GENERATION OF SYNTHETIC RADIOLOGICAL RECORDINGS

Description

[0001] The present disclosure relates to the technical field of radiology. Subjects of the present disclosure are a novel approach for training a machine learning model to generate synthetic radiological images based on measured radiological images and the use of the trained machine learning model to generate synthetic radiological images.

[0002] Medical imaging is the technique and process of imaging the interior of the body for clinical analysis and medical interventions, as well as for visually depicting the function of specific organs or tissues. Medical imaging is used to visualize internal structures hidden beneath the skin and bones and to diagnose and / or treat diseases.

[0003] Advances in imaging and machine learning have led to a rapid increase in the potential use of artificial intelligence in various medical imaging tasks, such as risk assessment, detection, diagnosis, prognosis, and therapy.

[0004] Machine learning models are used, among other things, to segment radiological images, to induce contrast enhancements in radiological images and / or to predict a radiological image in a temporal sequence of radiological images.

[0005] WO2019 / 074938A1, for example, discloses a method for reducing the amount of contrast agent when generating radiological images using an artificial neural network.

[0006] US 10,997,716 B2 also discloses a method for reducing the amount of contrast agent when generating radiological images using an artificial neural network.

[0007] WO2021 / 052896A1 discloses a method in which an MRI image of a patient's liver during the hepatobiliary phase is not generated by measurement, but is calculated (predicted) based on MRI images from one or more previous phases in order to shorten the patient's stay in the MRI scanner.

[0008] WO2021 / 197996A1 discloses a method in which MRI images are simulated after the application of a so-called blood pool contrast agent using a machine learning model.

[0009] In all of the above-mentioned methods, at least one measured radiological image of an examination area of ​​an object is fed to a trained machine learning model, and the model generates a synthetic radiological image. Both the supplied radiological image and the synthetic radiological image are representations of the examination area in a spatial representation.

[0010] It has been shown that synthetic radiological images often contain artifacts. Fine structures, in particular, are often not accurately depicted, or structures appear in areas of the synthetic radiological image that are not present in the represented tissue.

[0011] This revelation addresses these and other problems.

[0012] A first aspect of the present disclosure is a method for training a machine learning model to generate synthetic radiological images based on measured radiological images. The training method comprises the steps: Receiving and / or providing training data, wherein the training data for each examination object of a plurality of examination objects comprises a set of input data and target data, wherein each set comprises ∘ at least one input representation of an examination region of the examination object in a first state as input data and ∘ a target representation of the examination region of the examination object in a second state as well as a transformed target representation as target data, wherein the transformed target representation represents at least a part of the examination region of the examination object ∘ in the frequency domain if the target representation represents the examination region of the examination object in the spatial domain, or ∘ in the spatial domain if the target representation represents the examination region of the examination object in the frequency domain, Training a machine learning model,wherein the machine learning model is configured to generate a synthetic representation of the examination region of the examination object based on at least one input representation and model parameters, wherein the at least one input representation represents the examination region in a first time period before and / or after the application of an amount of contrast agent, and the synthetic representation represents the examination region in a second time period after the application of the amount of contrast agent, wherein the second time period lies chronologically behind the first time period, wherein the training for each examination object of the plurality of examination objects comprises: ∘ feeding the at least one input representation to the machine learning model, ∘ receiving a synthetic representation of the examination region of the examination object from the machine learning model,∘ Generating and / or receiving a transformed synthetic representation based on and / or to the synthetic representation, wherein the transformed synthetic representation represents at least a part of the examination area of ​​the examination object ▪ in the frequency domain if the synthetic representation represents the examination area of ​​the examination object in the spatial domain, or ▪ in the spatial domain if the synthetic representation represents the examination area of ​​the examination object in the frequency domain, ∘ Quantifying the deviations i) between at least a part of the synthetic representation and at least a part of the target representation and ii) between at least a part of the transformed synthetic representation and at least a part of the transformed target representation by means of an error function, ∘ Modifying model parameters with a view to reduced deviations,Outputting and / or storing the trained machine learning model and / or the model parameters and / or transmitting the trained machine learning model and / or the model parameters to a separate computer system. ,

[0013] Another subject of the present disclosure is a computer-implemented method (prediction method) for generating a synthetic radiological image using the trained machine learning model. The prediction method comprises the steps: Providing a trained machine learning model, wherein the trained machine learning model has been trained using training data to generate a synthetic representation of the examination region in a second state based on at least one input representation of an examination area of ​​an examination object in a first state, wherein the training data for each examination object of a plurality of examination objects comprise i) an input representation of the examination region, ii) a target representation and iii) a transformed target representation, wherein the transformed target representation represents at least a part of the examination area of ​​the examination object in the frequency domain if the target representation represents the examination area of ​​the examination object in the spatial domain, or in the spatial domain,if the target representation represents the examination region of the examination object in the frequency domain, wherein the machine learning model has been trained, for each examination object, to minimize deviations between i) at least a part of the synthetic representation and at least a part of the target representation and ii) between at least a part of a transformed synthetic representation and at least a part of the transformed target representation, receiving at least one input representation of an examination region of a new examination object in the first state, wherein the at least one input representation represents the examination region in a first time period before and / or after the application of an amount of contrast agent, inputting the at least one input representation of the examination region of the new examination object into the trained machine learning model,Receiving a synthetic representation of the examination area of ​​the new examination object in the second state from the machine learning model, wherein the synthetic representation represents the examination area in a second time period after the application of the amount of contrast agent, wherein the second time period is temporally behind the first time period, outputting and / or storing the synthetic representation and / or transmitting the synthetic representation to a separate computer system.

