform noise reduction of a computed tomography image by applying a neural network selected from a large number of neural networks.

[0011] The use of so-called "generative adversarial networks" for generating synthetic medical image data is known from the publication Yi, X. [et al.]: “Generative Adversaria! Network in Medical Imaging: A Review”, preprint 1809.07294v2 on arxiv.org.

[0012] From the publication KANG, E. [et al.]: Wavelet Domain Residual Network (WaveResNet) for Low-Dose-X- Ray CT Reconstruction, Preprint 1703.01383v1 on arxiv.org, March 4, 2017, a direct residual learning approach in the directional wavelet domain for improving the application of deep convolutional neural networks (CNNs) for low-dose X-ray CT is known.

[0013] Methods for the unsupervised transformation of videos into videos in the field of machine learning are known from the publication BASHKIROVA, D. [et al.]: Unsupervised Video-to-Video-Translation. Preprint 1806.03698v1 on arxiv.org, June 10, 2018.

[0014] Methods for image transformation with regard to X-ray doses are known from publication US 2019 / 0 035 118 A1.

[0015] The use of generative adversarial networks for noise reduction is known from the publication WOLTERINK, Jelmer M. [et al.]: Generative adversarial networks for noise reduction in low-dose CT. IEEE transactions on medical imaging, 2017, Vol. 36, No. 12, pp. 2536-2545.

[0016] The object of the invention is therefore to provide a way to achieve an increase in the signal-to-noise ratio without changing the image impression.

[0017] This problem is solved by a computer-implemented method for providing a result image data set, by a computer-implemented method for providing a trained generator function and / or a trained classifier function, by a provisioning system, by an X-ray device, by a computer program product, and by a computer-readable storage medium according to the independent claims. Advantageous embodiments are described in the dependent claims and in the following description.

[0018] The inventive solution to the problem is described below with regard to both the claimed devices and the claimed method. Features, advantages, or alternative embodiments mentioned herein are also applicable to the other claimed items and vice versa. In other words, the claims (which, for example, relate to a device) can also be further developed with the features described or claimed in connection with a method. The corresponding functional features of the method are thereby implemented by corresponding physical modules.

[0019] Furthermore, the inventive solution to the problem is described both with regard to methods and devices for providing result image data sets and with regard to methods and devices for providing trained generator functions and / or trained classifier functions. Features and alternative embodiments of data structures and / or functions in methods and devices for providing result image data sets can be transferred to analogous data structures and / or functions in methods and devices for providing trained generator functions and / or trained classifier functions. Analogous data structures can be characterized in particular by the use of the prefix "training".Furthermore, the trained generator functions and / or trained classifier functions used in methods and devices for providing result image datasets may have been adapted and / or provided, in particular, by methods and devices for providing trained generator functions and / or trained classifier functions.

[0020] The invention relates, in a first aspect, to a computer-implemented method for providing a result image data set. The method is based on receiving an input image data set of a first examination volume, in particular via an interface. Furthermore, a result image parameter is received or determined, in particular via the interface or a processing unit. A result image data set of the first examination volume is then determined by applying a trained generator function to the input data, in particular via the processing unit. Here, the input data is based on the input image data set and the result image parameter, and the result image parameter relates to a property of the result image data set.Furthermore, one parameter of the trained generator function is based on a GA algorithm (an acronym for "generative adversarial"). Additionally, the resulting image dataset is provided, particularly via the interface. Providing the resulting image dataset can include, in particular, displaying, transmitting, and / or storing the dataset.

[0021] An image dataset comprises at least one image; furthermore, an image dataset may also include additional data, particularly metadata. An image dataset may, in particular, be identical to the image itself. An image dataset may, in particular, be the result of a medical imaging examination, especially the result of a medical X-ray examination. In this case, the image dataset may also be referred to as an X-ray image dataset.

[0022] A two-dimensional image dataset comprises at least one two-dimensional image; in particular, a two-dimensional image dataset does not include any further image with a different dimension. A three-dimensional image dataset comprises at least one three-dimensional image; in particular, a three-dimensional image dataset does not include any further image with a different dimension. A four-dimensional image dataset comprises at least one four-dimensional image; in particular, a four-dimensional image dataset does not include any further image with a different dimension. A two-dimensional image dataset can, in particular, be identical to a two-dimensional image. A three-dimensional image dataset can, in particular, be identical to a three-dimensional image. A four-dimensional image dataset can, in particular, be identical to a four-dimensional image.

[0023] Metadata is information about the imaging examination that forms the basis for the image dataset, which is not images or image data, such as patient data or data about the protocol used.

[0024] An image dataset comprises a plurality of pixels or voxels. The terms pixel and voxel are used synonymously here, and therefore do not imply any dimensionality. Each pixel or voxel is assigned an intensity value, preferably corresponding to an X-ray absorption value.

[0025] The input image dataset is, in particular, an image dataset. The result image dataset is, in particular, an image dataset. The input image dataset and the result image dataset, in particular, have the same dimensionality.

[0026] An image dataset can be referred to as a real image dataset, in particular if it is the direct result of an imaging examination, or was generated directly from the raw data produced by the imaging examination. An image dataset can be referred to as a synthetic image dataset or a virtual image dataset if it is not a real image dataset, in particular if it is the result of applying a trained generator function. The input image dataset is, in particular, a real image dataset, while the output image dataset is, in particular, a virtual image dataset.

[0027] An image parameter refers in particular to a parameter of an image data set (for example, brightness, contrast, sharpness, signal-to-noise ratio) or a parameter of the acquisition of an image data set or a parameter of an image acquisition device during the acquisition of the image data set (for example, exposure time, aperture size, especially in the case of X-ray image data sets or X-ray devices also the X-ray voltage, the X-ray current, the X-ray dose, a designation of the X-ray source and / or a designation of the X-ray detector). The resulting image parameter is in particular an image parameter.

[0028] A real image parameter relates specifically to a real image data set. A synthetic image parameter relates specifically to a synthetic image data set. A synthetic image parameter can also relate to a parameter of an image acquisition device; in this case, it relates to a parameter of the image acquisition device that would be necessary or would have to be used to acquire a real image data set comparable to the synthetic image data set, in particular to acquire a real image data set identical to the synthetic image data set.

[0029] A trained function is, in particular, a function that maps input data to output data, where the output data continues to depend on at least one function parameter, and where the function parameter can be adapted by supervised learning, semi-supervised learning, and / or unsupervised learning. The input data and / or the output data can each, in particular, comprise at least one image dataset.

[0030] In particular, a trained generator function is a trained function, and a trained classifier function is a trained function. For a function to be considered trained, none of its parameters necessarily need to have been fitted; therefore, the term "trained function" can also be replaced by "trainable function." In particular, the term "trained generator function" can be replaced by "trainable generator function," and / or the term "trained classifier function" can be replaced by "trainable classifier function." Specifically, the terms "trained generator function" and "generator function" can be used synonymously, and / or the terms "trained classifier function" and "classifier function" can be used synonymously.

[0031] In a trained generator function, the output data includes at least one image dataset, while the input data may optionally include an image dataset. In a trained classifier function, the input data includes at least one image dataset, while the output data includes a classifier and / or one or more probability values. The classifier corresponds, in particular, to an estimate for an image parameter, and the probability value corresponds, in particular, to the probability that the image dataset in the input data is a real image dataset or a synthetic image dataset. The term "classifier function" can therefore be replaced, in particular, by the term "discriminator function" and / or by the term "discriminator and classifier function."

[0032] A GA algorithm ("GA" is an acronym for "generative adversarial") comprises a generator function and a classifier function. The generator function produces synthetic data (also called "virtual data"), and the classifier function distinguishes between synthetic and real data. Specifically, training the generator function and / or the classifier function ensures that, on the one hand, the generator function produces synthetic data that the classifier function incorrectly classifies as real, and on the other hand, that the classifier function can distinguish as accurately as possible between real and synthetic data. From a game-theoretic perspective, a GA algorithm can also be viewed as a zero-sum game.The training of the generator function and / or the classifier function is based in particular on minimizing one cost function at a time.

[0033] If the generator function and the classifier function are given by a network, especially an artificial neural network, then the GA algorithm is also called a GA network (also "GAN," an acronym for "generative adversarial network"). These are particularly well-known from the publication by Ian J. Goodfellow, "Generative Adversarial Networks," arXiv 1406.2661 (2014). Minimizing the cost function can be achieved, in particular, by backpropagation.

[0034] In particular, a parameter of the trained generator function is based on a GA algorithm or a GA network such that the parameter of the trained generator function is identical to a parameter of the generator function of the GA algorithm or the GA network. Specifically, a parameter of the trained generator function is based on a GA algorithm or a GA network such that the trained generator function is identical to the generator function of the GA algorithm or the GA network.

[0035] The inventors have discovered that by using a trained generator function, where one parameter of the trained generator function is based on a GA algorithm, result image datasets can be generated that have a predefined image appearance. This is because the use of a classifier function in the GA algorithm trains the generator function to achieve the predefined image appearance. The image appearance of the result image dataset can be specifically defined by the result image parameter.

