Evaluation of characterization data of x-ray detector

Abstract

The invention relates to a computer-implemented method for supporting the evaluation of characterization data (31) of an X-ray detector for an X-ray imaging system, in particular for a computed tomography system, said X-ray detector having a plurality of detector modules, the method comprises the following steps: receiving characterization data (31) of a detector module of the X-ray detector, at least part of the characterization data (31) being based on measurement data recorded by the detector module without an examination object; applying a trained algorithm to the characterization data (31), in which composite image data (14) is generated as output, which simulates image data (13) of an X-ray imaging system, in particular of a computed tomography system, recorded with an X-ray detector; composite image data (14) is provided.

Description

Technical Field

[0001] Regardless of the grammatical gender of a particular term, people with both masculine and feminine identities are included.

[0002] The present invention relates to a computer-implemented method for supporting the evaluation of characterization data of an X-ray detector for an X-ray imaging system, a method for quality control when manufacturing an X-ray detector for an X-ray imaging system, a computer program product or a storage medium, a system and a method for training a trainable algorithm. Background Art

[0003] X-ray detectors, such as computed tomography (CT) detectors, are typically constructed from multiple subcomponents, particularly detector modules. The image quality achievable with a corresponding X-ray detector is dependent on the installed detector modules. It has been shown that, in addition to the properties of the individual detector modules, other factors, such as the combination of the detector modules, can also influence image quality.

[0004] Therefore, image inspection of X-ray detectors (hereinafter also referred to as "detectors") is important, for example, to identify regularly occurring artifacts or other detector-induced errors. However, achieving image inspection requires multiple, very time-consuming steps. Running corresponding test chains to test the detectors can take a relatively long time. Therefore, verifying image quality is associated with considerable costs and time. Furthermore, changes to the detector, such as by replacing one or more detector modules, can carry the risk that the image quality will subsequently deteriorate or no longer meet the desired or required quality standards.

[0005] For this reason, one approach might be to make a statement about the expected image quality as early as possible, for example based on a few scans. For example, characterizing measurement data could be recorded. These characterizing measurement data could be analyzed, for example, using defined limit values ​​to make a statement about whether image artifacts are expected. The limit values ​​could be determined, for example, by experts based on empirical values. This makes it possible to estimate image quality, particularly without observing or even creating real image data. However, in many cases, such an analysis is qualitative rather than quantitative in nature. Therefore, the accuracy or reliability of the actual image quality is often limited. Consequently, detectors of insufficient quality could be overlooked using this analysis, or detectors whose quality is actually good enough could be required. Summary of the Invention

[0006] SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a method by which an early assessment of the image quality of an X-ray detector can be achieved with improved reliability.

[0007] This object is achieved by the method according to the invention, the computer program product or storage medium according to the invention, and the system according to the invention. Further features and advantages are apparent from the description and the drawings.

[0008] Below, the solution according to the present invention is described with reference to the claimed systems, products, and methods. Furthermore, the solution according to the present invention is subsequently described with reference to the claimed systems, products, and methods for supporting the evaluation of characterization data or for quality control in the manufacture of X-ray detectors, as well as to systems, products, and methods for providing a trained algorithm. Features, advantages, or alternative embodiments described herein with reference to one aspect of the present invention can be similarly transferred to the other aspects, and vice versa. In other words, embodiments of the system and / or product according to the present invention can be improved by features described or claimed in the context of the corresponding method. Functional features of the method can be implemented by physical units of the system and / or product. Features, advantages, or alternative embodiments. Embodiments for providing a trained algorithm can be improved by features described or claimed in the context of the system, product, and method for supporting the evaluation of characterization data or quality control in the manufacture of X-ray detectors. In particular, the data set of the system, product, and method for supporting the evaluation of characterization data or quality control in the manufacture of X-ray detectors can have the same properties and characteristics as the corresponding data set used in the system, product, and method for providing a trained algorithm. The trained algorithms provided by the corresponding methods, products and systems can be used in systems, products and methods for supporting the evaluation of characterization data or for quality control in the manufacture of X-ray detectors.

[0009] According to a first aspect of the present invention, a computer-implemented method is provided for supporting the evaluation of characterization data of an X-ray detector for an X-ray imaging system, in particular a computed tomography system, the X-ray detector having a plurality of detector modules. The method comprises the following steps:

[0010] receiving characterization data of a detector module of the X-ray detector, wherein at least a portion of the characterization data is based on measurement data of the detector module, which measurement data have been recorded without an examination object;

[0011] - applying the trained algorithm to the characterization data, wherein synthetic image data are generated as output, said synthetic image data simulating image data recorded by an X-ray imaging system, in particular a computed tomography system, using an X-ray detector;

[0012] -Provide composite image data.

[0013] Advantageously, the method according to the present invention can provide image data at a point in time during the X-ray detector construction process that precedes the point in time at which image data would typically be obtained via direct measurement. In particular, synthetic image data can typically be obtained significantly faster than image data generated via actual image measurement. Synthetic image data can be a particularly reliable estimate of actual image data. Synthetic image data therefore provides a more direct understanding of how the image quality of subsequently acquired image data will change. This allows for early and highly reliable information about image quality. Artifacts, such as ring artifacts or streak artifacts, can be easily identified in synthetic image data. Consequently, when errors are discovered, early action can be taken and corrections can be made. This allows for a better evaluation of the characterization data. The evaluation of the characterization data can, for example, include an assessment and / or evaluation of the characterization data. In the context of the present invention, the evaluation of the characterization data can, in particular, be an assessment of the quality of the X-ray detector or the quality of image data that can be acquired using the X-ray detector. In particular, the method according to the present invention can provide indicators of how changes to the X-ray detector, particularly those involving detector modules, will affect the detector early in the manufacturing and / or maintenance process using synthetic image data. On the one hand, this can speed up the production process. However, the method can also be used advantageously, for example, when replacing detector modules (for example, during maintenance) and when replacement is planned.

[0014] The term "characterization data" should be understood broadly within the scope of the present invention. Characterization data generally refers to data that can be used to characterize an X-ray detector and / or individual detector modules of an X-ray detector. The characterization data may be based on measurement data of the detector modules. The characterization data may, for example, be raw data, in particular corresponding to directly recorded measurement data of the detector modules. The characterization data may include further processed measurement data. The characterization data may include data not directly recorded by the detector modules. For example, the characterization data may include the temperature of the detector modules detected by a temperature sensor and / or the temperature of the surrounding environment of the detector modules. Preferably, the characterization data is recorded data where the X-ray imaging system is not yet fully configured and / or where image data of the examination object are not recorded using the X-ray detector. The examination object may also be referred to as a measurement object. The examination object may, for example, be a person, an animal, or a part of a person or an animal. The examination object may, for example, be an object, such as a piece of luggage during baggage inspection. The image data may, for example, be an image, in particular an image of the examination object. The image data may be raw data, in particular raw data of the examination object, which can be used within the scope of an image reconstruction method to create an image, in particular an image of the examination object.