[0014] Another subject of the present disclosure is a computer system comprising a receiving unit, a control and computing unit, and an output unit, wherein the control and computing unit is configured to cause the receiving unit to receive at least one input representation of an examination region of a new examination object in a first state, wherein the at least one input representation represents the examination region in a first time period before and / or after the application of a quantity of a contrast agent, wherein the control and computing unit is configured to input the received at least one input representation into a trained machine learning model, ∘ wherein the trained machine learning model has been trained using training data to generate a synthetic representation of the examination region in a second state based on at least one input representation of an examination region of an examination object in a first state,∘ wherein the training data for each examination object of a plurality of examination objects comprises i) an input representation of the examination area, ii) a target representation and iii) a transformed target representation, ∘ wherein the transformed target representation represents at least a part of the examination area of ​​the examination object in the frequency domain if the target representation represents the examination area of ​​the examination object in the spatial domain, or in the spatial domain if the target representation represents the examination area of ​​the examination object in the frequency domain, ∘ wherein the machine learning model has been trained for each examination object,To minimize deviations between i) at least a part of the synthetic representation and at least a part of the target representation and ii) between at least a part of a transformed synthetic representation and at least a part of the transformed target representation, wherein the control and computing unit is configured to receive a synthetic representation of the examination region of the new examination object in the second state from the machine learning model, wherein the synthetic representation represents the examination region in a second time period after the application of the amount of contrast agent, wherein the second time period lies chronologically behind the first time period, wherein the control and computing unit is configured to cause the output unit to output the synthetic representation and / or to store it and / or to transmit it to a separate computer system.

[0015] A further subject of the present disclosure is a computer program product comprising a computer program that can be loaded into a working memory of a computer system and causes the computer system to perform the following steps: Receiving at least one input representation of an examination region of a new examination object in a first state, wherein the at least one input representation represents the examination region in a first time period before and / or after the application of an amount of contrast agent, inputting the at least one input representation of the examination region of the new examination object into a trained machine learning model, ∘ wherein the trained machine learning model has been trained using training data to generate a synthetic representation of the examination region in a second state based on at least one input representation of an examination region of an examination object in a first state, ∘ wherein the training data for each examination object of a plurality of examination objects i) an input representation of the examination region,ii) a target representation and iii) a transformed target representation, ∘ wherein the transformed target representation represents at least a part of the examination area of ​​the examination object in the frequency domain if the target representation represents the examination area of ​​the examination object in the spatial domain, or in the spatial domain if the target representation represents the examination area of ​​the examination object in the frequency domain, ∘ wherein the machine learning model has been trained, for each examination object, to minimize deviations between i) at least a part of the synthetic representation and at least a part of the target representation and ii) between at least a part of a transformed synthetic representation and at least a part of the transformed target representation,Receiving a synthetic representation of the examination area of ​​the new examination object in the second state from the machine learning model, wherein the synthetic representation represents the examination area in a second time period after the application of the amount of contrast agent, wherein the second time period is temporally behind the first time period, outputting and / or storing the synthetic representation and / or transmitting the synthetic representation to a separate computer system.

[0016] The invention is explained in more detail below, without distinguishing between the subject matters of the invention (training method, prediction method, computer system, computer program product). Rather, the following explanations are intended to apply analogously to all subject matters of the invention, regardless of the context in which they occur (method, computer system, computer program product).

[0017] If steps are mentioned in a particular order in this description or in the claims, this does not necessarily mean that the invention is limited to that order. Rather, it is conceivable that the steps may be performed in a different order or even in parallel; unless a step builds on another step, which absolutely requires that the subsequent step be performed (which will become clear in individual cases). The specified sequences thus represent preferred embodiments of the invention.

[0018] With the aid of the present invention, representations of an examination region of an examination object can be predicted. Such a predicted representation is also referred to in this disclosure as a synthetic representation.

[0019] The "object of investigation" is usually a living being, preferably a mammal, most preferably a human.

[0020] The "area of ​​investigation" is a part of the object under investigation, for example an organ or part of an organ such as the liver, brain, heart, kidney, lung, stomach, intestine, pancreas or part of the aforementioned organs or several organs or another part of the body.

[0021] In a preferred embodiment of the present invention, the examination area is the liver or a part of the liver of a human.

[0022] The examination area, also known as the recording volume (English: field of view , FOV), represents a volume that is depicted in radiological images. The examination area is typically defined by a radiologist, for example, on an overview image (English: localizer ). Of course, the examination area can alternatively or additionally be determined automatically, for example based on a selected protocol.

[0023] A "representation of the examination area" is usually the result of a radiological examination.

[0024] "Radiology" is the branch of medicine that deals with the application of primarily electromagnetic radiation and (including ultrasound diagnostics, for example) mechanical waves for diagnostic, therapeutic, and / or scientific purposes. In addition to X-rays, other ionizing radiation such as gamma radiation or electrons is also used. Since a key application is imaging, other imaging techniques such as sonography and magnetic resonance imaging (MRI) are also considered radiology, even though these techniques do not involve the use of ionizing radiation. The term "radiology" within the meaning of the present invention thus encompasses, in particular, the following examination methods: computed tomography, magnetic resonance imaging, and sonography.

[0025] In a preferred embodiment of the present invention, the radiological examination is a magnetic resonance imaging or computed tomography examination. Most preferably, the radiological examination is a magnetic resonance imaging examination.

[0026] Magnetic resonance imaging, abbreviated MRI or MRI (English: Magnetic Resonance Imaging ) is an imaging technique that is used primarily in medical diagnostics to depict the structure and function of tissues and organs in the human or animal body.

[0027] In MR imaging, the magnetic moments of protons in a subject are aligned in a basic magnetic field, resulting in macroscopic magnetization along a longitudinal direction. This magnetization is then deflected from its rest position by applying radiofrequency (RF) pulses (excitation). The return of the excited states to their rest position (relaxation), or the magnetization dynamics, is subsequently detected as relaxation signals using one or more RF receiver coils.

[0028] For spatial encoding, rapidly switched magnetic gradient fields are superimposed on the basic magnetic field. The acquired relaxation signals or the detected MR data are initially available as raw data in the frequency domain (so-called k-space data) and can be transformed into the spatial domain (image space) by subsequent inverse Fourier transformation.