[0036] According to a further aspect of the invention, the method further comprises receiving or determining an input image parameter, wherein the input image parameter relates to a property of the input image data set, and wherein the input data is furthermore based on the input image parameter. The receiving or determination of the input image parameter is performed in particular by means of the interface or the processing unit. An input image parameter is in particular an image parameter.

[0037] The input image parameter can be contained within the input image data record, for example, in the metadata of the input image data record. The input image data record can also be calculated or determined from the input image data record.

[0038] The inventors realized that by using the input image parameter in the input data of the trained generator function, it is no longer necessary for the trained generator function to derive it from the input image dataset. This allows for more efficient training of the generator function (since, for a given number of parameters, less output data needs to be calculated, and therefore less training data is required). Furthermore, the method is less error-prone, as it allows for the use of an exact value of the input image parameter, rather than a potentially erroneous one.

[0039] According to a further aspect of the invention, the input image data set is an X-ray image data set of the first examination volume, and the output parameter relates to one or more of the following properties of the output image data set: X-ray dose of the output image data set, noise level of the output image data set, and the X-ray source and / or X-ray detector corresponding to the output image data set. Optionally, particularly when an input image parameter is received or determined, the input image data set relates to one or more of the following properties of the input image data set: X-ray dose of the input image data set, noise level of the input image data set, and the X-ray source and / or X-ray detector corresponding to the input image data set.

[0040] The X-ray dose of the input image dataset corresponds specifically to the X-ray dose used when acquiring the input image dataset, or the X-ray dose absorbed by the first examination volume during the acquisition of the input image dataset. The X-ray dose of the result image dataset corresponds specifically to the X-ray dose that would be used in a hypothetical acquisition of the result image dataset, or the hypothetical X-ray dose absorbed by the first examination volume during a hypothetical acquisition. It is important to note that the result image dataset does not originate from an actual acquisition of the first examination volume, but is determined based on the input image dataset. Therefore, no actual acquisition is associated with the result image dataset. The X-ray dose is thus the X-ray dose that would be necessary to produce the corresponding image impression of the result image dataset.

[0041] The noise level of an image data set corresponds in particular to the signal-to-noise ratio of the image data set.

[0042] The X-ray source and / or X-ray detector corresponding to the input image dataset is, in particular, the X-ray source and / or X-ray detector that was used when acquiring the input image dataset. The X-ray source and / or X-ray detector corresponding to the result image dataset is, in particular, the X-ray source and / or X-ray detector that would be used in a hypothetical acquisition of the result image dataset. It should be noted that the result image dataset does not originate from an actual acquisition of the first examination volume, but was determined based on the input image dataset. Therefore, no actual acquisition is assigned to the result image dataset. The X-ray source and / or X-ray detector corresponding to the result image dataset is therefore the one that would be used in a hypothetical acquisition of the result image dataset.the X-ray source and / or X-ray detector that would be necessary to generate the corresponding image impression of the result image data set.

[0043] If the result image parameter includes the X-ray dose of the result image dataset, then the optional input image parameter also includes the X-ray dose of the input image dataset. If the result image parameter includes the noise level of the result image dataset, then the optional input image parameter also includes the noise level of the input image dataset. If the result image parameter includes the X-ray source and / or X-ray detector corresponding to the result image dataset, then the optional input image parameter also includes the X-ray source and / or X-ray detector corresponding to the input image dataset.

[0044] The inventors recognized that by using the X-ray dose or noise level as output image parameters, the actual X-ray dose used can be kept low, while simultaneously providing an X-ray image dataset with a low noise level. By using the X-ray source and / or the X-ray detector as output image parameters, image impressions can be generated that correspond to an image taken by another X-ray device different from the one used. This allows, in particular, the use of image processing routines that are specifically tailored to images from the other X-ray device.

[0045] According to another possible aspect, the invention is further based on the fact that a comparison image data set of a second examination volume is received, in particular via the interface, wherein the first examination volume and the second examination volume overlap and / or are identical. Furthermore, a comparison image parameter is determined, in particular by means of the processing unit, wherein the comparison image parameter relates to a property of the comparison image data set. The input image parameter and the comparison image parameter are then compared, and the determination of the result image data set is carried out if the comparison image parameter differs from the input image parameter. Here, the result image parameter used corresponds to the comparison image parameter.

[0046] The comparison image dataset is, in particular, an image dataset, and the comparison image parameter is, in particular, an image parameter. The comparison image parameter relates, in particular, to one or more of the following properties of the comparison image parameter: the X-ray dose of the comparison image parameter, the noise level of the comparison image parameter, and the X-ray source and / or X-ray detector corresponding to the comparison image parameter. The X-ray dose of the comparison image dataset corresponds, in particular, to the X-ray dose used during the acquisition of the comparison image dataset or to the X-ray dose absorbed by the second examination volume during the acquisition of the comparison image dataset. The X-ray source and / or X-ray detector corresponding to the comparison image dataset is, in particular, the X-ray source and / or X-ray detector that was used during the acquisition of the comparison image dataset.

[0047] The inventors recognized that by comparing the input image data with a given reference image set, a multiple set of image data with an identical or similar visual appearance can be generated, while simultaneously determining a result image data set only needs to be performed if a deviation between the input image parameter and the reference image parameter necessitates it. This avoids unnecessary calculations. This is particularly advantageous when the described method is to be performed on input image data sets within an image stream (for example, during X-ray monitoring of a surgical procedure).

[0048] According to a further aspect, the invention is based on the fact that a comparison image dataset of a second examination volume is received, in particular via the interface, wherein the first examination volume and the second examination volume overlap and / or are identical. Furthermore, a comparison image parameter is determined, in particular by means of the processing unit, by applying a trained classifier function to the comparison image dataset, wherein the comparison image parameter relates to a property of the comparison image dataset, and wherein a parameter of the trained classifier function is based on the GA algorithm. Furthermore, the input image parameter and the comparison image parameter are compared, and the determination of the result image dataset is carried out if the comparison image parameter differs from the input image parameter. Here, the result image parameter used corresponds to the comparison image parameter.In particular, the input image parameter can also be determined by applying the trained classifier function to the input image data set.

[0049] The inventors have recognized that by using the trained classifier function, it is particularly possible to use comparison image parameters or input image parameters that are not contained in the comparison image dataset or the input image dataset itself, or that cannot be derived from the comparison image dataset or the input image dataset by a simple calculation.

[0050] According to a further aspect of the invention, the method further comprises adjusting an imaging parameter of an imaging unit based on a comparison of the input image parameter and the comparison image parameter. The imaging unit may, in particular, be an X-ray device, and the imaging parameter, in the case of an X-ray device, may, in particular, be the X-ray voltage, the X-ray current, and / or an exposure time.

[0051] Adjusting the imaging parameter can particularly involve controlling the imaging parameter to reduce the difference between the input image parameter and the comparison image parameter.

[0052] The inventors recognized that by adjusting or regulating the imaging parameter, the image appearance of successive image data sets can be standardized. In particular, this can improve the application of the described method for subsequent image data sets from the imaging unit, as artifacts can be reduced when the input image parameter and the output image parameter are similar.

[0053] According to a further aspect of the invention, the method further comprises determining an input frequency data set based on the input image data set, wherein the input frequency data set is a representation of the input image data set in the frequency domain, wherein the input data is based on the input frequency data set, and wherein the application of the trained generator function to the input data produces a result frequency data set, wherein the result frequency data set is a representation of the result image data set in the frequency domain. The determination of the input frequency data set is carried out in particular by means of the processing unit.

[0054] In particular, the input frequency data set is a Fourier transform of the input image data set, and the result frequency data set is a Fourier transform of the result image data set. Specifically, the input frequency data set can also be a wavelet decomposition of the input image data set, and the result frequency data set can be a wavelet decomposition of the result image data set.

[0055] In particular, the input data can be based on the input frequency data set in such a way that the input data includes the input frequency data set below a predetermined cutoff frequency (low-frequency components), and that the input data does not include the input frequency data set above the predetermined cutoff frequency (high-frequency component).

[0056] The inventors have realized that by applying the method in the frequency domain instead of the real domain, the edges (corresponding to the high-frequency components) can be better preserved in the image data sets, and thus the structural information can be preserved in the resulting image data set.

[0057] According to a further aspect of the invention, the result image parameter is tailored to a trained image processing function. The result image parameter is particularly tailored to the trained image processing function if the trained image processing function is configured to process image data sets with properties described by the result image parameter.

[0058] The image processing function can, in particular, be a trained function, especially an artificial neural network. The trained image processing function is specifically designed to process image datasets with properties described by the result image parameter, provided that the training image datasets used to train the image processing function, or a subset thereof, exhibit the properties described by the result image parameter.

[0059] The inventors realized that by tailoring the result image parameter to the trained image processing function, it can also be applied to image data that differs from the training image datasets used for training. In particular, the trained image processing function can therefore be trained more efficiently based on a smaller number of training image datasets, and its application is less error-prone.

[0060] According to a further aspect of the invention, the method further comprises receiving a mask image data set from a third examination volume, wherein the first examination volume and the third examination volume overlap and / or are identical, and wherein the input data is further based on the mask image data set. The reception is performed in particular via the interface. Specifically, in this aspect, the input image data set is a fill image data set, and the result image data set is a difference image data set.