[0015] The term "X-ray detector" should be understood broadly within the scope of the present invention. Generally speaking, an X-ray detector refers to an X-ray detector for detecting X-rays. An X-ray detector is provided for use in an X-ray imaging system. In particular, an X-ray detector is provided for use in a computed tomography system. An X-ray detector comprises a plurality of detector modules. The X-ray detector may be configured to convert X-ray radiation into electrical signals using the detector modules. The plurality of detector modules may preferably be arranged in a matrix. For example, an X-ray detector may include 10-200, preferably 20-100, detector modules. The detector modules of an X-ray detector may be of the same type or type. However, it has been shown that in practice, detector modules of the same type often have at least minor differences. Such differences may lead to artifacts during imaging. It is not always easy to predict whether such artifacts will occur and to what extent. It has been shown that, on the one hand, detector modules that are easily distinguishable in practice may result in artifacts, such as ring artifacts, while, on the other hand, detector modules that are inherently defective may still produce good image data. Advantageously, the synthetic image data obtained using the method according to the present invention can better predict the occurrence of artifacts.

[0016] The method includes the step of receiving characterization data for detector modules of an X-ray detector. The characterization data can be received, for example, via an interface. For example, the characterization data can be retrieved from a database. The characterization data can be retrieved locally and / or from a network and / or remote connection, such as via the Internet. The characterization data can be input by a user into corresponding processing software designed to perform the method according to the present invention. Alternatively, the method according to the present invention can be performed on a computing unit of an X-ray imaging system. For example, the characterization data can be generated by the X-ray imaging system itself and forwarded to the computing unit. The characterization data can be generated when the detector modules of the X-ray detector are already installed together in the X-ray detector. Alternatively, the characterization data for the detector modules can be generated in whole or in part when the detector modules are not yet installed together in the X-ray detector and / or when the detector modules are installed in a configuration different from their current assigned arrangement. The characterization data is particularly data suitable for characterizing at least some of the properties of the detector modules and / or the X-ray detector, individually and / or in their entirety. At least some of the characterization data is based on measurement data recorded by the detector modules without an examination object. Preferably, all characterization data can be based on data that was not recorded in the presence of an examination object during a method corresponding to the normal, set operation of the X-ray detector. Measurement data recorded by the detector module without an examination object can, for example, be or include air jet data of the detector module. Measurement data recorded by the detector module without an examination object can, for example, be data recorded in the X-ray radiation of the detector module. The characterization data can be provided, for example, as vectors or matrices. For example, different detector modules can be coded numerically.

[0017] A trained algorithm is applied to the characterization data. The trained algorithm can be based on machine learning in particular. The term "trained algorithm" can include different aspects of machine learning in particular. A trained algorithm can also be referred to as a trained function. In particular, the algorithm can adapt to new situations and recognize and extrapolate patterns based on training data. In general, the parameters of the algorithm can be adapted through training. This training can include, for example, supervised learning, semi-supervised learning, active learning, self-supervised learning, unsupervised learning, and / or reinforcement learning. In particular, the parameters of the algorithm can be adapted iteratively through multiple training steps. In particular, during training, a specific loss function can be optimized, in particular minimized. For example, the trained algorithm can include an artificial neural network, a support vector machine, a decision tree (in particular, a random forest), and / or a Bayesian network. Additionally or alternatively, the algorithm can be based on a k-means algorithm, Q learning, an evolutionary algorithm, a Monte Carlo tree search, and / or association analysis. In particular, within the scope of neural network training, a backpropagation algorithm (error feedback algorithm) can be used. The neural network can be, in particular, a deep neural network (DNN), a convolutional neural network (CNN), or a convolutional deep neural network. The trained algorithm preferably comprises a generative model.The trained algorithm is configured or trained to output synthetic image data.

[0018] The term "synthetic image data" is to be understood in a broad sense within the scope of the present invention. Image data generally describe data from which at least one visual medium, in particular a two-dimensional or three-dimensional image, can be generated. Synthetic image data are in particular image data that are completely or partially artificially generated. In particular, the synthetic image data can be generated by an algorithm. Within the scope of the present invention, the synthetic image data correspond to image data recorded by an X-ray imaging system using an X-ray detector. In particular, the synthetic image data can correspond to image data recorded by a computed tomography system using an X-ray detector. Compared to real image data that can be obtained using corresponding measurements, synthetic image data can be obtained significantly faster using the method according to the present invention. Within the scope of the present invention, it is recognized that characterization data that can be generated already at an early stage during the production of the X-ray detector can be decisive for artifacts that appear in the image that is actually generated later. Advantageously, the present invention utilizes this situation by generating synthetic image data based on the characterization data.

[0019] The synthetic image data is provided. Within the scope of the present invention, this provision is generally to be understood in a broad sense. For example, the output can be configured, in particular for the user, via an output medium. The output medium can be, for example, a screen, a projector, or a printer. This provision can include, for example, output for further processing, such as forwarding to another program and / or an external device. This provision can also include, for example, storage on an external or internal data carrier.

[0020] According to an embodiment, the characterization data can be assigned to the arrangement of detector modules in the X-ray detector. For example, the characterization data can be set separately for each detector module, wherein the characterization data of each detector module are assigned to a position in the arrangement of the detector modules. For example, the assignment can be set by coordinate data. For example, the assignment can be set by the order of the characterization data. For example, the first characterization data in a set of characterization data can be assigned to the detector module in the first position in the X-ray detector, and the last characterization data in the set of characterization data can be assigned to the detector module in the last position in the X-ray detector. For example, the first position can be in the upper left corner of the X-ray detector and / or the last position can be in the lower right corner of the X-ray detector. Here, the characterization data can be set as a vector or a matrix, for example. For example, different detector modules can be coded by numbers.

[0021] According to an embodiment, the characterization data includes one or more of the following:

[0022] - the response of the detector pixels of the detector module to radiation in the absence of an examination object,

[0023] - the temporal noise characteristics of the individual detector pixels of the detector module,

[0024] - signal instabilities of the detector module caused by incident radiation,

[0025] - a list of detector pixels of the detector module that are marked as defective,

[0026] the expected influence of the orientation of the collimator, in particular of a tilted collimator, on the detection of the radiation signal by the detector module,

[0027] -Dependence of the detector response of the detector module on the thermal influence parameters.