[0029] A representation of the examination region within the meaning of the present disclosure may be an MRI image, a computer tomography, an ultrasound image or the like, or the representation of the examination region may be generated from one or more MRI images, computer tomography, ultrasound images or the like.

[0030] A representation of the examination area within the meaning of the present disclosure can be a representation in the spatial space (image space) or a representation in the frequency space.

[0031] In a spatial representation, also referred to in this description as a spatial representation, the examination area is typically represented by a large number of image elements (pixels or voxels), which can be arranged in a grid, for example, with each image element representing a part of the examination area. A widely used format in radiology for storing and processing spatial representations is the DICOM format. DICOM (Digital Imaging and Communications in Medicine) is an open standard for storing and exchanging information in medical image data management.

[0032] In a representation in frequency space, also referred to in this description as a frequency-space representation or frequency-space representation, the area under investigation is represented by a superposition of fundamental oscillations. For example, the area under investigation can be represented by a sum of sine and cosine functions with different amplitudes, frequencies, and phases. The amplitudes and phases can be plotted as a function of frequency, for example in a two- or three-dimensional representation. Typically, the lowest frequency (origin) is placed in the center. The further one moves away from this center, the higher the frequencies. Each frequency can be assigned an amplitude, with which the frequency is represented in the frequency-space representation, and a phase, which indicates the extent to which the respective oscillation is shifted compared to a sine or cosine oscillation.

[0033] A representation in spatial space can be converted (transformed) into a representation in frequency space, for example, using a Fourier transform. Conversely, a representation in frequency space can be converted (transformed) into a representation in spatial space, for example, using an inverse Fourier transform.

[0034] Details about position-space representations and frequency-space representations and their respective conversion into each other are described in numerous publications, see for example: https: / / see.stanford.edu / materials / lsoftaee261 / book-fall-07.pdf.

[0035] With the help of a trained machine learning model, a synthetic representation of the examination area of ​​the examination object in a second state can be generated based on at least one representation of an examination area of ​​an examination object in a first state. The at least one representation on the basis of which the (trained) machine learning model generates the synthetic representation is also referred to in this description as the input representation.

[0036] The synthetic representation typically represents the same area of ​​investigation of the same object of investigation as the at least one input representation on whose generation the synthetic representation is based. The input representation represents the area of ​​investigation in a first state; the synthetic representation represents the area of ​​investigation in a second state. The first state and the second state are different states.

[0037] The synthetic representation represents the area of ​​investigation at a different time than the input representation.

[0038] The machine learning model is configured and trained to generate, based on at least one input representation of an examination region of an examination subject, which represents the examination region at a first time before and / or after the application of a quantity of contrast agent, a synthetic representation that represents the examination region at a second time after the application of the quantity of contrast agent. Such a trained machine learning model can be used, for example, as described in WO2021 / 052896A1, to reduce the amount of time an examination subject must spend in an MRI or CT scanner.

[0039] A "machine learning model" can be understood as a computer-implemented data processing architecture. The model can receive input data and produce output data based on this input data and model parameters. Through training, the model can learn a relationship between the input data and the output data. During training, model parameters can be adjusted to produce a desired output for a given input.

[0040] When training such a model, the model is presented with training data from which it can learn. The trained machine learning model is the result of the training process. The training data includes input data and the correct output data (target data) that the model is supposed to generate based on the input data. During training, patterns are recognized that map the input data to the target data.

[0041] During the training process, the input data of the training data is fed into the model, and the model generates output data. The output data is compared with the target data. Model parameters are modified so that the deviations between the output data and the target data are reduced to a (defined) minimum. An optimization method such as a gradient descent method can be used to modify the model parameters to reduce the deviations.

[0042] The deviations can be calculated using an error function (engl.: loss function ) can be quantified. Such an error function can be used to calculate an error ( loss ) for a given pair of output and target data. The goal of the training process may be to modify (adjust) the parameters of the machine learning model so that the error is reduced to a (defined) minimum for all pairs in the training dataset.

[0043] For example, if the output and target data are numbers, the error function can be the absolute difference between these numbers. In this case, a high absolute error may indicate that one or more model parameters need to be significantly changed.

[0044] For example, for output data in the form of vectors, difference metrics between vectors such as the mean square error, a cosine distance, a norm of the difference vector such as a Euclidean distance, a Chebyshev distance, an Lp norm of a difference vector, a weighted norm, or another type of difference metric between two vectors can be chosen as the error function.

[0045] For higher-dimensional outputs, such as two-dimensional, three-dimensional, or higher-dimensional outputs, an element-wise difference metric can be used. Alternatively or additionally, the output data can be transformed, e.g., into a one-dimensional vector, before calculating an error value.

[0046] In the present case, the machine learning model is trained using training data to generate a synthetic representation of an examination area of ​​an examination object in a second state based on at least one representation of the examination area of ​​the examination object in a first state.

[0047] The training data includes a set of input data and target data for each of a plurality of study objects.

[0048] The term multitude means at least ten, preferably more than one hundred.

[0049] Each set of input data and target data comprises at least one input representation of an examination area of ​​the examination object in a first state as input data. Each set of input data and target data further comprises a target representation of the examination area of ​​the examination object in a second state and a transformed target representation as target data.

[0050] The area of ​​investigation is usually the same for all objects under investigation.

[0051] The transformed target representation represents at least part of the examination area of ​​the examination object in the second state.

[0052] In the event that the target representation represents the examination area of ​​the examination object in the spatial space, the transformed target representation represents at least part of the examination area of ​​the examination object in the frequency space.

[0053] In the event that the target representation represents the examination area of ​​the examination object in the frequency domain, the transformed target representation represents at least part of the examination area of ​​the examination object in the spatial domain.

[0054] The transformed target representation in the frequency domain can be generated from a target representation in the spatial domain, for example by Fourier transformation.