[0061] A mask image dataset is, in particular, an image dataset of an examination volume that does not contain any contrast medium at the time the mask image dataset is acquired. Specifically, a mask image dataset is an X-ray image dataset. A fill image dataset is, in particular, an image dataset of an examination volume that contains contrast medium at the time the fill image dataset is acquired. A difference image dataset is, in particular, an image dataset of an examination volume that represents the difference between a fill image dataset and a mask image dataset.

[0062] The inventors recognized that by using the trained generator function, a result image dataset or a difference image dataset can be generated, particularly with a lower signal-to-noise ratio. Furthermore, the trained generator function can also compensate for deviations in the image geometry between the fill image dataset and the mask image dataset without requiring resource-intensive registration between the images.

[0063] According to another aspect of the invention, the parameter of the trained generator function is based on a cyclic consistency cost function and / or an information loss cost function. An alternative term for cyclic consistency cost function is reconstruction cost function. According to another possible aspect of the invention, the parameter of the trained generator function is based on an adversarial cost function and / or a classification cost function. According to yet another possible aspect of the invention, a parameter of the trained classifier function is based on an adversarial cost function and / or a classification cost function.

[0064] A parameter of a trained function is based on a cost function, in particular if the parameter has been adjusted or changed to minimize or maximize this cost function.

[0065] A cyclic consistency cost function is, in particular, a cost function based on a comparison of two applications of the trained generator function to the input data. Specifically, the cyclic consistency cost function is based on a norm of the difference between the input image dataset and the application of the trained generator function to input data comprising the result image dataset and the input image parameter. The norm can be, in particular, a 1-norm or a 2-norm, and can be evaluated pixel-wise and / or voxel-wise.

[0066] By using a cyclic consistency cost function, applying the trained generator function twice (each time with appropriate image parameters) can be done as an identity mapping. This allows the trained generator function to produce synthetic image datasets for given image parameters particularly well.

[0067] An information loss cost function is, in particular, a cost function based on noise averaging; specifically, the information loss cost function is based on an averaging of noise in differently sized sub-areas of the result image data set OD and / or the input image data set.

[0068] By using an information loss cost function, it is possible to exploit the fact that real noise has a mean value of zero, or that the noise-averaged intensity value of pixels and / or voxels in a region of an image dataset corresponds to the mean intensity value. In particular, using this cost function makes it possible to preserve the low-frequency and high-frequency components of the input image dataset in the resulting image dataset.

[0069] An adversarial cost function is, in particular, a cost function that measures the goodness of the trained classifier function's ability to distinguish between real and synthetic data. An adversarial cost function for the trained generator function can, in particular, have a first value if the resulting image dataset is recognized as a synthetic image dataset by the trained classifier function, and a second value if the resulting image dataset is recognized as a real image dataset by the trained classifier function, with the first value being greater than the second value.An adversarial cost function for the trained classifier function can, in particular, have a third value if the result image dataset is recognized by the trained classifier function as a synthetic image dataset, and a fourth value if the result image dataset is recognized by the trained classifier function as a real image dataset, where the fourth value is greater than the third value. An adversarial cost function for the trained classifier function can, in particular, have a fifth value if a real comparison image dataset is recognized by the trained classifier function as a synthetic image dataset, and a sixth value if the real comparison image dataset is recognized by the trained classifier function as a real image dataset, where the fifth value is greater than the sixth value.

[0070] By using an adversarial cost function, the trained generator function and the trained classifier function can be trained in such a way that the trained generator function produces result image datasets that are hardly distinguishable from real image datasets by the trained classifier function, and that the trained classifier function can simultaneously distinguish as accurately as possible between real and synthetic image datasets.

[0071] A classification cost function is, in particular, a cost function based on a comparison (especially a difference) of the result image parameter and a calculated image parameter of the result image dataset. Specifically, the classification cost function reaches its optimum (especially its minimum) when the result image parameter and the calculated image parameter of the result image dataset are identical.

[0072] By using a classification cost function, the trained generator function can be made to produce result image datasets that are described as accurately as possible by the result image parameter. Similarly, the trained classifier function can be made to ensure that the image parameter determined by the trained classifier function corresponds as closely as possible to the actual image parameter.

[0073] According to another possible aspect of the invention, the trained generator function and / or the trained classifier function comprise a convolutional layer. According to another possible aspect of the invention, the trained generator function and / or the trained classifier function comprise a deconvolutional layer. According to another possible aspect of the invention, the trained generator function and / or the trained classifier function comprise a residual block.

[0074] A convolution layer, in particular, replicates the mathematical operation of a convolution with one or more convolution kernels, where the elements of the convolution kernel correspond to weights of the neural network. A deconvolution layer, in particular, replicates the mathematical operation of a deconvolution with one or more convolution kernels. In a residual block, a node layer of the neural network is connected not only to the immediately following layer but also to one of the subsequent layers.

[0075] The inventors recognized that convolutional layers are particularly well-suited for identifying and processing features of image datasets. Specifically, different convolutional kernels can be used to analyze various features of the processed image datasets (e.g., edges or gradients). Deconvolutional layers, in particular, can convert predefined features (e.g., edges or gradients) back into their corresponding image datasets. A suitable combination of convolutional and deconvolutional layers can act as an autoencoder. Furthermore, the inventors discovered that using residual blocks improves the training of layers near the input layers of the neural network and enables the solution of problems involving vanishing gradients.

[0076] According to a further aspect of the invention, the input image data set comprises a temporal sequence of input image data of the first investigation volume, and wherein the result image data set comprises a temporal sequence of result image data of the first investigation volume.

[0077] The inventors recognized that image noise, in particular, is also statistically independent to a good approximation in temporally independent images. Therefore, by using a temporal sequence of input image data as input data for the trained generator function, a result image dataset with the specified result image parameter can be generated particularly well. This also applies if the multiple result image datasets have different imaging geometries (for example, if they were captured with respect to different projection directions).

[0078] According to another aspect of the invention, the trained generator function and / or the trained classifier function was provided by a method for providing a trained generator function and / or a trained classifier function according to the invention, in particular according to the second aspect of the invention.

[0079] In a second aspect, the invention relates to a method for providing a trained generator function and / or a trained classifier function. Here, an input image dataset and a comparison image dataset of a first investigation volume are received, in particular via a training interface. Furthermore, also in particular via the interface, a result image parameter is received, wherein the result image parameter relates to a property of the comparison image dataset.

[0080] The procedure further comprises determining a result image dataset of the first investigation volume by applying a trained generator function to input data, in particular by means of a computing unit, wherein the input data is based on the input image dataset and the result image parameter. Furthermore, also in particular by means of the computing unit, a result classifier and a comparison classifier are determined by applying a trained classifier function to the result image dataset and the comparison image dataset. In particular, the result classifier is determined by applying the trained classifier function to the result image dataset, and the comparison classifier is determined by applying the trained classifier function to the comparison image dataset.

[0081] The method further comprises adjusting a parameter of the trained generator function and / or a parameter of the trained classifier function based on the result classifier and the comparison classifier, in particular based on a comparison of the result classifier and the comparison classifier. Furthermore, the trained generator function and / or the trained classifier function are provided, whereby providing may in particular include storing, transmitting, and / or displaying the trained generator function and / or the trained classifier function.

[0082] In the procedure for providing a trained generator function and / or a trained classifier function, the input image dataset can also be referred to as the training input image dataset, the output image dataset can also be referred to as the training output image dataset, and the comparison image dataset can, in particular, also be referred to as the training comparison image dataset. Furthermore, in this procedure, the output image parameter can also be referred to as the training output image parameter.

[0083] An image classifier of an image dataset includes, in particular, a probability value and / or an image parameter of the image dataset. The probability value corresponds, in particular, to the probability that the image dataset corresponds to a real image dataset.

[0084] The result classifier is specifically an image classifier of the result image dataset. The comparison classifier is specifically an image classifier of the comparison image dataset.

[0085] The parameter is adjusted, in particular, by a general algorithm. This adjustment can be achieved by optimizing a cost function, for example, using backpropagation. The cost function can be based on the outcome classifier and the comparison classifier. Specifically, the cost function can be based on the probability values ​​of the outcome classifier and the comparison classifier, and / or the cost function can be based on a comparison of the image parameters of the outcome classifier and the comparison classifier.The cost function can be based, in particular, on the probability value of the outcome classifier by basing the cost function on a deviation of the probability value of the outcome classifier from 0, and the cost function can be based, in particular, on the probability value of the comparison classifier by basing the cost function on a deviation of the probability value of the comparison classifier from 1.

[0086] The inventors have recognized that the described method can provide a trained generator function that produces result image datasets whose properties correspond to the respective comparison image datasets and have a similar overall impression.

[0087] Because the result image parameter describes a property of the comparison image dataset, and the training takes place within a GA algorithm whose optimization can generate result image datasets with similar or identical properties to the comparison image datasets, the properties of the result image dataset are also described by the result image parameter. Therefore, it is possible to provide a trained generator function using the described method, which can generate artificial result image datasets with predefined parameters.