[0028] The response of the detector pixels of a detector module to radiation in the absence of an object to be examined can, for example, be the response of the detector pixels to radiation from an air scan. In other words, preferably defined X-ray radiation can be directed at the detector module or individual detector modules, and it can be detected which signal the detector module records based on this. The X-ray radiation can be defined in particular with respect to its intensity and / or frequency. For example, a specific X-ray spectrum with a specific intensity can be directed at the detector module. Detecting the detector response to radiation can be a good indicator of how the corresponding detector module functions, without having to perform a complete measurement. Therefore, this detection of the detector response can be performed particularly early when constructing the X-ray detector or the entire X-ray imaging system. By means of the method according to the invention, relatively reliable statements about the performance of the entire X-ray detector can still be made in advance with the help of the responses of the detector modules.

[0029] It may happen that a detector module has a changing noise characteristic over time. This may also have an impact on the subsequent performance of the X-ray detector. By detecting the noise characteristic over time, this change over time can also be taken into account within the scope of the method according to the present invention.

[0030] Detector modules can react differently to prolonged incident radiation. In particular, for example, in the case of prolonged high-dose incident X-ray radiation, signal instabilities may occur during detection by the detector module or by individual detector modules. Radiation-induced signal instabilities can be indicated, for example, by the temporal course of the measurement signal via the corresponding detector module, in particular in the case of high-dose radiation. Different courses of the measurement signal can represent different degrees of radiation-induced signal instability. For example, a uniform, in particular high-dose, radiation signal can be recorded, and deviations from a constant course represent signal instabilities.

[0031] It has been shown that individual defective detector pixels can have varying degrees of impact on the quality of an X-ray detector. For example, an X-ray detector can, in some cases, still function perfectly well even when individual detector pixels are defective. However, the image quality of the X-ray detector may vary depending on the relative position of the defective detector pixels, in particular the number of defective detector pixels. The list of detector pixels marked as defective can each include a measure of the severity and / or type of the defect. For example, a detector pixel can be marked as defective if its signal response differs significantly from the signal responses of other surrounding detector pixels within defined boundaries. A measure of the severity of the defect can, for example, be the relative deviation of the signal response from the average value of the signal responses of the surrounding detector pixels. The method according to the present invention has been shown to provide a good estimate of the impact of individual defective detector pixels on the overall image quality.

[0032] Tilted collimators can, for example, lead to additional scattering effects. For example, a shift in the focus of the X-ray tube can result in the collimator being projected as a shadow. It may be helpful to estimate the scattering effects and other effects of the tilted collimators of the individual detector modules. A collimator is tilted in particular when it is not oriented precisely to the X-ray source, in particular the focus of the X-ray tube. When mounting the collimator on the detector module, for example by gluing and / or screwing, and when screwing the detector module into the detector mechanism, tolerances may generally exist which ensure that the orientation is not precisely matched. If the collimator is unfavorable, the reaction to scattered radiation may change. This can, for example, lead to brightness differences in the scan in relation to the projection and the signal, as a result of which artifacts may appear in the image during image reconstruction.

[0033] The detector modules can react differently to thermal influences. The thermal influencing variable can be, for example, the temperature and / or conditions influencing the temperature. The thermal influencing variable can be, for example, the temperature of the respective detector module and / or the temperature of the environment of the respective detector module. The thermal influencing variable can be, for example, the rotational speed of a fan for cooling the X-ray detector and / or the respective detector module. The thermal influencing variable can be, for example, the dose of incident X-ray radiation. A large amount of incident X-ray radiation can lead to an increase in the temperature of the detector module. It has been shown that the different reactions of different detector modules to thermal influences or to thermal influencing variables can have an impact on the image quality of the X-ray detector. Advantageously, with the method according to the invention, by using corresponding characterization data concerning the correlation of the detector response of the detector module with the thermal influencing variable, a prediction of the impact on the image quality can be made with good accuracy.

[0034] According to embodiments, the trained algorithm comprises a trained generative artificial intelligence. In particular, the generative artificial intelligence may include a diffusion model with at least one denoising block, which is used to generate synthetic image data. Diffusion models are typically generative probabilistic models that can be used to generate new data, particularly image data. Diffusion models are typically based on adding noise to training data, particularly image data used for training, during training and then removing it again. The noise may be, for example, Gaussian noise. However, other types of noise are also conceivable. After training, the diffusion model can be used to generate new synthetic data, particularly synthetic image data, from random noise. During training of the diffusion model, noise is added to a reference image in a forward or diffusion process such that continuous noise is added to the reference image. This results in a continuous noise addition to the reference image. Typically, the result of this diffusion process is a completely noisy distribution (corresponding to a completely noisy image). Following the diffusion process, an artificial neural network is trained in an inverse process to gradually remove the noise, thereby generating denoised image data corresponding to the reference image, thereby undoing the noise addition caused by the diffusion process. The inverse process or the training of the inverse process is also carried out step by step, in particular by training the neural network for each diffusion step to undo the corresponding diffusion step. Finally, the individual steps can be combined so that noise can be removed from the image. Thus, the neural network is trained to generate image data from noise. During training, the inverse process is approximated by adapting the trainable parameters of the neural network. The inverse process can include one or more denoising blocks. The denoising block can be adapted in particular for performing denoising. Optionally, multiple denoising blocks can be provided. The denoising block can be adapted in particular for sequential application to the data to be denoised. This can produce overall even better results if necessary. One or more denoising blocks can preferably each include an artificial neural network, in particular a convolutional neural network. For example, one or more denoising blocks can be based on a U-Net structure. The concept of U-Net is described in particular in "U-Net: Convolutional Networks for Biomedical Image Segmentation" arXiv:1505.04597 [cs.CV], Ronneberger O, Fischer P, Brox T, 2015, and can also be applied analogously in the context of the present invention. For example, it can be provided that the resolution is halved in the U-Net until a low final resolution of, for example, 2×2 is achieved. For example, two ResNet blocks, which in particular contain attention heads, can be provided for each resolution. The concept can in particular be token-based, wherein, for example, 32 dimensions can be provided for each token. ResNet ("residual neural network") is in particular a model based on deep learning, in which a weight layer learns a residual function related to the input of the layer.The "layers" of a model or neural network are often also called "tiers".

[0035] However, other structures of the artificial neural network for the denoising block are also basically conceivable. The inverse process can be associated with a condition as an additional input parameter. This association with the condition can be called adjustment. Within the scope of the present invention, the condition can in particular include characterization data as an additional input parameter. The condition can be implemented, for example, with the aid of an embedding function and / or with the aid of a cross-attention mechanism. The training data can therefore include real image data and respectively assigned characterization data, so that a set of training data in particular includes training pairs of real image data and characterization data. Therefore, the characteristics of the detector characterized based on multiple scans or measurement results from the inspection process can be input as training data into the generation of synthetic image data.