[0055] The transformed target representation in the spatial domain can be generated from a target representation in the frequency domain, for example by inverse Fourier transformation.

[0056] When training the machine learning model, at least one input representation of the study area is fed into the machine learning model for each of the multiple study objects. The model generates a synthetic representation based on the at least one input representation of the study area and model parameters.

[0057] In case the at least one input representation supplied to the model represents the study area in the spatial space, the synthetic input representation preferably (but not necessarily) also represents the study area in the spatial space.

[0058] In case the at least one input representation supplied to the model represents the study area in the frequency domain, the synthetic representation preferably (but not necessarily) also represents the study area in the frequency domain.

[0059] However, the machine learning model can also be configured and trained to generate a synthetic representation of the examination area based on the at least one input representation of the examination area, which i) represents the examination area in frequency space if the at least one representation represents the examination area in spatial space, or ii) represents the examination area in spatial space if the at least one representation represents the examination area in frequency space. In other words, the machine learning model can be configured and trained to (among other things) perform a transformation from spatial to frequency space or vice versa.

[0060] A transformed synthetic representation is generated from or to the synthetic representation. If the synthetic representation represents the examination region in a second state in the spatial domain, the transformed synthetic representation represents at least a portion of the examination region in the second state in the frequency domain. If the synthetic representation represents the examination region in the second state in the frequency domain, the transformed synthetic representation represents at least a portion of the examination region in the second state in the spatial domain. The generation of the transformed synthetic representation can occur by transforming the synthetic representation and / or the machine learning model can be configured and trained to generate a transformed synthetic representation based on the at least one input representation.

[0061] Using an error function, the deviations i) between at least part of the synthetic representation and at least part of the target representation and ii) between at least part of the transformed synthetic representation and at least part of the transformed target representation are quantified.

[0062] The error function can have two terms: a first term for quantifying the deviations between at least part of the synthetic representation and at least part of the target representation, and a second term for quantifying the deviations between at least part of the transformed synthetic representation and at least part of the transformed target representation. The terms can be added together in the error function, for example. The terms can be weighted differently in the error function. The following equation provides an example of a (total) error function. L to quantify the deviations: L = λ 1 ⋅ L 1 + λ 2 ⋅ L 2

[0063] This is L the (total) error function, L 1 a term that represents the deviations between the synthetic representation and the target representation, L2 is a term that quantifies the deviations between the transformed synthetic representation and the transformed target representation, and λ 1 and λ 2 are weighting factors that can, for example, take values ​​between 0 and 1 and give the two terms a different weight in the error function. It is possible that the weighting factors are kept constant or varied during training of the machine learning model.

[0064] Examples of error functions that can generally be used to implement the present invention are L1 error function ( L1 loss ), L2 error function ( L2 loss ), Lp error function (Lp loss), structural similarity index measure ( structural similarity index measure (SSIM)), VGG error function (VGG loss), perception error function ( perceptual loss ) or a combination of the above-mentioned functions or other error functions. Further details on error functions can be found in the scientific literature (see, for example, R. Mechrez et al.: The ContextualLoss for Image Transformation with Non-Aligned Data, 2018, arXiv:1803.02077v4; H. Zhao et al.: Loss Functions for Image Restoration with Neural Networks, 2018, arXiv:1511.08861v3).

[0065] Fig. 1 shows an example and schematic of the process of training the machine learning model. Training is carried out using training data. In Fig. 1 Training data TD for an examination object is shown. The training data TD comprises, as input data, a first input representation R1 of an examination region of the examination object and a second input representation R2 of the examination region of the examination object. The first input representation R1 and the second input representation R2 represent the examination region in spatial space. The first input representation R1 represents the examination region in a first state, for example, at a first time before or after the application of a contrast agent. The second input representation R2 represents the examination region in a second state, for example, at a second time after the application of a contrast agent.

[0066] The training data TD further includes a target representation TR as target data. The target representation TR also represents the examination region in spatial space. The target representation TR represents the examination region in a third state, for example, at a third time after the application of a contrast agent.

[0067] The machine learning model MLM is trained to predict the target representation TR based on the first input representation R1 and the second input representation R2, as well as on the model parameters MP. The first input representation R1 and the second input representation R2 are fed to the machine learning model as input data. The machine learning model is configured to generate a synthetic representation SR.

[0068] The synthetic representation SR is discussed in Fig. 1 shown example using a transformationT (e.g. a Fourier transform) a transformed synthetic representation SR T< Analogously, the target representation TR is transformed into T a transformed target representation TR T< The transformed synthetic representation SR T< and the transformed target representation TR T< are frequency domain representations of the investigation area in the third state.

[0069] A first error function L 1 is used to quantify the deviations between the synthetic representation SR and the target representation TR. A second error function Lz is used to quantify the deviations between the transformed synthetic representation SR T< and the transformed target representation TR T< to quantify. The error functions L 1 and L2 calculated errors are combined into a total error function L to form a total error (e.g. by addition).

[0070] Model parameters MP are modified to reduce the total error. Reducing the total error can be achieved using an optimization method, for example, a gradient method.

[0071] The process is repeated for a large number of test objects until the total error reaches a predefined (desired) minimum.

[0072] In a preferred embodiment, the transformed synthetic representation and the transformed target representation are each reduced to one or more predefined parts. This is illustrated schematically in Fig. 2 shown.

[0073] The Fig. 2 The process shown corresponds to the process described in Fig. 1 shown, with the following differences: The transformed target representation TR T< is limited to a defined proportion of TR T , P < reduced. The representation TR T , P < is a part of the transformed target representation TR. The representation TR T , P < can be determined by a function P from the transformed target representation TR T< be generated, for example, by setting all values ​​of frequencies outside the predefined part to zero. Analogously, the transformed synthetic representation SR T< is limited to a defined proportion of SR T , P < reduced. The proportions to which the transformed target representation TR T< and the transformed synthetic representation SR T< reduced, usually correspond to each other, ie they usually affect the same frequencies. In the Fig. 2 The process shown is both the transformed target representation TR T< as well as the transformed synthetic representation SR T< reduced to a portion that contains the low frequencies in an area (e.g. rectangle or square) around the center (see the dashed white frames). In this area, predominantly contrast information is encoded).