[0088] In a third aspect, the invention relates to a provision system for providing a result image data set, comprising an interface and a computing unit, - wherein the interface is designed to receive an input image data set of a first examination volume, - wherein the interface and / or the computing unit are configured to receive or determine a result image parameter, - wherein the computing unit is further trained to determine a result image data set of the first investigation volume by applying a trained generator function to input data, wherein the input data is based on the input image data set and the result image parameter, wherein the result image parameter relates to a property of the result image data set, and wherein a parameter of the trained generator function is based on a GA algorithm, - and where the interface is still configured to provide the result image data set.

[0089] Such a provisioning system can, in particular, be configured to execute the previously described methods according to the invention for providing a result image data set and their aspects. The provisioning system is configured to execute these methods and their aspects by providing the interface and the processing unit to perform the corresponding process steps.

[0090] In a fourth aspect, the invention relates to an X-ray device comprising an X-ray source and an X-ray detector, further comprising a delivery system according to the third aspect of the invention. The X-ray device may, in particular, be a C-arm X-ray device or a computed tomography scanner.

[0091] In a possible fifth aspect, the invention relates to a training system for providing a trained generator function and / or a trained classifier function, comprising a training interface and a training computing unit. - wherein the training interface is designed to receive an input image data set and a comparison image data set of a first examination volume, - wherein the training interface is further configured to receive a result image parameter, where the result image parameter relates to a property of the comparison image data set, - wherein the training computing unit is trained to determine a result image data set of the first investigation volume by applying a trained generator function to input data, wherein the input data is based on the input image data set and the result image parameter, - wherein the training computing unit is further trained to determine a result classifier and a comparison classifier by applying a trained classifier function to the result image dataset and the comparison image dataset, - wherein the training computing unit is further trained to adjust a parameter of the trained generator function and / or the trained classifier function based on the result classifier and the comparison classifier, - wherein the training interface is still configured to provide the trained generator function and / or the trained classifier function.

[0092] Such a training system can, in particular, be configured to execute the previously described methods according to the invention for providing a trained generator function and / or a trained classifier function. The training system is configured to execute these methods and their aspects by providing a training interface and a training processing unit capable of performing the corresponding process steps.

[0093] In a sixth aspect, the invention relates to a computer program product comprising a computer program that can be directly loaded into a memory of a provisioning system, with program sections to execute all steps of the method for providing a result image data set and its aspects when the program sections are executed by the provisioning system; and / or which can be directly loaded into a training memory of a training system, with program sections to execute all steps of the method for providing a trained generator function and / or a trained classifier function and its aspects when the program sections are executed by the training system.

[0094] In particular, the invention relates to a computer program product comprising a computer program that can be directly loaded into a memory of a provisioning system, with program sections to execute all steps of the method for providing a result image data set and its aspects when the program sections are executed by the provisioning system.

[0095] In particular, the invention relates to a computer program product comprising a computer program that can be directly loaded into a training memory of a training system, with program sections to execute all steps of the method for providing a trained generator function and / or a trained classifier function and its aspects when the program sections are executed by the training system.

[0096] In a seventh aspect, the invention relates to a computer-readable storage medium on which program sections readable and executable by a provisioning system are stored in order to execute all steps of the method for providing a result image data set and its aspects when the program sections are executed by the provisioning system; and / or on which program sections readable and executable by a training system are stored in order to execute all steps of the method for providing a trained generator function and / or a trained classifier function and its aspects when the program sections are executed by the training system.

[0097] In particular, the invention relates to a computer-readable storage medium on which program sections readable and executable by a provisioning system are stored in order to execute all steps of the method for providing a result image data set and its aspects when the program sections are executed by the provisioning system.

[0098] In particular, the invention relates to a computer-readable storage medium on which program sections readable and executable by a training system are stored in order to execute all steps of the method for providing a trained generator function and / or a trained classifier function and its aspects when the program sections are executed by the training system.

[0099] In an eighth aspect, the invention relates to a computer program or a computer-readable storage medium comprising a trained generator function and / or a trained classifier function, provided by a method for providing a trained generator function and / or a trained classifier function.

[0100] A largely software-based implementation has the advantage that existing deployment and training systems can be easily retrofitted via a software update to operate according to the invention. Such a computer program product may, in addition to the computer program itself, optionally include additional components such as documentation and / or additional components, as well as hardware components such as hardware keys (dongles, etc.) for using the software.

[0101] The properties, features, and advantages of this invention described above, as well as the manner in which they are achieved, will become clearer and more readily understandable in connection with the following description of the exemplary embodiments, which are explained in more detail in conjunction with the drawings. This description does not limit the invention to these exemplary embodiments. Identical components are designated with identical reference numerals in various figures. The figures are generally not to scale. They show: Fig. 1. An examination volume and various image datasets of the examination volume, Fig. 2 a first data flow diagram of the methods according to the invention and its aspects, Fig. 3 a second data flow diagram of the methods according to the invention and its aspects, Fig. 4 a third data flow diagram of the methods according to the invention and its aspects, Fig. 5 a fourth data flow diagram of the methods according to the invention and its aspects, Fig. 6 a first embodiment of the method for providing a result image data set, Fig. 7 a second embodiment of the method for providing a result image data set, Fig. 8 a third embodiment of the method for providing a result image data set, Fig. 9 a fourth embodiment of the method for providing a result image data set, Fig. 10 an embodiment of a method for providing a trained generator function and / or a trained classifier function, Fig. 11 a deployment system, Fig. 12 a training system, Fig. 13 an X-ray device.

[0102] Fig. Figure 1 shows an investigation volume VOL and various image data sets of the investigation volume VOL. In the illustrated embodiment, the investigation volume VOL is an artificial investigation volume, and the image data sets are simulations of an imaging of the artificial investigation volume.

[0103] The displayed image data sets of the examination volume VOL are characterized by two image parameters p1 and p2. The first image parameter p1 relates to the X-ray dose used during imaging, and the second image parameter p2 describes the X-ray device used during imaging (i.e., the X-ray source XSYS.SRC and the X-ray detector XSYS.DTC used during imaging).

[0104] Here, three different values ​​are shown for the first image parameter p1 (p 1,1 “ultra-low dose” or “very low dose”, p 1,2 “low dose” and p1,3 “medium dose” or “medium dose”), and three different values ​​are also shown for the second image parameter p2 (p 2,1 “System A”, p 2,2 “System B” and p 2,3 “System C”). For each of the nine combinations of the first image parameter p1 and the second image parameter p2, an image data set is presented that is described by this combination of the first image parameter p1 and the second image parameter p2.

[0105] For example, an input image data record ID is defined by the parameter combination (p 1,1 , p 2,1 ) described. However, if a trained image processing function or a doctor is adapted to a different parameter combination, for example the parameter combination (p 1,3 , p 2,2 ), so a result image data set OD can be generated using the invention, which is described by this parameter combination.

[0106] Fig. Figure 2 shows a first data flow diagram of the methods according to the invention and its aspects. In this first data flow diagram, the trained generator function GF is applied to two different input data sets. In the first case, the input data sets comprise the input image data set ID (denoted by the formula letter X) and the result image parameter OP (denoted by the formula letter p). Y ) and optionally the input image parameter IP (denoted by the formula letter p) X In this case, the result image data set OD (denoted by the formula letter Y) is calculated as Y = G(X, p Y , p X ) or as Y = G(X, p Y In the second case, the input data includes the result image data set OD, the input image parameter IP, and optionally the result image parameter OP. In this case, a comparison input image data set ID-S (designated by the formula letter X) is used. s ) calculated as X s = G(Y, p X , p Y) or as X s = G(Y, p X , p Y ). Overall, the comparison input image dataset ID-S is therefore X s = G(G(X, p Y , p X ), p X , p Y ) or as X s = G(G(X, p Y ) , p X ). Both the result image dataset OD and the comparison input image dataset ID-S are, in particular, synthetic image datasets.

[0107] The trained classifier function DF is also applied to two different input data sets in this first data flow diagram. In the first case, the trained classifier function DF is applied to a comparison image data set OD-C (denoted by the formula letter Y). c ) applied, and calculates a probability value 0 ≤ q(Y c ) ≤ 1 and optionally a classifier c(Y c), which corresponds to an estimated value for the image parameter of the comparison image dataset OD-C. The probability value q(Y c Here, corresponds to the probability estimated by the trained classifier function DF that the input value is a real image dataset. In the second case, the trained classifier function DF is applied to the resulting image dataset OD and calculates a probability value 0 ≤ q(Y) ≤ 1 and optionally a classifier c(Y). Thus, DF(Y) = q(Y) or DF(Y) = (q(Y), c(Y)). The comparison image dataset OD-C is, in particular, a real image dataset.

[0108] In this first data flow diagram, it should be noted that when providing the result image data set OD, only a part of the first data flow diagram is used, namely the determination of the result image data set OD as Y = G(X, p). Y , p X ) or as Y = G(X, p YThe other parts of the first data flow diagram are used specifically only during the training of the generator function GF or the training of the classifier function DF.