[0036] Advantageously, a very large number of noisy images can be generated from a set of training images. As a result, the artificial neural network can very accurately learn the image domain from which it should generate samples. It has been demonstrated that using diffusion models, particularly through self-supervised learning, it is possible to develop a sufficient understanding of real image data from X-ray detectors to reliably generate high-quality synthetic image data.

[0037] For example, the architecture of the diffusion model can be based on a decoder-only transformer, which in particular has a cross-attention mechanism for adjusting the data. For example, the attention head can be arranged essentially conventionally. The feedforward network at the end of the block can consist of hidden layers and can in particular have a ReLU (rectified linear unit) or an ELU (exponential linear unit) as an activation function. A LayerNorm (layer normalization) or RMSNorm (root mean square layer normalization) operation can, for example, be used to normalize these values. In summary, a plurality of such blocks can be arranged in particular in the form of a U-Net in order to produce a synthetic image.

[0038] The diffusion model may include at least one latent space, an encoder for transferring image data from the image domain to the at least one latent space, a decoder for transferring data from the at least one latent space to the image domain, an embedding function for embedding representation data into the at least one latent space, and at least one denoising block. The at least one denoising block may preferably be located in the at least one latent space. The latent space may also be referred to as an embedding space. In particular, a latent space of the image domain and a latent space of the representation data may be provided. The latent space of the image domain is preferably a continuous representation, in particular an embedding, having a lower dimensionality than the image domain. Therefore, the latent space can be considered a smaller spatial representation of the image data than the image domain representation. The reduced complexity, corresponding to the lower dimensionality, enables efficient generation of image data using the diffusion model. The transfer of data from the image domain to the latent space, in particular into the image domain latent space, and vice versa, may be performed using the encoder or decoder. The image data may be summarized in the form of coordinates in the latent space, in particular in the image domain latent space. By adjusting the latent space using the representation data, the latent space of the representation data can be learned during training with respect to detector characteristics. The encoder is configured and / or can be trained to map the image data into the latent space, in particular into the image domain latent space. The decoder is configured and / or can be trained to generate image data, i.e., in particular, synthesized image data, from coordinates in a latent space. For example, a ResNet structure can be provided for the encoder and / or decoder. The encoder's ResNet structure can be designed so that it downsamples the input data to a lower resolution. Preferably, the downsampling can be provided in multiple blocks. In particular, the encoder's ResNet structure can be designed to embed image data having a higher resolution value into a lower resolution value. For example, the ResNet structure can be designed to embed image data having 512×512×1 pixels into 32×32×1 values. The downsampling step can, for example, include two ResNet layers, in particular convolutional layers, for example, with a kernel size of 3, wherein the number of channels is gradually doubled. In other words, the number of channels can be doubled when the resolution is halved. For example, self-attention blocks with 8 or 16 dimensions can be constructed for resolutions of 32 and 16. At the end of the encoder chain, the output can be normalized with group normalization and processed with a further Conv2D layer (e.g. kernel size 3) to channels corresponding to the output resolution, e.g. 32×32×16 channels. Group normalization is known in the prior art, in particular by dividing the channels into groups and calculating the mean and variance for normalization within each group. The decoder can be constructed in the opposite direction to the encoder, in particular by using corresponding upsampling instead of downsampling. An embedding function can be constructed and / or trained to extract detector characteristics from the representation data and represent them in a latent space, in particular the latent space of the representation data.For example, the characterization data can be represented as vectors in a latent space. For example, different detector modules can be encoded as numbers. Additionally or alternatively, the image data in the latent space can be represented as vectors. The image data to which noise is added can be set accordingly to randomly selected latent (noise) vector image information. A denoised latent vector can be generated from the latent noise vector using at least one denoising block. For example, a cross-attention mechanism can be used to take into account the embedded characterization data when generating the denoised latent vector. The cross-attention mechanism can, for example, be based on a transformer model. Preferably, provision can be made for a plurality of denoising blocks to be implemented in sequence, in particular until the noise is sufficiently removed.

[0039] According to embodiments, a generative network (particularly comprising at least one latent space, an encoder, a decoder, an embedding function, and at least one denoising block as components) can be trained overall using a set of training data. The set of training data can each include pairs of image data and representation data associated with each other. In particular, the components of the generative network can be optimized simultaneously. According to embodiments, a diffusion model includes a latent space, an encoder for transferring image data from an image domain into the latent space, a decoder for transferring data from the latent space into the image domain, an embedding function for embedding the representation data into the latent space, and at least one denoising block. The generative artificial intelligence is trained such that the encoder, decoder, and embedding function are first trained separately, and then the at least one denoising block is trained using the trained encoder, decoder, and embedding function. Thus, the encoder, decoder, and embedding function can be pre-trained before the at least one denoising block is trained using the pre-trained components. The encoder, decoder, and / or embedding function can be trained using self-supervised training. The autoencoder structure can be configured so that the encoder and decoder are trained simultaneously, so that the image data can be reconstructed. Similarly, the embedding of the representation data can be trained. For example, individual parts of the data can be masked and the corresponding network is asked to predict the corresponding data. Thus, each component in the component can learn to extract the corresponding basic information from the data. It can be provided that the latent space of the component is optimized for structural similarity, such as cosine similarity. In particular, the vectors generated by the individual components can be optimized so that the vectors of different components have similar angles to each other (especially cosine similarity). Similarity can be optimized, for example, with the help of a loss function. For example, contrastive training, especially self-supervised contrastive training, can be set up. The loss function can be, for example, a symmetric cross-entropy loss function. This training can be set up similarly to that described in Radford, A., "Learning Transferable Visual Models From Natural Language Supervision", 2021, doi:10.48550 / arXiv.2103.00020. It has been shown that this pre-training of the individual components can lead to better results because each individual component itself can already be optimized and the individual problems of the component can be better solved. In particular, it has been shown that less training data is required to obtain generally good results. For example, a neural network can be trained with a maximum of approximately 10,000 images. For example, the training data and test data can be divided into approximately 80:20 ratios to obtain good results. Training can be performed, for example, via 5-fold cross validation.

[0040] According to embodiments, the synthetic image data corresponds to a synthetic phantom image, in particular a synthetic water phantom image. A water phantom image is, in particular, an image recorded using an X-ray imaging system, in particular a computed tomography system, wherein a water phantom serves as the examination subject. Typically, a water phantom is a container filled with water, such as a Perspex container filled with distilled water. A water phantom can be used as a representative example of real living tissue. Advantageously, water phantom images can, on the one hand, serve as a good measure of the image quality of an X-ray detector, and, on the other hand, can be relatively easy to generate a large amount of training data for training algorithms. In particular, when recording water phantom images, fewer radiation protection regulations must be considered than, for example, when the examination subject is living. For example, pairs of characterization data and water phantom images, each of which is assigned to an X-ray detector having an arrangement of detector modules, can be used to train the algorithm. Corresponding characterization data can be assigned to each detector module of the X-ray detector.