[0074] In the Fig. 2 In the example shown, the entire frequency range is not used in the calculation using the error function L2. The representations in the frequency space are reduced to a range of low frequencies; the higher frequencies are discarded. The low frequency range predominantly encodes contrast information, while the higher frequency range predominantly encodes information about fine structures. This means that the machine learning model in the Fig. 2 In the example shown, in addition to generating the synthetic representation SR, the brain is trained to correctly reproduce the low frequencies in particular, thus placing particular emphasis on contrast information.

[0075] Fig. 3 shows another example where the entire frequency range is not considered, but rather a focus is placed on a defined frequency range. Fig. 3 The process shown corresponds to the process described in Fig. 2 is shown, with the difference that the function Pensures that the transformed target representation TR T< and the transformed synthetic representation SR T< are reduced to a higher frequency component, while low frequencies are discarded (the low frequency values ​​in this example are set to zero, which is represented by the color black). This means that the machine learning model in the Fig. 3 In the example shown, in addition to generating the synthetic representation SR, the system is trained to correctly reproduce the higher frequencies in particular, thus placing particular emphasis on the correct reproduction of fine structures.

[0076] It should be noted that the part to which the frequency space is reduced may also include other than those in Fig. 2 and Fig. 3 can take on the shapes and sizes shown. Furthermore, multiple parts can be selected and the frequency-domain representations restricted to multiple parts (ranges). For example, it is possible to divide the frequency domain into different ranges and generate frequency-domain representations of the study area for several or all ranges, each of which contains only the frequencies of the respective range. In this way, different frequency ranges can be weighted / considered differently during training.

[0077] In the Fig. 2 and Fig. 3 In the examples shown, the representations R1, R2, TR, and SR are representations in the spatial domain. It is also possible that one or more of these representations are frequency-domain representations.

[0078] Are the transformed synthetic representation SR T< and the transformed target representation TR T< Representations in spatial space, a part (or several parts) of the representations can also be selected, and the representations can be reduced to this part (these parts). In such a case, the error calculation focuses on features in spatial space that are found in the selected part (or parts).

[0079] Fig. 4 , Fig. 5 , Fig. 6 , Fig. 7 and Fig. 8 show further embodiments of the training procedure.

[0080] In the Fig. 4 In the embodiment shown, the training data for each examination object comprises at least one input representation R1 of an examination area in a first state and a target representation TR of the examination area in a second state. The at least one input representation R1 and the target representation TR can each be a representation in the spatial domain or a representation in the frequency domain. In the embodiment shown in Fig. 4 In the example shown, the input representation R1 and the target representation TR are representations in spatial space. The machine learning model MLM is trained to predict the target representation TR based on at least one input representation R1. Using a transformation T, a transformed input representation R2 is generated based on the input representation R1. Analogously, a transformed target representation TR is generated using the transformation T. T< based on the target representation TR. In the Fig. 4 In the example shown, the transformed input representation R2 and the transformed target representation TR T< Representations of the examination area in the frequency domain. The transformed input representation R2 represents the examination area (like the input representation R1) in the first state; the transformed target representation TR T< represents the investigation area (like the target representation TR) in the second state.

[0081] In Fig. 4 It is shown that the input representation R1 is transformed by an inverse transformation T -1< can be obtained based on the transformed input representation R2. Analogously, a target representation TR can be obtained by the inverse transformation T -1< based on the transformed target representation TR T< The inverse transformation T-1< is the inverse transformation of the transformation T, which is denoted by the index -1. The reference to the inverse transformation is intended to clarify that the training data must contain at least one input representation (R1 or R2) and one target representation (TR or TR7); the other (R2 or R1, or TR T< or TR) can be obtained from the given representation by transformation or inverse transformation. This applies generally and not only to the Fig. 4 illustrated embodiment.

[0082] In the Fig. 4 In the embodiment shown, both the input representation R1 and the transformed input representation R2 are fed to the machine learning model MLM. The machine learning model MLM is configured to generate both a synthetic representation SR and a transformed synthetic representation SR T< to generate.

[0083] Using a first error function L 1, the deviations between at least part of the synthetic representation SR and at least part of the target representation TR are quantified. Using a second error function L 2, the deviations between at least part of the transformed synthetic representation SR T< and at least part of the transformed target representation TR T< quantified. The Fig. 4 The dashed white frames drawn in the transformed representations are intended to express that the calculation of the error is carried out using the error function as in relation to Fig. 3 described does not have to be based on the entire frequency domain representation, but that the transformed representations can be reduced to one frequency range (or several frequency ranges). Analogously, the spatial domain representations SR and TR can be reduced to one part (or several parts) in the error calculation (also represented by the dashed white frames). This applies analogously to the other embodiments and not only to the Fig. 4 illustrated embodiment.

[0084] The error functions L 1 and L 2 calculated errors are combined into a total error function L to form a total error (e.g. by addition).

[0085] Model parameters MP are modified to reduce the overall error. This can be achieved using an optimization method, such as a gradient method.

[0086] The process is repeated for a large number of test objects until the total error reaches a predefined (desired) minimum.