[0109] In the following, B denotes the space of image data sets, and P denotes the space of parameters. For example, for two-dimensional image data sets, the space of image data sets can be B = R m·n be (where R denotes the real numbers, and m and n the number of pixels and voxels respectively for each of the dimensions), for three-dimensional image datasets the space of image datasets can be B = R i·j·k be (where i, j, and k denote the number of pixels or voxels for each dimension, respectively). If the image parameters correspond to a signal-to-noise ratio or an X-ray dose, the space of image parameters can in particular be P = R + If the image parameters correspond to a used X-ray detector XSYS.DTC, the space of image parameters can be the quantity M DThis can be any designation of X-ray detectors. If the image parameters include both a signal-to-noise ratio and a used X-ray detector XSYS.DTC, the space of image parameters can in particular be P = R + × M D be.

[0110] Therefore, the trained generator function GF is a function GF: B × P → B or GF: B × P 2 → B, and the trained classifier function DF is a function DF: B → [0, 1] or DF: B → [0, 1] × P.

[0111] Fig. Figure 3 shows a second data flow diagram of the methods according to the invention and its aspects. The second data flow diagram includes elements of the first data flow diagram for determining the result image data set OD.

[0112] Unlike the first data flow diagram, the input image parameter IP and the output image parameter OP are determined computationally. The trained classifier function DF can be used for this purpose; alternatively, the input image parameter IP and the output image parameter OP can also be determined computationally.

[0113] The input image parameter IP is determined here based on the input image data set ID, for example by p x = c(X), with (q(X), c(X)) = DF(X). The result image parameter OP is determined here based on a comparison image data set CD (which is described by the formula letter Z), for example by p Y = c(Z), with (q(Z), c(Z)) = DF(Z). In particular, in this case, the determination and provision of the result image data set (OD) only takes place if the result image parameter OP and the input image parameter IP differ from each other, or if the deviation exceeds a given threshold.

[0114] Fig. Figure 4 shows a third data flow diagram of the methods according to the invention and its aspects. The third data flow diagram includes elements of the first data flow diagram.

[0115] In contrast to the first data flow diagram, the trained generator function GF here is a function GF: B 2 × P → B 2 or GF: B 2 × P 2 → B 2In particular, the trained generator function GF represents an input image data set ID, a mask image data set MD (described by the formula letter M), a result image parameter OP, and optionally an input image parameter IP. Here, the input image data set ID is specifically a fill image data set. The input image data set ID and the mask image data set have the same dimensionality and the same spatial extent. The trained generator function GF generates a modified mask image data set (denoted by the formula letter M') and a modified fill image data set (denoted by the formula letter X'), where at least the modified fill image data set is described by the result image parameter OP. Therefore, (X', M') = GF(X, M, p). Y , p X ) or (X', M') = GF(X, M, p Y , p X). The resulting image data set OD, which in this case is a difference image data set, can then be calculated, for example, by Y = X' - M'.

[0116] In particular, the mask image data set MD and the input image data set ID (corresponding to a fill image data set) may not be registered, but by applying the trained generator function GF the modified mask image data set and the modified fill image data set can then be registered.

[0117] In a second application, the generator function GF creates a comparison fill image data set in the third data diagram (denoted by the formula letter X). s ) and your comparison mask image data set (designated by the formula letter M) s ) as (X s , M s ) = GF(X', M', p X , p Y ) = GF(GF(X, M, p Y , p X ), p X , p Y ) or (X s , M s ) = GF(X', M', pX ) = GF(GF(X, M, p Y ), px) .

[0118] The third data flow diagram can also include the additional elements of the second data flow diagram.

[0119] Fig. Figure 5 shows a fourth data flow diagram of the methods according to the invention and its aspects. In contrast to the first, second, and third data flow diagrams, in this fourth data flow diagram the trained generator function GF is not applied directly to the input image data set ID, but to an input frequency data set IFD.

[0120] Here, a frequency representation F(ID) of the input image data set ID is generated using a transformation function TF (for example, a Fourier transform or a wavelet decomposition). The frequency representation F(ID) of the output data set comprises several frequency components F X (1) , ..., F X (n), where the multiple frequency components are ordered in ascending order according to the superscript index. In this embodiment, the input frequency data set IFD corresponds to the first frequency component F. X (1) regarding the lowest frequency. The resulting frequency data set OFD is then determined by applying the trained generator function GF to the input frequency data set IFD, here by F. Y (1) = GF(F X (1) , p Y , p X ) or F Y (1) = GF(F X (1) , p Y ). The other frequency components of the frequency representation F(OD) of the result image data set OD can then be transformed by another frequency function FF, for example as (F Y (2) , ..., F Y (n) ) = FF(F X (2) , ..., F X (n)The actual result image data set OD is then generated by applying the inverse transformation function TF. -1 determined. Therefore, in summary, Y = TF -1 (GF(F X (1) , p Y , p X ), FF(F X (2) , ..., F X (n) )) = TF -1 (GF(TF(X) (1) , p Y , px), FF(TF(X) (2) , ..., TF(X) (n) )) or Y = TF -1 (GF(F X (1) , p Y ), FF(F X (2) , ..., F X (n) )) = TF -1 (GF(TF(X) (1) , p Y ) , FF(TF(X) (2) , ..., TF(X) (n) )) .

[0121] Fig. Figure 6 shows a first embodiment of the method for providing an OD result image data set. In particular, this first embodiment implements the method described in Figure 6. Fig. 2. First data flow diagram shown.

[0122] The first step of the illustrated first embodiment is receiving the REC-ID of an input image data set ID of a first examination volume VOL via an interface IF. In particular, the input image data set ID can be acquired beforehand using an X-ray device.

[0123] In this embodiment, the input image data set ID is a two-dimensional image data set with dimensions of 512x512 pixels. Specifically, the input image data set ID is in DICOM format (an acronym for "Digital Imaging and Communications in Medicine"). Alternatively, the input image data set ID can have a different dimension (especially three-dimensional or four-dimensional), a different size, and / or a different format.

[0124] The next step in the first embodiment described is receiving or determining a result image parameter OP (REC-DET-OP). In this embodiment, the result image parameter OP is received via the IF interface.

[0125] In this embodiment, the result image parameter OP comprises a signal-to-noise ratio and a type designation of an X-ray detector XSYS.DTC. The described method is intended to determine a result image data set OD, which is described by the signal-to-noise ratio and is similar to an image acquired using the specified X-ray detector XSYS.DTC. In other words, the result image parameter OP relates to the signal-to-noise ratio of the result image data set OD and an X-ray detector XSYS.DTC associated with the result image data set OD—that is, properties of the result image data set OD.

[0126] The result image parameter OP is, in particular, a pair comprising a number (corresponding to the signal-to-noise ratio) and a string (corresponding to the type of the X-ray detector XSYS.DTC), and is therefore, in particular, a 2-tuple or a two-dimensional vector. Specifically, a result image parameter OP, given by (3.0, “RD 1”), can comprise the signal-to-noise ratio 3.0 and the type designation “RD 1”.

[0127] Alternatively, the result image parameter OP can also comprise only a signal-to-noise ratio. Alternatively, the result image parameter OP can also comprise only a type designation of an X-ray detector XSYS.DTC. Alternatively, the result image parameter OP can also comprise an X-ray dose of the result image data set OD, a noise level of the result image data set OD, or an X-ray source XSYS.SRC corresponding to the result image data set OD, or a combination of these and / or other properties of the result image data set OD. In general, the result image parameter OP can therefore be represented by an n-tuple or by an n-dimensional vector, where in particular n = 1 or n > 1.

[0128] The next step of the first embodiment is to determine the DET-OD of a result image data set OD of the first investigation volume VOL by applying a trained generator function GF to input data, where the input data is based on the input image data set ID and the result image parameter OP. The determination of the DET-OD of the result image data set OD is performed, in particular, using a processing unit CU. Here, one parameter of the trained generator function GF is based on a GA algorithm.

[0129] In this first embodiment, the input data comprises the input image data set ID and the result image parameter OP. Specifically, the result image data set OD is generated by directly applying the trained generator function GF to the input image data set ID and the result image parameter OP. In particular, the input data includes no further data besides the input image data set ID and the result image parameter OP. Specifically, the result image data set OD can be determined by Y = GF(X, p). Y ) , wherein the in Fig. The notation introduced in 2 is used.

[0130] Alternatively, the input image data set ID can also be processed section by section. In this example, the input image data set ID, which is two-dimensional and has dimensions of 512x512 pixels, can be divided into 16 sections, each containing two-dimensional image data or image data sets with dimensions of 128x128 pixels. Other divisions into sections are also possible, for example, 64 sections with dimensions of 64x64 pixels. It is also possible for the different sections to overlap at their edges. Division into different sections is also possible with a three-dimensional input image data set ID and a four-dimensional input image data set ID.

[0131] If processing is performed in sections, the trained generator function GF can be applied multiple times. Specifically, the trained generator function GF can be applied exactly once to each section. Applying the trained generator function GF generates sections of the result image data set OD, for example, as Y. i = GF(X i , p Y ) , where X i one of the sections of the input image data set ID describes, and where Y i describes the corresponding section of the result image data set OD. In particular, the procedure in this case may include an additional step that involves combining or assembling the result image data set OD based on the sections of the result image data set OD.

[0132] Furthermore, in this first embodiment, a parameter of the trained generator function GF is based on a GA algorithm, in that the trained generator function GF was trained together with a trained classifier function DF. In particular, the trained generator function GF is the result of a method according to the invention for providing a trained generator function GF and / or a trained classifier function DF.