[0041] According to an embodiment, the method comprises the following further steps:

[0042] receiving real image data, which are measured by an X-ray detector and which in particular correspond to the synthetic image data;

[0043] - optionally, registering the real image data with the synthetic image data,

[0044] -Retrain the trained algorithm using real image data as training data.

[0045] Optionally, retraining can also be performed using real image data based on the characterization data. The real image data and the associated characterization data can be archived for retraining purposes. Algorithms or other trainable algorithms can be retrained, retrained, and / or fine-tuned using real image data. Advantageously, this allows for increasingly accurate representations of parameters influencing image quality. For example, provision can be made for the additional training data thus obtained to be added to the existing training data. Retraining can be performed similarly to initial training.

[0046] A further aspect of the present invention is a method for quality control during the production of an X-ray detector for an X-ray imaging system, in particular for a computed tomography system, the X-ray detector having a plurality of detector modules, wherein the method comprises the following steps:

[0047] - recording and / or creating characterization data of a detector module of the X-ray detector, wherein at least part of the characterization data is recorded by measurements with the detector module without an examination object;

[0048] - carrying out the method according to the invention using the recorded and / or created characterization data;

[0049] - analyzing the synthetic image data and estimating the quality of the X-ray detector based on the synthetic image data.

[0050] Characterization data can be generated, for example, using an X-ray source for generating X-rays. The X-ray source can include, for example, an X-ray tube. Characterization data can optionally be generated using sensors, particularly temperature sensors. The quality of the X-ray detector can be estimated, for example, based on an evaluation of image artifacts and / or an evaluation of whether and / or to what extent image artifacts are present. Image artifacts can include, for example, ring artifacts or streak artifacts. The extent of image artifacts can be assessed, for example, based on the intensity and / or frequency of the artifacts. Optionally, additionally or alternatively, the extent of the artifacts can be assessed based on their location in the image data. The estimation of the quality of the X-ray detector can be automatically determined, for example, using an evaluation algorithm. The evaluation algorithm can be implemented and / or configured on a computing unit, for example, which can be part of a computer and / or part of an X-ray imaging system. The evaluation algorithm can be part of a computer program, particularly a computer program as described herein. Alternatively, the quality estimation can be set by a user, for example, through optical analysis of the composite image data.

[0051] Depending on the embodiment, the characterization data may be associated with an arrangement of detector modules in the X-ray detector. A further step may be provided for virtually or physically assembling the detector modules in the X-ray detector according to the arrangement or an arrangement of the detector modules in the X-ray detector. This step may be provided, for example, before or after recording and / or generating the characterization data.

[0052] According to an embodiment, multiple sets of characterization data are recorded and / or created, wherein each set of characterization data corresponds to a different arrangement of detector modules and / or a different combination of detector modules, and synthetic image data are generated and analyzed for each of the multiple sets of characterization data. Advantageously, the most suitable arrangement can thus be found from a plurality of possible arrangements. In particular, it has been found to be sufficient to create characterization data for each individual detector module, thereby allowing analysis of different arrangements of detector modules and / or different combinations of detector modules. The method may optionally include the further step of selecting an arrangement and / or combination of detector modules based on the analysis of the synthetic image data, in particular such that the selection is performed according to the arrangement and / or combination of detector modules based on an assessment of image quality. If a module replacement is provided, then, for example, it is already possible to assess which of the multiple available detector modules is suitable before the module replacement.

[0053] Another aspect of the present invention is a computer-implemented method for training a trainable algorithm, comprising:

[0054] - receiving input training data in the form of characterization data for an X-ray detector of an X-ray imaging system, in particular a computed tomography system;

[0055] - receiving output training data, in particular in the form of real image data, which have been recorded using an X-ray detector, wherein the output training data are correlated with the input training data;

[0056] - training a trainable algorithm based on input training data and output training data, wherein the trainable algorithm is particularly based on machine learning;

[0057] -Provide trained algorithms.

[0058] For example, the weights of a trainable algorithm can first be randomly initialized. To generate training data, for example, characterization data of an X-ray detector can be created using an arrangement of detector modules, and then real image data can be generated using the manufactured X-ray detector. These characterization data can be related in particular to the corresponding real image data. The input training data can form a set of training data together with the output training data. Optionally, a database of already existing characterization data and the corresponding real image data can be used, as long as such data exists. The training data can be designed, for example, as vectors and / or matrices. For example, the image data can be set with a resolution of 512×512 to 16384×16384. The image data can be set, for example, in HU scale (Henriquez scale).

[0059] For example, it can be provided that for training, a certain amount of real image data and characterization data is used based on a certain number of inspected and constructed detectors in the order of 50. Generally, a plurality of image data and associated characterization data can be obtained from each constructed detector. This generally allows sufficient information to be collected for different artifacts. Alternatively or additionally, the training data can be provided in the form of synthetic data. For example, synthetic data can be provided by modifying the characterization data and image data with the help of a model. Synthetic training data can be provided in particular for pre-training. Training parameters can be determined and optimized, for example, within the scope of hyperparameter optimization, for example by grid search. In some cases, it can be provided that the training data is reduced, especially when it is proven in individual cases that better training results can be achieved as a result. For example, image data with a resolution of 512×512 pixels can be reduced to 256×256 pixels.

[0060] According to embodiments, the trainable algorithm comprises a trainable generative artificial intelligence, in particular as described herein. Preferably, the trainable algorithm comprises a diffusion model. In particular, the generative artificial intelligence may comprise a diffusion model having at least one denoising block and a diffusion process, which is used to generate synthetic image data. Advantageously, noisy images or image data can be generated from a set of training images to a certain extent without restriction. This allows the algorithm to very precisely learn the image domain from which it should generate sample or synthetic image data.

[0061] According to embodiments, a trainable algorithm, in particular including generative artificial intelligence, is generally trained using a set of training data. The trained algorithm as a whole may in particular include at least one latent space, an encoder, a decoder, an embedding function, and at least one denoising block. In particular, provision may be made for these components to be trained simultaneously within the scope of the training.