[0087] The synthetic representation SR and the transformed synthetic representation SR T< must be convertible into each other in the same way as the target representation TR and the transformed target representation TR T< . It is possible to introduce a further error function that evaluates the quality of such a transformation. It is therefore possible to generate a transformed synthetic representation from the synthetic representation SR by the transformation T and to estimate the deviations between this transformed synthetic representation and the transformed synthetic representation generated by the machine learning model using a third error function. L 3. Alternatively or additionally, it is possible to use the inverse transformation T-1< from the transformed synthetic representation SR T< to generate a synthetic representation and to estimate the deviations between this synthetic representation generated by inverse transformation and the synthetic representation generated by the machine learning model by means of a third error function L 3. This is shown schematically in Fig. 5 using the example of generating a transformed synthetic representation SR T *< by transformation T from the synthetic representation SR. In the Fig. 5 The example shown quantifies the error function L 3 the deviations between the transformed synthetic representation SR generated by transformation T *< and the transformed synthetic representation SR generated by the machine learning model MLM T< . The error functions L 1 , L 2 and L3 are used in a total error function L combined (e.g. by addition). In the total error function L The individual terms can carry different weights, whereby the weights can be kept constant or varied over the course of the training.

[0088] A variation of the Fig. 4 and Fig. 5 The advantage of the embodiments shown is that the machine learning model MLM is not fed both input representations R1 and R2, but only one of the two. The machine learning model can be configured and trained to generate the other one.

[0089] Fig. 6 schematically shows another embodiment of the training method. In this embodiment, two machine learning models are used, a first model MLM1 and a second model MLM2. The first machine learning model MLM1 is trained to generate a target representation TR and / or a transformed target representation TR from an input representation R1 and / or R2. T< The second machine learning model MLM2 is trained to predict the target representation SR and / or the transformed target representation SR T< to predict (reconstruct) the original input representation R1 and / or R2 again. The machine learning models thus perform a cycle that can improve the prediction quality of the first model MLM1 (which is used in the subsequent prediction).

[0090] The first machine learning model MLM1 is supplied with at least one input representation R1 and / or at least one transformed input representation R2 as input data. The first model MLM1 is configured to use a synthetic representation SR and / or a transformed synthetic representation SR T< based on the input data and model parameters MP1. A transformed synthetic representation SR T< can also be achieved through transformation T the synthetic representation SR. A synthetic representation SR can also be generated by inverse transformation T -1< of the transformed synthetic representation SR T< be generated.

[0091] Deviations between the synthetic representation SR and the target representation TR can be quantified using an error function L 1 1<. Deviations between the transformed synthetic representation SR T< and the transformed target representation TR T< can be determined using an error function L 2 1< can be quantified.

[0092] Deviations between a transformation T obtained synthetic representation and the synthetic representation SR generated by the model MLM1 can be compared using an error function L 3 1< can be quantified (in Fig. 6 not shown). Alternatively or additionally, deviations between a transformed synthetic representation obtained by transformation T and the transformed synthetic representation SR generated by the model MLM1 T< using an error function L 4 1< can be quantified (in Fig. 6 not shown).

[0093] The error functions L 1 1< , L 2 1< and, if applicable, L 3 1< and / or L4 1< can be combined to form a total error function (in Fig. 6 not shown).

[0094] Model parameters MP1 can be modified to reduce the overall error.

[0095] The synthetic representation SR and / or the transformed synthetic representation SR T< is / are fed to the second machine learning model MLM2. The second model MLM2 is configured to reconstruct (predict) the first input representation R1 and / or the second input representation R2. The second model MLM2 is configured to generate a predicted first input representation R1* and / or a predicted second input representation R2* based on the synthetic representation SR, the transformed synthetic representation SR T< and based on model parameters MP2. A second input representation R2* can also be generated by transformation T of the first input representation R1* and / or a first input representation R1* can also be generated by inverse transformation T -1< of the second input representation R2*.

[0096] Deviations between the predicted first input representation R1* and the first input representation R1 can be estimated using an error function L 1 2< can be quantified. Deviations between the predicted second input representation R2* and the second input representation R2 can be quantified using an error function L 2 2< can be quantified.

[0097] Deviations between an inverse transformation T -1< and the input representation R1* generated by the model MLM2 can be compared using an error functionL 3 2< can be quantified (in Fig. 6 not shown). Alternatively or additionally, deviations between an input representation R2* obtained by transformation and the input representation R2* generated by the model MLM2 can be calculated using an error function L 4 2< can be quantified (in Fig. 6 not shown).

[0098] The error functions L 1 2< , L 2 2< and, if applicable, L 3 2< and / or L 4 2< can be combined to form a total error function (in Fig. 6 not shown).

[0099] Model parameters MP2 can be modified to reduce the overall error.

[0100] Fig. 7 shows schematically another example of a method for training a machine learning model. Fig. 8 shows the use of the trained machine learning model for prediction. The Fig. 7 and Fig. 8 The example shown concerns the acceleration of a magnetic resonance imaging examination of the liver of an examination subject using a hepatobiliary contrast agent.

[0101] An example of a hepatobiliary contrast agent is the disodium salt of gadoxetic acid (Gd-EOB-DTPA disodium), which is described in US Patent No. 6,039,931A and is commercially available under the brand names Primovist ®< and Eovist ®<.

[0102] A hepatobiliary contrast agent can be used to detect tumors in the liver. The blood supply to healthy liver tissue is primarily via the portal vein, while the hepatic artery supplies most primary tumors. Accordingly, after intravenous bolus injection of a contrast agent, a time delay can be observed between the signal enhancement of the healthy liver parenchyma and the tumor.

[0103] In addition to malignant tumors, benign lesions such as cysts, hemangiomas, and focal nodular hyperplasia (FNH) are frequently found in the liver. For appropriate treatment planning, these lesions must be differentiated from malignant tumors. Primovist ®< can be used to detect benign and malignant focal liver lesions. It provides information about the nature of these lesions using T1-weighted MRI. Differentiation is based on the different blood supply to the liver and tumor and the temporal course of contrast enhancement.

[0104] With the contrast enhancement achieved by Primovist® during the inflow phase, typical perfusion patterns are observed, providing information for characterizing lesions. The visualization of vascularity helps characterize lesion types and determine the spatial relationship between tumor and blood vessels.