[0133] The trained generator function GF is in particular a neural network, especially a convolutional neural network (an English technical term is "convolutional neural network") or a network comprising a convolution layer (an English technical term is "convolution layer").

[0134] The trained classifier function DF is also, in particular, a neural network, specifically a convolutional neural network or a network comprising a convolution layer.

[0135] The final step of the illustrated embodiment is the provision of the PROV-OD result image data set OD. In this first embodiment, the result image data set OD is transmitted to a requesting unit via the interface. Alternatively, the result image data set OD can also be displayed or saved.

[0136] Fig. Figure 7 shows a second embodiment of the method for providing an OD result image data set. This second embodiment also implements the method described in the Fig. 2. First data flow diagram shown.

[0137] In addition to the already mentioned regarding the Fig. The second embodiment further comprises the determination and / or reception of an input image parameter IP, in addition to the steps described in Section 6 and the first embodiment (which may include, in particular, all advantageous configurations and further developments described therein). This additional step can be performed, in particular, using the processing unit CU and / or the interface IF.

[0138] The input image parameter IP describes a property of the input image data record ID. In particular, the input image parameter IP and the result parameter OP have the same structure, i.e., the input image parameter IP describes the same properties of an image data record (with respect to the input image data record ID) as the result image parameter OP describes (with respect to the result image data record OD).

[0139] In this second embodiment, the input image parameter IP specifically includes the signal-to-noise ratio of the input image data set ID and the type of X-ray detector XSYS.DTC used to acquire the input image data set ID. Specifically, the signal-to-noise ratio of the input image parameter IP differs from the signal-to-noise ratio of the result image parameter OP, and / or the type of X-ray detector XSYS.DTC of the input image parameter IP differs from the type of X-ray detector XSYS.DTC of the result image parameter OP.

[0140] In this second embodiment, the input image parameter IP is received via the IF interface. Alternatively, the input image parameter IP can be determined based on the input image data record ID. This can be achieved, in particular, by applying a trained classifier function DF to the input image data record ID. Alternatively, other methods of determining the input image parameter IP are also possible; for example, the input image parameter IP can be determined based on the metadata of the input image data record ID.

[0141] In particular, one parameter of the trained classifier function DF is based on the GA algorithm, on which the parameter of the trained generator function GF is also based. Specifically, the trained classifier function DF and the trained generator function GF were trained together. Specifically, the trained classifier function DF was provided using a method according to the invention for providing a trained generator function GF and / or a trained classifier function DF.

[0142] The second embodiment includes, as an optional step, the adjustment of an imaging parameter of an imaging unit XSYS by ADJ-PRM based on a comparison of the input image parameter IP and the comparison image parameter CP. In this example, the X-ray dose of the X-ray device is increased or decreased if the deviation between the signal-to-noise ratio of the input image parameter IP and the signal-to-noise ratio of the comparison image parameter CP exceeds a predefined value. Specifically, the X-ray dose of the X-ray device XSYS can be increased if the signal-to-noise ratio of the input image parameter IP is lower than the signal-to-noise ratio of the comparison image parameter CP, and if the deviation or the magnitude of the deviation exceeds the predefined value.In particular, the X-ray dose of the XSYS X-ray device can be reduced if the signal-to-noise ratio of the input image parameter IP is greater than the signal-to-noise ratio of the comparison image parameter CP, and if the deviation or the amount of the deviation exceeds the specified value.

[0143] Fig. Figure 8 shows a third embodiment of the method for providing an OD result image data set. This third embodiment implements the method described in the Fig. 3. Second data flow diagram shown.

[0144] The third embodiment includes the aspects relating to the Fig. 6 and the Fig. 7 or the steps described with regard to the first and second embodiments, which may in particular include all the advantageous training and further development described therein.

[0145] The third embodiment further includes receiving REC-CD of a comparison image data set OD-C from a second examination volume, wherein the first examination volume VOL and the second examination volume overlap and / or are identical. Receiving REC-CD of the comparison image data set OD-C is performed in particular via the IF interface.

[0146] The comparison image dataset OD-C can, in particular, have the same dimensionality as the input image dataset ID. Furthermore, the comparison image dataset OD-C can, in particular, have the same extent, measured in voxels or pixels, as the input image dataset ID with respect to each of the dimensions. However, it is also possible that the input image dataset ID and the comparison image dataset OD-C have different extents.

[0147] In this embodiment, the comparison image dataset OD-C is also a two-dimensional image dataset with dimensions of 512x512 pixels. Furthermore, in this embodiment, the first examination volume VOL and the second examination volume are identical.

[0148] A further step of the third embodiment is the determination of a comparison image parameter CP by applying the trained classifier function DF to the comparison image dataset OD-C, wherein the comparison image parameter CP relates to a property of the comparison image dataset OD-C, and wherein a parameter of the trained classifier function DF is based on the GA algorithm. The determination of the comparison image parameter OD-C is performed in particular using the processing unit CU.

[0149] A further step of the third embodiment is the comparison of the input image parameter IP and the comparison image parameter CP, in particular by means of the computing unit CU.

[0150] In this embodiment, the determination of DET-OD of the result image data set OD is only performed if the comparison image parameter CP differs from the input image parameter IP. In particular, a deviation can occur if the difference between the comparison image parameter CP and the input image parameter IP exceeds a predefined threshold. If the comparison image parameter CP does not differ from the input image parameter IP, the process ends without determining DET-OD of the result image data set OD. When determining DET-OD of the result image data set OD, the result image parameter OP is selected or set such that it corresponds to the comparison image parameter CP.

[0151] Fig. Figure 9 shows a fourth embodiment of the method for providing an OD result image data set. The third embodiment implements the method described in the Fig. 5. Fourth data flow diagram shown.

[0152] The third embodiment includes the aspects relating to the Fig. 6 and the Fig. 7 or the steps described with regard to the first and second embodiments, which may in particular include all the advantageous training and further development described therein.

[0153] The fourth embodiment further includes determining the DET-IFD of an input frequency data set IFD based on the input image data set ID. Here, the input frequency data set IFD is a representation of the input image data set ID in the frequency domain (in particular, a component of the representation of the input image data set ID in the frequency domain). In this embodiment, the input frequency image data set IFD is a component of the wavelet decomposition TF of the input image data set ID; alternatively, a Fourier transform or a Laplace transform can also be used, for example.

[0154] Applying wavelet decomposition TF to the input image data set ID generates a frequency representation F(ID) of the input image data set ID. The frequency representation F(ID) of the input image data set ID comprises several frequency components F X (1) , ..., F X (n), where the multiple frequency components are ordered in ascending order according to the superscript index. In this embodiment, the input frequency data set IFD corresponds to the first frequency component F. X (1) regarding the lowest frequency.

[0155] In this embodiment, the input data is still based on the input frequency data set IFD. By applying the trained generator function GF to the input data, a result frequency data set OFD is generated, where the result frequency data set OFD is a representation of the result image data set OD (in particular, a component of the representation of the result image data set OD in the frequency domain) in the frequency domain.

[0156] The other frequency components of the frequency representation F(OD) of the result image data set OD can then be determined by another frequency function FF based on the other frequency components of the frequency representation F(ID) of the input image data set ID. The result image data set ID can then be determined by an inverse wavelet decomposition TF. -1 The frequency representation F(OD) of the result image data set can be determined.

[0157] Fig. Figure 10 shows an embodiment of a method for providing a trained generator function GF and / or a trained classifier function DF.

[0158] The first step of the illustrated embodiment is the reception of an input image data set ID and a comparison image data set OD-C of a first investigation volume VOL. The reception of TREC-ID-OD-C is performed, in particular, using a training interface TIF. In the illustrated embodiment, both the comparison image data set OD-C and the input image data set ID are two-dimensional image data sets, where the extent of the input image data set ID, measured in pixels, corresponds to the extent of the comparison image data set OD-C in each dimension. Alternatively, both the comparison image data set OD-C and the input image data set ID can be three-dimensional or four-dimensional image data sets; furthermore, the extent of the comparison image data set OD-C and the input image data set ID can also differ.

[0159] The comparison image data set OD-C will also be referred to by the formula letter Y in the following text. c The input image data set ID is designated by the formula letter X.

[0160] A further step of the illustrated embodiment is the reception of a result image parameter OP via TREC-OP, where the result image parameter OP relates to a property of the comparison image data set OD-C. The reception of the result image parameter OP via TREC-OP is performed, in particular, using the training interface TIF. In this embodiment, the result image parameter OP relates to the signal-to-noise ratio of the comparison image data set OD-C and the X-ray detector XSYS.DTC used during the acquisition of the comparison image data set OD-C. The result image parameter OP is subsequently denoted by the formula letter p. Y designated.

[0161] A further step in the illustrated embodiment is the determination of the TDET-OD of a result image data set OD of the first investigation volume VOL by applying a trained generator function GF to input data, where the input data is based on the input image data set ID and the result image parameter OP. The determination of the TDET-OD of the result image data set OD is performed, in particular, using a training processing unit TCU.