[0062] According to an embodiment, the diffusion model includes a latent space, an encoder for transferring image data from the image domain into the latent space, a decoder for transferring data from the latent space into the image domain, an embedding function for embedding representation data into the latent space, and at least one denoising block, wherein the generative artificial intelligence is trained such that the encoder, decoder, and embedding function are first trained separately, and then the at least one denoising block is trained using the trained encoder, decoder, and embedding function. The training can each be based on self-supervised learning. Various parts of the training data can be masked out and predicted by the corresponding components.

[0063] Another aspect of the present invention is a computer program product or a storage medium, in particular a non-volatile storage medium, comprising instructions that, when executed by a computer, cause the computer to implement the steps of the method described herein, in particular a computer-implemented method for supporting the evaluation of characterization data of an X-ray detector for an X-ray imaging system, a method for quality control during the manufacture of X-ray detectors for an X-ray imaging system, and / or a method for training a trainable algorithm. All advantages and features of the proposed method can be similarly transferred to the computer program product or storage medium, and vice versa. The computer program product can, for example, be stored on a computer-readable storage medium, in particular a non-volatile storage medium. The storage medium can, for example, be a hard drive, an SSD, a flash memory, an online server, or the like.

[0064] A further aspect of the present invention is a system comprising an interface for receiving characterization data from a detector module of an X-ray detector, and a computing unit connected to the interface and configured to implement the proposed method as described herein. All advantages and features of the method and the computer program product or storage medium can be similarly transferred to the system, and vice versa.

[0065] Unless explicitly stated otherwise, all embodiments described herein can be combined with each other. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Embodiments will be described below with reference to the accompanying drawings.

[0067] Figure 1 shows a flow chart of a computer-implemented method for supporting evaluation of characterization data of an X-ray detector for an X-ray imaging system according to an embodiment of the present invention,

[0068] Figure 2 shows a flowchart of a computer-implemented method for training a trainable algorithm according to an embodiment of the present invention,

[0069] Figure 3 A schematic diagram showing the structure of a trained algorithm or a trainable algorithm according to an embodiment of the present invention is shown,

[0070] Figure 4 Schematically shows Figure 4 The working principle of the diffusion model shown,

[0071] Figure 5 shows a flow chart of a computer-implemented method for training a trainable algorithm based on a diffusion model according to an embodiment of the present invention,

[0072] Figure 6 shows a schematic structure of a system according to an embodiment of the present invention, and

[0073] Figure 7 A method for quality control when manufacturing an X-ray detector for an X-ray imaging system having a plurality of detector modules is shown according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] Figure 1A flowchart illustrates a computer-implemented method for supporting the evaluation of characterization data 31 of an X-ray detector for an X-ray imaging system, according to an embodiment of the present invention. The X-ray imaging system may, in particular, be a computed tomography system. The X-ray detector includes a plurality of detector modules. In a first step 101 of the method, characterization data 31 of the detector modules of the X-ray detector is received. At least a portion of the characterization data 31 is based on measurement data recorded by the detector modules without an examination object. In particular, the characterization data 31 may be associated with the arrangement of the detector modules in the X-ray detector. In other words, the characterization data 31 may include information determining the arrangement and / or configuration of the detector modules in the X-ray detector. For example, the characterization data 31 may include the response of the detector pixels of the detector modules to radiation without an examination object. Additionally or alternatively, the characterization data 31 may include, for example, the temporal noise characteristics of individual detector pixels of the detector modules. Additionally or alternatively, the characterization data 31 may include, for example, signal instabilities of the detector modules caused by incident radiation. The characterization data 31 may include a list of detector pixels of the detector modules that have been marked as defective. Additionally or alternatively, the characterization data 31 may, for example, include the expected effect of the orientation of a collimator, particularly a tilted collimator, on the detection of radiation signals by the detector module. Additionally or alternatively, the characterization data 31 may, for example, include a correlation between the detector response of the detector module and a thermally influencing variable. In a further step 102, a trained algorithm is applied to the characterization data 31. The trained algorithm may, in particular, include a trained generative artificial intelligence. For example, the trained generative artificial intelligence may be or may include a diffusion model. The trained algorithm generates as output synthetic image data 14, which simulates image data 13 recorded by an X-ray detector of an X-ray imaging system. The synthetic image data 14 may include raw data corresponding to raw data recorded by the X-ray detector. The synthetic image data 14 may include an image corresponding to a reconstructed image recorded by the X-ray detector. The synthetic image data 14 may, for example, correspond to a synthetic phantom image, particularly a synthetic water phantom image. The synthetic image data 14 is provided in a further step 103. This provision can include, for example, storage of the synthetic image data 14, transmission via a network to another system and / or data storage, and / or output to a user interface. Optionally, the method or further embodiments of the method described herein can include further steps 104, 105 for retraining the trained algorithm. To this end, in step 104, real image data 13 measured by a subsequently constructed X-ray detector integrated into the X-ray imaging system is received. The real image data 13 corresponds in particular to the synthetic image data 14.This should be understood in particular to mean that the real image data 13 and the synthetic image data 14 represent or are intended to represent the same object and / or were generated under measurement conditions corresponding to the measurement conditions simulated for the synthetic image data 14. This is because the measurement conditions simulated for the synthetic image data 14 can, for example, correspond to the measurement conditions of the real image data 13, in which the real image data 13 were recorded under measurement conditions in which the training data for the trained algorithm were also recorded. The real image data 13 can optionally be registered with the synthetic image data 14. In a further step 105, the trained algorithm is retrained using the real image data 13 as training data. The retrained algorithm, i.e., optionally adapted, can then be used, in particular, for further applications of the method.

[0075] Figure 2 A flow chart of a computer-implemented method for training a trainable algorithm according to an embodiment of the present invention is shown. In a first step 201, input training data are received in the form of characterization data 31 of an X-ray detector for an X-ray imaging system, in particular for a computed tomography system. In a further step 202, output training data are received in the form of real image data 13 recorded by the X-ray detector. Preferably, the output training data are related to the input training data or are associated with these input training data. In particular, the output training data and some training data can be associated as training pairs. In a further step 203, the trainable algorithm is trained based on the input training data and the output training data. The trainable algorithm can in particular be based on machine learning. For example, the trainable algorithm can be a diffusion model. In a further step 204, the now trained algorithm is provided.