[0105] In T1-weighted MRI images, Primovist ® leads to a significant signal enhancement in the healthy liver parenchyma < 10-20 minutes after injection (in the hepatobiliary phase), while lesions containing no or few hepatocytes, e.g. metastases or moderately to poorly differentiated hepatocellular carcinomas (HCCs), appear as darker areas.

[0106] Tracking the distribution of contrast agent over time offers a good opportunity for the detection and differential diagnosis of focal liver lesions. However, the examination takes a comparatively long time. During this time, patient movement should be avoided as much as possible to minimize motion artifacts in the MRI images. This prolonged restriction of movement can be uncomfortable for the patient.

[0107] Accordingly, WO2021 / 052896A1 already proposes not to generate MRI images of a patient's liver during the hepatobiliary phase using measurement technology, but to predict them on the basis of MRI images from one or more previous phases.

[0108] Fig. 7 and Fig. 8 demonstrate an improved approach compared to the method described in WO2021 / 052896A1: The training of the machine learning model (MLM) is based on training data. The training data includes, for each examination object of a plurality of examination objects, one or more input representations (R1, R2) of the liver or part of the liver during a first time period before and / or after the application of a hepatobiliary contrast agent and at least one target representation (TR) of the liver or part of the liver during a second time period after the application of the hepatobiliary contrast agent.

[0109] The input representations R1,2 and the target representation TR are typically generated using an MRI scanner. The hepatobiliary contrast agent can, for example, be administered as a weight-adjusted bolus into a vein in the arm of the subject.

[0110] The input representations R1,2 may represent the liver or part of the liver during the native phase, the arterial phase, the portal venous phase and / or the transition phase (as described in WO2021 / 052896A1 and the references cited therein).

[0111] The target representation shows the liver or part of the liver during the hepatobiliary phase, for example, 10 to 20 minutes after the application of the hepatobiliary contrast agent.

[0112] In the Fig. 7 The input representations R1,2 and the target representation TR shown are representations of the liver or part of the liver in spatial space. The input representations R1,2 represent the examination area in different states, namely at different times in the native phase, the arterial phase, the portal venous phase, and / or the transition phase. The target representation TR represents the examination area in a further state, namely at a time in the hepatobiliary phase.

[0113] The input representations R1,2 are fed (possibly after preprocessing, which may include, for example, motion correction, co-registration and / or color space conversion) to a machine learning model MLM.

[0114] The machine learning model is configured to generate a synthetic representation SR based on the input representations R1,2 and on model parameters MP.

[0115] From the synthetic representation SR a transformed synthetic representation SR is created by means of a transformation T (e.g. a Fourier transformation). T< generated.

[0116] The transformed synthetic representation SR T< is a representation of the liver or part of the liver during the hepatobiliary phase in frequency space.

[0117] The transformed synthetic representation SR T< is done using a function P reduced to a predefined proportion, thereby creating a proportionally transformed synthetic representation SR T,P< generated. The function P reduces the transformed synthetic representation SR T< to a predefined frequency range. In this example, frequencies outside a circle with a defined radius around the origin of the transformed synthetic representation SR T< set to zero, leaving only the frequencies within the circle. In other words, the higher frequencies, which encode fine-structure information, are erased, leaving only the lower frequencies, which encode contrast information.

[0118] The same procedure is followed with the target representation TR: From the target representation TR, a transformation T (e.g. a Fourier transform) a transformed target representation TR T< The transformed target representation TR T< is a representation of the liver or part of the liver during the hepatobiliary phase in the frequency domain. The transformed target representation TR T< is done using the function P reduced to a predefined proportion, whereby a proportional transformed target representation TR T , P < generated. The function P2 reduces the transformed target representation TR T< to the same frequency range as the transformed synthetic representation SR T< In this example, frequencies outside a circle with a defined radius around the origin of the transformed target representation TR T< set to zero, leaving only the frequencies within the circle. In other words, the higher frequencies, which encode fine-structure information, are erased, leaving only the lower frequencies, which encode contrast information.

[0119] To evaluate the prediction quality of the machine learning model, an error is calculated using an error function LIn this example, the error function is L consists of two terms, a first error function L 1 and a second error function L 2 .

[0120] The first error function L 1 quantifies the deviations between the synthetic representation SR and the target representation TR.

[0121] The second error function L 2 quantifies the deviations between the proportional transformed synthetic representation SR T< and the proportional transformed target representation TR T< .

[0122] The terms for L 1 and L 2 can be included in the total error function L for example, they can be given weighting factors and added together as shown in equation 1 above.

[0123] In an optimization procedure, e.g. a gradient method, the model parameters MP can be modified with a view to reducing the error calculated using the total error function L.

[0124] The described process is repeated for the remaining objects of the multitude of objects to be examined. The training can be terminated when the total error function L calculated errors reach a defined minimum, ie the prediction quality reaches a desired level.

[0125] Fig. 8 shows the use of the trained machine learning model for prediction. The model can be used as in relation to Fig. 7 The trained machine learning model MLM t< is fed one or more input representations R1,2* of the liver or part of the liver of a new study object. The term "new" means that the input representations R1,2* have not already been used in training the model.

[0126] The input representations R1,2* represent the liver or part of the liver of the new examination subject before / and or after the application of a hepatobiliary contrast agent, which may, for example, have been administered in the form of a bolus into an arm vein of the new examination subject.

[0127] The input representations R1,2* represent the liver or part of the liver of the new object of investigation in the native phase, the arterial phase, the portal venous phase and / or the transition phase.

[0128] The input representations R1,2* represent the liver or part of the liver of the new object of investigation in spatial space.

[0129] The trained machine learning model is configured and trained to predict a synthetic representation SR* based on the input representations R1,2*.

[0130] The synthetic representation SR* represents the liver or part of the liver during the hepatobiliary phase, for example, 10 to 20 minutes after the application of the hepatobiliary contrast agent.