[0162] In the illustrated embodiment, the result image data set OD is also a two-dimensional image data set. The extent of the result image data set OD, measured in pixels, is identical with respect to each dimension to the extent of the input image data set ID, also measured in pixels. Alternatively, if the input image data set ID is a three-dimensional image data set, then the result image data set OD is also a three-dimensional image data set. Alternatively, if the input image data set ID is a four-dimensional image data set, then the result image data set OD is also a four-dimensional image data set.

[0163] The result image data set OD (hereinafter also described by the formula letter Y) is defined in this example as Y = GF(X, p y ) calculated. Alternatively, an input image parameter IP (denoted by the formula letter p) can still be used in the procedure. x) can be received or determined, where the input image parameter IP is a property of the input image record ID. In this case, the result image record OD can be determined by Y = GF(X, p). y , p x ) are determined.

[0164] A further step of the illustrated embodiment is the determination of a result classifier and a comparison classifier by applying a trained classifier function DF to the result image data set OD and the comparison image data set OD-C, in particular using the training processing unit TCU. The result classifier is subsequently denoted by the formula letter c(Y) and the comparison classifier by the formula letter c(Y). c ) described.

[0165] In the illustrated embodiment, the trained classifier function DF is a function DF: B → [0, 1] × P. The result classifier is obtained, in particular, by (q(Y), c(Y)) = DF(Y), where q(Y) corresponds to the probability determined by the trained classifier function DF that the result image data set OD is a real image data set. The comparison classifier is obtained, in particular, by (q(Y) c ), c(Y c )) = DF(Y c ), where q(Y c ) the probability determined by the trained classifier function DF corresponds to the fact that the comparison image dataset OD-C is a real image dataset.

[0166] A further step of the illustrated embodiment is the adjustment TADJ of a parameter of the trained generator function GF and / or the trained classifier function DF based on a comparison of the result classifier and the comparison classifier, in particular by means of the training computing unit TCU.

[0167] The following section describes possible cost functions K. (GF) the trained generator function GF and possible cost functions K (DF) The trained classifier function DF is described. However, it is of course always possible to use other cost functions. Here, the cost functions are described in the context of either the trained generator function GF or the trained classifier function DF. This should be understood to mean that elements of the cost functions can also be used for the other trained function. The individual components of the cost functions K (GF) , K(DF) They can be weighted differently, especially by using predetermined weights.

[0168] If the trained classifier function DF is a function DF: B → [0, 1], the cost function K can be (DF) The trained classifier function DF, for example, has an adversarial cost function K. (DF) A = - BCE(DF(Y c ), 1) - BCE(DF(Y), 0) = - BCE(q(Y c ),1) - BCE(q(Y), 0), where BCE denotes the binary cross-entropy (an English technical term is "binary cross-entropy") with BCE(z, z) = z'log(z) + (1-z')log(1-z). In particular, the adversarial cost function is given by K (DF) A = - log(DF(Y c )) - log (1 - DF(Y)) = - log(q(Y c )) - log(1 - q(Y)) = - log(q(Y c )) - log(1 - q(GF(X, p y))). When this cost function is minimized, the trained classifier function DF is trained to distinguish as well as possible between real image data (corresponding to the comparison image datasets OD-C) and synthetic image data generated by the trained generator function GF (corresponding to the result image datasets OD).

[0169] For the cost function K (GF) The trained generator function GF can in particular also be an adversarial cost function K. (GF) A = - BCE(DF(Y), 1) = - log(q(Y)) = - log(q(GF(X, p y ))) are used. When this cost function is minimized, the trained generator function GF is trained to produce such result image data sets OD that are incorrectly classified as real image data by the trained classifier function DF.

[0170] If the trained classifier function DF is a function DF: B → [0, 1] × P, then the cost function K can be (DF) in addition to the already defined adversarial cost function K (DF) A a further share (classification cost function K) P ) exhibiting deviations in image parameters. For example, it is possible to define a classification cost function K. P to use with K P = |c(Y) - c(Y c )| n = |c(GF(X,p y )) - c(Y c ) | n or K P = |c(Y) - p y | n = c(GF(X, p y )) - p y | n , where |A| n denotes the n-norm of A, in particular a 1-norm or a 2-norm. The classification cost function K P can also be found in the cost function K (GF)The trained generator function GF can be used. In particular, if an image parameter can be measured directly based on an image data set, the classifier c(Y) can be determined either by applying the trained classifier function DF or by direct calculation.

[0171] If the result image parameter OP is an element from a set (e.g., the X-ray detector used XSYS.DTC), then p can Y defined as a unit vector whose dimension corresponds to the cardinality of the set and which can only take on elements 0 and 1. In this case, c(Y) and c(Y) are also c ) Vectors of the same dimension, either of the same form as p Y or with entries between 0 and 1, each corresponding to a probability. A norm |c(Y) - c(Y c ) | nIn this case, it can be understood as the sum of the norms of the individual components. This also applies if the result image parameter OP is an object with a higher dimension; in this case, too, a norm |c(Y) - c(Y) can be used. c )| n can be understood as the sum of the norms of the individual dimensions.

[0172] The cost function K (GF) The trained generator function GF can also be a cyclic consistency cost function K. CC and an information loss cost function K IL include the contribution K CC The cyclic consistency cost function is given by K CC = |X - X c | m = |X - GF(GF(X), p y ) p x ) | m , or K CC = |X - GF(GF(X), p y , p x ) p x , p y )| m , where |A| m The m-norm of A is denoted. In particular, m = 1 and m = 2 can be chosen. Since X and X cFor image datasets, the norm can be calculated pixel-wise or voxel-wise. Minimizing the cyclic consistency cost function ensures that applying the trained generator function GF twice with swapped image parameters closely resembles an identity mapping (and thus acts similarly to an autoencoder), resulting in image data generation with fewer errors for other parameters.

[0173] The contribution K IL The information loss cost function is calculated as KIL=∑i∈π∑s(i)∈σ(i)|〈GF(X,py)−X〉s(i)|1 where π is the set of pixels or voxels of X or GF(X, p). y ) denotes, where σ(i) denotes a set of neighborhoods of pixel i (for example, the 3x3 neighborhood, the 5x5 neighborhood and the 7x7 neighborhood in the two-dimensional case, or the 3x3x3 neighborhood, the 5x5x5 neighborhood and the 7x7x7 neighborhood in the three-dimensional case), and <GF(X, p y ) - X> s(i)the mean of the difference GF(X, p) y ) - X in the neighborhood s(i) of pixel i. Furthermore, |A| denotes l The l-norm of A, where m = 1 and m = 2 can be chosen in particular. The use of an information loss cost function is especially advantageous when the image noise is not independent of the image signal, and when simultaneously the input image data set ID is divided into several sub-areas, and the trained generator function GF is applied to each of these sub-areas, and the resulting image data set OD is composed from the results of the application to the sub-areas ("patch processing"). In this case, the information loss cost function of the trained generator function GF allows the signal-dependent component of the image noise to be better understood and simulated.

[0174] The entire cost function K to be minimized (GF)The trained generator function GF can therefore be given by K (GF) = µ (GF) A ·K (GF) A + µ (GF) P ·K P + µ cc · K cc + µ IL ·K IL , where µ is a weighting factor (which can also be 0). The entire cost function K to be minimized (DF) The trained classifier function DF can therefore be given by K (DF) = p (DF) A ·K (DF) A + µ (DF) P ·K P, where µ is a weighting factor (which can also be 0). The parameters of the trained generator function GF and the trained classifier function DF can be adjusted, for example, by backpropagation, as is known to those skilled in the art; other minimization methods are, of course, also applicable. If the trained generator function GF and the trained classifier function DF are neural networks, the respective cost function can be determined, in particular, by adjusting parameters of the trained generator function GF and / or the trained classifier function DF.

[0175] The final step of the illustrated embodiment is the provision of TPROV of the trained generator function GF and / or the trained classifier function DF. In this embodiment, both the trained generator function GF and the trained classifier function DF are stored. Alternatively or additionally, the trained generator function GF and the trained classifier function DF can also be transmitted to a receiver.

[0176] Fig. 11 shows a deployment system PRVS, Fig. Figure 12 shows a training system TRS. The depicted provisioning system PRVS is configured to execute a method according to the invention for providing a result image data set OD. The depicted training system TRS is configured to execute a method according to the invention for providing a trained generator function GF and / or a trained classifier function DF. The provisioning system PRVS comprises an interface IF, a processing unit CU, and a storage unit MU; the training system TRS comprises a training interface TIF, a training processing unit TCU, and a training storage unit TMU.

[0177] The deployment system PRVS and / or the training system TRS can be, in particular, a computer, a microcontroller, or an integrated circuit. Alternatively, the deployment system PRVS and / or the training system TRS can be a real or virtual cluster of computers (a technical term for a real cluster is "cluster," a technical term for a virtual cluster is "cloud"). The deployment system PRVS and / or the training system TRS can also be configured as a virtual system running on a real computer or a real or virtual cluster of computers (a technical term for this is "virtualization").

[0178] An interface (IF) and / or a training interface (TIF) can be a hardware or software interface (e.g., PCI bus, USB, or FireWire). A processing unit (CU) and / or a training processing unit (TCU) can consist of hardware or software elements, such as a microprocessor or an FPGA (Field Programmable Gate Array). A memory unit (MU) and / or a training memory unit (TMU) can be implemented as random access memory (RAM) or as permanent mass storage (hard drive, USB flash drive, SD card, solid-state drive).