[0076] Figure 3A schematic diagram illustrates the structure of a trained or trainable algorithm according to an embodiment of the present invention. In this example, the algorithm comprises generative artificial intelligence based on a diffusion model. The diffusion model can be implemented and trained similarly to the method described in "High-Resolution Image Synthesis with Latent Diffusion Models" by Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer, CVPR 2022. The diffusion model comprises a latent space 2, an encoder 11 for transferring image data from an image domain 1 into the latent space 2, a decoder 12 for transferring data from the latent space 2 into the image domain 1, and an embedding function 32 for embedding representation data 31 from the original data domain 3 into the latent space 2. A denoising block 22, at least one further denoising block 23, and a diffusion process 21 are provided in the latent space 2. The encoder 11 maps the image data from the image domain 1 into the latent space 2, particularly as coordinates or vectors. Decoder 12 maps coordinates from latent space 2 into image domain 1 or generates image data, particularly an image, in image domain 1 from coordinates in latent space 2. In this example, image data 13 and 14 are water phantom images. For example, a ResNet architecture can be provided for each encoder and decoder. The encoder's ResNet architecture can be designed to downsample input data in multiple blocks to a lower resolution. For example, the ResNet architecture can be designed to embed image data with 512×512×1 pixels into 32×32×1 values. The downsampling step can, for example, include two ResNet layers, particularly convolutional layers with a kernel size of 3, in which the number of channels is gradually doubled. In other words, the number of channels can be doubled as the resolution is halved. For example, self-attention blocks with 8 or 16 dimensions can be constructed for resolutions of 32 and 16. At the end of the encoder chain, the output can be normalized using group normalization and processed to, for example, 32×32×16 channels using a further Conv2D layer (kernel size 3). The decoder can be constructed in the opposite direction to the encoder, similarly using upsampling instead of downsampling. On the other hand, the representation data 31 is embedded into the latent space 2 using an embedding function 32. At least one denoising block 22, 23 is responsible for the generative process for generating the synthetic image data 14. During training, noise is added to the real image data 13 mapped into the latent space 2 by the encoder 11 via a diffusion process 21 and then removed again by the denoising blocks 22, 23. The noise can be, for example, Gaussian noise.In a diffusion process 21, the real image mapped into the latent space 2 (which serves as a reference image) is noisy in that successive noises are added to the reference image until there is generally a distribution Z of fully noisy images corresponding to the fully noisy image. T After the diffusion process 21, the denoising blocks 22 and 23 are trained in an inverse process to gradually remove the noise, so that a denoised distribution Z is generated, so that the synthetic image data 14 denoised by the decoder 12 is finally generated based on the real image data 13. Here, the noise addition of the diffusion process 21 is undone. After the first denoising block 22, there is initially a distribution Z with partially added noise. T-1, which is ultimately transformed into a fully denoised distribution Z after applying at least one further denoising block 23. The denoising blocks 22, 23 are thus trained to generate image data 14 from noise. During training, this inverse process is approximated by adapting the trainable parameters of the denoising blocks 22, 23. The inverse process in the denoising blocks 22, 23 is associated (adjusted) with a condition as an additional input parameter, wherein the condition includes representation data 31 as an additional input parameter, which is embedded in the latent space 2 by means of an embedding function 32 and enters the inverse process via a cross-attention mechanism 24. This cross-attention mechanism 24 can be implemented, for example, as described in "High-Resolution Image Synthesis with Latent Diffusion Models" by Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser and Bjorn Ommer, CVPR, 2022. The training data can therefore include multiple training pairs of real image data 14 and representation data 31. After training, a diffusion model can be used to generate new synthetic data, in particular synthetic image data 14, from random noise. Starting from a randomly selected latent vector for the image information, a new denoised latent vector is generated via a cross-attention mechanism 24, which is fed with the representation data 31 embedded in the latent space 2. This process is repeated multiple times, particularly by adding at least one additional denoising block 23, until the noise has been sufficiently removed and image data 14, in particular an image, can be generated by the decoder 12. In the illustrated embodiment, the denoising block 22 and the at least one additional denoising block 23 are each constructed using a U-Net architecture. This U-Net architecture can be constructed similarly to the architecture of the encoder 11 and decoder 12. For example, it can be provided that the resolution in the U-Net is halved until a 2×2 resolution is achieved. Two ResNet blocks, in particular including attention heads, can be provided for each resolution. This concept can be token-based, in particular, with 32 dimensions per token. However, alternatives to the U-Net architecture are also conceivable that allow for the integration of a suitable cross-attention mechanism 24 and can be trained efficiently. Thus, in the latent space 2, the input of the decoder 12 can be created from the detector representation of the representation data 31 by a back diffusion process (inverse process 25) of the denoising blocks 22, 23, and finally the image data for the X-ray detector can be predicted in the form of synthetic image data 14. Via the representation data 31, the characteristics of the X-ray detector or the detector modules of the X-ray detector flow into the generation process for generating the synthetic image data 14.In the example shown here, the characterization data 31 includes the air jet of the entire x-ray detector (the first image shown in the characterization data 31), wherein each stripe shown in the image is assigned to a detector module. In addition, further characterization data 31, such as detector measurement data of a detector module with a moving focus, may be provided as input. While the characterization data 31 is presented visually here as an image, it can be provided that the characterization data 31 is actually input as raw data or as numerical values.

[0077] In this embodiment, the method according to the present invention is based on the fact that a sufficient understanding of the image data of the X-ray detector can be developed by self-supervised learning via a diffusion process 21, which is summarized in the form of coordinates in a latent space 2. Simultaneously, the latent space 2 of detector characteristics is learned by adjustment using characterization data 31. The algorithmic design allows the characterization data 31 to be linked to the image data 13, 14. This utilizes the fact that the detector characteristics represented by the characterization data 31 may be decisive for artifacts that may be observed in the image.

[0078] Figure 4 Schematically shows Figure 3 The working principle of the diffusion model shown. During training, continuous Gaussian noise is added to the water phantom image in a diffusion process 21 (forward process). In the example shown here, the water phantom image is largely artifact-free. In the inverse process 25, the noise is continuously removed again by denoising blocks 22 and 23. The inverse process is adjusted using characterization data 31. After training, the characterization data 31 can be used together with (e.g., randomly generated) noise to generate synthetic image data 14 in the inverse process 25.

[0079] Figure 5A flow chart of a computer-implemented method for training a trainable algorithm based on a diffusion model according to an embodiment of the present invention is shown. In a first step 300, pre-training steps 301-303 are first performed by training the encoder 11 in step 301, training the decoder 12 in another step 302, and training the embedding function 32 in yet another step 303. Pre-training of the individual components can be achieved, in particular, by unsupervised learning, wherein individual parts of the data are to be masked and predicted by the network. In this case, each of the components 11, 12, 32 configured to transition into or out of the latent space 2 learns to extract basic information from the data. These pre-training steps can, for example, be performed in parallel or (optionally partially) sequentially. Optionally, after training the individual components, in a further step 304, the mapping can be optimized for structural similarity (e.g., via cosine similarity). Subsequently, in further steps 310-340, at least one denoising block 22, 23 of the diffusion model is trained using the trained encoder 11, decoder 12, and embedding function 32. To this end, in step 310, input training data in the form of characterization data 31 of an X-ray detector for an X-ray imaging system, in particular a computed tomography system, is received. In a further step 320, output training data in the form of real image data 13 recorded by the X-ray detector is received. In particular, the output training data and some training data can be associated as training pairs. In a further step 330, the diffusion model as a whole is trained based on the input training data and the output training data. In a further step 340, the now fully trained diffusion model is provided.