[0179] The interface IF and / or the training interface TIF can, in particular, comprise multiple sub-interfaces that execute different steps of the respective procedures. In other words, the interface IF and / or the training interface TIF can also be understood as a plurality of interfaces IF or a plurality of training interfaces TIF, respectively. The processing unit CU and / or the training processing unit TCU can, in particular, comprise multiple sub-processing units that execute different steps of the respective procedures. In other words, the processing unit CU and / or the training processing unit TCU can also be understood as a plurality of processing units CU or a plurality of training processing units TCU, respectively.

[0180] Fig.Figure 13 shows an X-ray device XSYS connected to a delivery system PRVS. In the illustrated embodiment, the X-ray device XSYS is a C-arm X-ray unit XSYS. The C-arm X-ray unit XSYS comprises an X-ray source XSYS.SRC for emitting X-rays. Furthermore, the C-arm X-ray unit XSYS comprises an X-ray detector XSYS.DTC for receiving X-rays. The X-ray source XSYS.SRC and the X-ray detector XSYS.DTC are attached to the two opposite ends of the C-arm XSYS.ARM. The C-arm XSYS.ARM of the C-arm X-ray unit XSYS is mounted on a stand XSYS.STC. The stand XSYS.STC includes drive elements designed to change the position of the C-arm XSYS.ARM. In particular, the C-arm XSYS.ARM can be rotated about two different axes. The C-arm X-ray unit also includes a control and evaluation unit XSYS.CTRL and a patient positioning device XSYS.The patient tray (PAT) is a platform on which a patient can be positioned. Using the control and evaluation unit XSYS.CTRL, the position of the C-arm XSYS.ARM can be adjusted, the C-arm XSYS.ARM can be rotated around the examination volume VOL, and X-ray image data sets of the examination volume VOL can be acquired. Alternatively, the PRVS delivery system can be implemented as part of the control and evaluation unit XSYS.CTRL.

Claims

[1] Computer-implemented method for providing an output image data set (OD), comprising - Receiving (REC-ID) an input image data set (ID) of a first examination volume (VOL), wherein the input image data set (ID) is an X-ray image data set of the first examination volume (VOL), - Receiving or determining (REC-DET-IP) an input image parameter (IP), where the input image parameter (IP) relates to a property of the input image record (ID), - Receiving or determining (REC-DET-OP) a result image parameter (OP), - Determining (DET-OD) a result image data set (OD) of the first investigation volume (VOL) by applying a trained generator function (GF) to input data, where the input data is based on the input image data set (ID) and the result image parameter (OP), where the input data continues to be based on the input image parameter (IP), where the result image parameter relates to one or more of the following properties of the result image data set (OD): -- X-ray dose of the result image dataset (OD), -- Noise level of the result image data set (OD), - X-ray source (XSYS.SRC) and / or X-ray detector (XSYS.DTC) corresponding to the result image data set (OD), and wherein a parameter of the trained generator function (GF) is based on a generative adversarial algorithm, - Provisioning (PROV-OD) of the result image data set (OD). [2] Method according to claim 1, wherein the input image parameter (IP) relates to one or more of the following properties of the input image data set (ID): - X-ray dose of the input image data set (ID), - Noise level of the input image data set (ID), - X-ray source (XSYS.SRC) and / or X-ray detector (XSYS.DTC) corresponding to the input image data set (ID). [3] Method according to any of the foregoing claims, further comprising - Receiving (REC-CD) a comparison image data set (OD-C) of a second examination volume, where the first examination volume (VOL) and the second examination volume overlap and / or are identical, - Determining (DET-CP) a comparison image parameter (CP) by applying a trained classifier function (DF) to the comparison image dataset (OD-C), where the comparison image parameter (CP) relates to a property of the comparison image dataset (OD-C), and wherein one parameter of the trained classifier function (DF) is based on the generative adversarial algorithm, - Comparing (CMP-IP-CP) the input image parameter (IP) and the comparison image parameter (CP), wherein the determination of the result image data set (OD) is performed when the comparison image parameter (CP) differs from the input image parameter (IP), and wherein the result image parameter (OP) corresponds to the comparison image parameter (CP). [4] Method according to claim 3, further comprising: - Adjusting (ADJ-PRM) an imaging parameter of an imaging unit based on the comparison of the input image parameter (IP) and the comparison image parameter (CP). [5] Method according to any of the foregoing claims, further comprising: - Determining (DET-IFD) an input frequency data set based on the input image data set (ID), where the input frequency data set is a representation of the input image data set (ID) in the frequency domain, where the input data is based on the input frequency data set, wherein the application of the trained generator function (GF) to the input data produces a result frequency data set, where the result frequency data set is a representation of the result image data set (OD) in the frequency space. [6] Method according to one of the preceding claims, wherein the result image parameter (OP) is matched to a trained image processing function. [7] Method according to any of the foregoing claims, further comprising: - Receiving (REC-MD) a mask image dataset of a third examination volume, wherein the first examination volume (VOL) and the third examination volume overlap and / or are identical, with the input data still being based on the mask image dataset. [8] Method according to any of the preceding claims, wherein the parameter of the trained generator function (GF) is based on a cyclic consistency cost function and / or on an information loss cost function. [9] Method according to any of the preceding claims, wherein the input image data set (ID) comprises a temporal sequence of input image data of the first examination volume (VOL), and wherein the result image data set (OD) comprises a temporal sequence of result image data of the first examination volume (VOL). [10] Method for providing a trained generator function (GF) and / or a trained classifier function (DF), comprising: - Receiving (TREC-ID-OD-C) an input image dataset (ID) and a comparison image dataset (OD-C) of a first examination volume (VOL), wherein the input image dataset (ID) is an X-ray image dataset of the first examination volume (VOL), - Receiving or determining an input image parameter (IP), where the input image parameter (IP) relates to a property of the input image record (ID), - Receiving (TREC-OP) a result image parameter (OP), where the result image parameter relates to one or more of the following properties of the comparison image data set (OD-C): -- X-ray dose of the comparison image dataset (OD-C), -- Noise level of the comparison image data set (OD-C), -- X-ray source (XSYS.SRC) and / or X-ray detector (XSYS.DTC) corresponding to the comparison image data set (OD-C), - Determining (TDET-OD) a result image data set (OD) of the first examination volume (VOL) by applying the trained generator function (GF) to input data, wherein the input data is based on the input image data set (ID) and the result image parameter (OP), and furthermore, the input data is based on the input image parameter (IP). - Determining (TDET-CL) a result classifier and a comparison classifier by applying the trained classifier function (DF) to the result image dataset (OD) and the comparison image dataset (OD-C), - Adjusting (TADJ) a parameter of the trained generator function (GF) and / or the trained classifier function (DF) based on the result classifier and the comparison classifier, - Deploying (TPROV) the trained generator function (GF) and / or the trained classifier function (DF). [11] Provisioning system (PRVS) for providing a result image data set (OD), comprising an interface (IF) and a computing unit (CU), - wherein the interface (IF) is configured for receiving (REC-ID) an input image data set (ID) of a first examination volume (VOL), wherein the input image data set (ID) is an X-ray image data set of the first examination volume (VOL), - wherein the interface (IF) and / or the computing unit (CU) are configured to receive or determine (REC-DET-OP) an input image parameter (IP) and a result image parameter (OP), wherein the input image parameter (IP) relates to a property of the input image data set (ID), - wherein the computing unit (RU) is further trained to determine (DET-OD) a result image data set (OD) of the first investigation volume (VOL) by applying a trained generator function (GF) to input data, wherein the input data is based on the input image data set (ID) and the result image parameter (OP), wherein the input data is further based on the input image parameter (IP), and wherein the result image parameter relates to one or more of the following properties of the result image data set (OD): -- X-ray dose of the result image dataset (OD), -- Noise level of the result image data set (OD), -- X-ray source (XSYS.SRC) and / or X-ray detector (XSYS.DTC) corresponding to the result image data set (OD), and wherein a parameter of the trained generator function (GF) is based on a generative adversarial algorithm, - whereby the interface is still configured to provide (PROV-OD) the result image data set (OD). [12] X-ray device (XSYS) comprising an X-ray source (XSYS.SRC) and an X-ray detector (XSYS.DTC), further comprising a delivery system (PRVS) according to claim 11. [13] Computer program product comprising a computer program which can be directly loaded into a memory (MU) of a deployment system (PRVS), comprising program sections to execute all steps of the method according to any one of claims 1 to 9 when the program sections are executed by the deployment system (PRVS); and / or which can be directly loaded into a training memory (TMU) of a training system (TRS), comprising program sections to execute all steps of the method according to claim 10 when the program sections are executed by the training system (TRS). [14] Computer-readable storage medium on which program sections readable and executable by a deployment system (PRVS) are stored to execute all steps of the method according to any one of claims 1 to 9 when the program sections are executed by the deployment system (PRVS); and / or on which program sections readable and executable by a training system (TRS) are stored to execute all steps of the method according to claim 10 when the program sections are executed by the training system (TRS).

Images