[0080] Figure 6 1 shows a schematic structure of a system according to an embodiment of the present invention. The system comprises an interface 50 for receiving characterization data 31 from a detector module of an X-ray detector, and a computing unit 60 connected to the interface 50 and configured to implement the method described herein. For outputting the composite image data 14, a further interface 70 is provided, which is connected to the computing unit 60. Optionally, the interface 50 and the further interface 70 can be integrated into a common interface.

[0081] Figure 7A method for quality control during the manufacture of an X-ray detector for an X-ray imaging system, in particular for a computed tomography system, having a plurality of detector modules, is shown according to an embodiment of the present invention. In a first step 400, characterization data 31 of the detector modules of the X-ray detector are recorded and / or created, wherein at least a part of the characterization data 31 is recorded by measurements with the detector modules without an object to be examined, wherein the characterization data 31 are in particular assigned to an arrangement of the detector modules in the X-ray detector. Optionally, a plurality of groups of characterization data 31 can be recorded and / or created, wherein each group of characterization data 31 corresponds to a different arrangement of the detector modules and / or a different combination of the detector modules. In further steps 401-403, steps 401-403 of the method are implemented using the recorded and / or created characterization data 31, as for example with reference to Figure 1 Here, the steps 401-403 shown here correspond in particular to the steps in Figure 1 101 - 103 of the method shown in FIG. 4. If a plurality of sets of characterization data 31 have been recorded and / or created, synthetic image data 14 can be generated for each set of characterization data 31. In a further step 410, the synthetic image data 14 are analyzed and the quality of the X-ray detector is estimated based on the synthetic image data 14. If a plurality of sets of characterization data 31 have been recorded and / or created and synthetic image data 14 have been generated for each set of characterization data 31, the synthetic image data 14 can be analyzed separately for each set of characterization data 31. Similar to Figure 1 In addition to the retraining steps 104-105 in the embodiment, the retraining steps 404-405 may also be optionally provided.

Claims

1. A computer-implemented method for supporting the evaluation of characterization data (31) of an X-ray detector for an X-ray imaging system, in particular for a computed tomography system, the X-ray detector having a plurality of detector modules, wherein: The method comprises the following steps: - receiving characterization data (31) of a detector module of the X-ray detector, wherein at least a portion of the characterization data (31) is based on measurement data recorded by the detector module without an examination object; - applying the trained algorithm to the characterization data (31), wherein synthetic image data (14) are generated as output, which simulate image data (13) recorded by an X-ray imaging system, in particular a computed tomography system, using the X-ray detector; - providing said composite image data (14).

2. The method according to claim 1, wherein The characterization data (31) are assigned to an arrangement of detector modules in the X-ray detector.

3. The method according to claim 1 or 2, wherein The characterization data (31) includes one or more of the following: - a response of the detector pixels of the detector module to radiation in the absence of an examination object, - the noise characteristics of the individual detector pixels of the detector module over time, - signal instabilities of the detector module caused by incident radiation, - a list of detector pixels of said detector module marked as defective, the expected influence of the orientation of the collimator, in particular of a tilted collimator, on the detection of radiation signals by the detector module, - Dependence of the detector response of the detector module on thermally influencing variables.

4. A method according to any one of the preceding claims, wherein The trained algorithm comprises a trained generative artificial intelligence, wherein the generative artificial intelligence comprises in particular a diffusion model with at least one denoising block (22, 23), with which the synthetic image data (14) are generated.

5. The method according to claim 4, wherein The diffusion model includes: at least one latent space (2), an encoder (11) for transferring image data from the image domain (1) into the at least one latent space (2), a decoder (12) for transferring data from said at least one latent space (2) into said image domain (1), and an embedding function (32) for embedding the representation data (31) into the at least one latent space (2), and at least one denoising block (22, 23), The generative artificial intelligence is trained in such a way that first the encoder (11), the decoder (12) and the embedding function (32) are trained individually, wherein the at least one denoising block (22, 23) is then trained with the aid of the trained encoder (11), the decoder (12) and the embedding function (32).

6. A method according to any one of the preceding claims, wherein The synthetic image data (14) correspond to a synthetic phantom image, in particular a synthetic water phantom image.

7. The method according to any one of the preceding claims, comprising the following further steps: - receiving real image data (13), which are measured by the X-ray detector and which in particular correspond to the synthetic image data (14); - optionally, registering the real image data (13) with the synthetic image data (14), - retraining the trained algorithm using the real image data (13) as training data.

8. A method for quality control during the production of an X-ray detector for an X-ray imaging system, in particular for a computed tomography system, the X-ray detector having a plurality of detector modules, wherein: The method comprises the following steps: - recording and / or creating characterization data (31) of a detector module of the X-ray detector, wherein at least part of the characterization data (31) is recorded by measurements with the detector module without an examination object; - carrying out a method according to any of the preceding claims using the recorded and / or created characterization data (31); - analyzing the composite image data (14) and estimating the quality of the X-ray detector based on the composite image data (14).

9. The method according to claim 8, wherein The characterization data (31) are assigned to an arrangement of detector modules in the X-ray detector.

10. The method according to claim 8 or 9, wherein: A plurality of sets of characterization data (31) are recorded and / or created, wherein each set of characterization data (31) corresponds to a different arrangement of the detector modules and / or a different combination of the detector modules, and wherein synthetic image data (14) are generated and analyzed for each set of characterization data (31).

11. A computer-implemented method for training a trainable algorithm, comprising: - receiving input training data in the form of characterization data (31) of an X-ray detector for an X-ray imaging system, in particular a computed tomography system; - receiving output training data recorded by the X-ray detector, in particular in the form of real image data (13), wherein the output training data is correlated with the input training data; - training a trainable algorithm based on the input training data and the output training data, wherein the trainable algorithm is particularly based on machine learning; -Provide trained algorithms.

12. A computer program product or storage medium, in particular a non-volatile storage medium, comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to any one of claims 1 to 7, the method according to any one of claims 8 to 10 and / or the method according to claim 11.

13. A system comprising: An interface (50) for receiving characterization data (31) of a detector module of an X-ray detector; and A computing unit (60) is connected to the interface and is configured to carry out the method according to any one of claims 1 to 7, the method according to any one of claims 8 to 10 and / or the method according to claim 11.

Images