[0131] For example, the machine learning models according to the present disclosure may be an artificial neural network or may include one or more such artificial neural networks.

[0132] An artificial neural network comprises at least three layers of processing elements: a first layer with input neurons (nodes), an Nth layer with at least one output neuron (node), and N-2 inner layers, where N is a natural number and greater than 2.

[0133] The input neurons serve to receive the input representations. Typically, there is one input neuron for each pixel or voxel of an input representation if the representation is a spatial representation in the form of a raster graphic, or one input neuron for each frequency present in the input representation if the representation is a frequency-domain representation. Additional input neurons may be present for additional input values ​​(e.g., information about the examination region, the object under investigation, conditions that prevailed when the input representation was generated, information about the state represented by the input representation, and / or information about the time or time period at / in which the input representation was generated).

[0134] The output neurons can be used to output a synthetic representation that represents the area of ​​interest in a different state.

[0135] The processing elements of the layers between the input neurons and the output neurons are connected in a predetermined pattern with predetermined connection weights.

[0136] Preferably, the artificial neural network is a so-called convolutional neural network (CNN for short) or it includes one.

[0137] A CNN usually consists essentially of filters (convolutional layer) and aggregation layers (pooling layer) that repeat alternately, and at the end of one or more layers of "normal" fully connected neurons (dense / fully connected layer).

[0138] The neural network can be trained, for example, using a backpropagation method. The goal is to achieve the most reliable mapping of the input representation(s) to the synthetic representation. The quality of the prediction is described by an error function. The goal is to minimize the error function. In the backpropagation method, an artificial neural network is trained by changing the connection weights.

[0139] In the trained state, the connection weights between the processing elements contain information regarding the dynamics of the relationship between the input representation(s) and the synthetic representation, which can be used to predict a synthetic representation of the new study object's study domain based on one or more representations of the study object's study domain. The term "new" means that representations of the new study object's study domain were not already used in training the machine learning model.

[0140] A cross-validation method can be used to split the data into training and validation sets. The training set is used for backpropagation training of the network weights. The validation set is used to test the predictive accuracy of the trained network when applied to unknown (new) data.

[0141] The artificial neural network may have an autoencoder architecture; e.g., the artificial neural network may have an architecture such as the U-Net (see e.g. O. Ronneberger et al.: U-net: Convolutional networks for biomedical image segmentation, International Conference on Medical image computing and computer-assisted intervention, pages 234-241, Springer, 2015, https: / / doi.org / 10.1007 / 978-3-319-24574-4_28).

[0142] The artificial neural network can be a Generative Adversarial Network (GAN).

[0143] The artificial neural network can be a Regularized Generative Adversarial Network (see e.g. Q. Li et al.: RegGAN: An End-to-End Network for Building Footprint Generation with Boundary Regularization, Remote Sens. 2022, 14, 1835).

[0144] The artificial neural network can be a Conditional Adversarial Network (see e.g. P. Isola et al.: Image-to-Image Translation with Conditional Adversarial Networks, arXiv:1611.07004 [cs.CV]).

[0145] The artificial neural network can be a transformer network (see, for example, D. Karimi et al.: Convolution-Free Medical Image Segmentation using Transformers, arXiv:2102.13645 [eess.IV]).

[0146] Fig. 9 shows, by way of example and schematically, a computer system according to the present disclosure and which can be used for training the machine learning model and / or for using the trained machine learning model for prediction.

[0147] A "computer system" is an electronic data processing system that processes data using programmable computing instructions. Such a system typically includes a "computer," the unit containing a processor for performing logical operations, and peripherals.

[0148] In computer technology, "peripherals" refers to all devices connected to a computer that serve to control the computer and / or act as input and output devices. Examples include monitors, printers, scanners, mice, keyboards, drives, cameras, microphones, speakers, etc. Internal connectors and expansion cards are also considered peripherals in computer technology.

[0149] The Fig. 9 The computer system (1) shown comprises an input unit (10), a control and computing unit (20) and an output unit (30).

[0150] The control and computing unit (20) serves to control the computer system (1), to coordinate the data flows between the units of the computer system (1) and to carry out calculations.

[0151] The control and computing unit (20) is configured: to cause the receiving unit (10) to receive at least one input representation of an examination area of ​​a new examination object, to input the received at least one input representation into a trained machine learning model, wherein the trained machine learning model has been trained as described in this description, to receive a synthetic representation of the examination area of ​​the new examination object from the machine learning model, to cause the output unit (30) to output the synthetic representation and / or to store it and / or to transmit it to a separate computer system.

[0152] Fig. 10 shows an exemplary and schematic illustration of a further embodiment of the computer system according to the invention.

[0153] The computer system (1) comprises a processing unit (21) connected to a memory (22). The processing unit (21) and the memory (22) form a control and calculation unit as described in Fig. 9 is shown.

[0154] The processing unit (21) (engl.: processing unit) may comprise one or more processors alone or in combination with one or more memories. The processing unit (21) may be ordinary computer hardware capable of processing information such as digital images, computer programs and / or other digital information. The processing unit (21) typically consists of an arrangement of electronic circuits, some of which may be embodied as an integrated circuit or as a plurality of interconnected integrated circuits (an integrated circuit is sometimes referred to as a "chip"). The processing unit (21) may be configured to execute computer programs that may be stored in a main memory of the processing unit (21) or in the memory (22) of the same or another computer system.

[0155] The memory (22) may be ordinary computer hardware capable of storing information such as digital images (e.g. representations of the examination area), data, computer programs and / or other digital information either temporarily and / or permanently. The memory (22) may comprise volatile and / or non-volatile memory and may be permanently installed or removable. Examples of suitable memories are RAM (Random Access Memory), ROM (Read-Only Memory), a hard disk, flash memory, a removable computer diskette, an optical disc, magnetic tape or a combination of the above. Optical discs may include read-only compact discs (CD-ROM), read / write compact discs (CD-R / W), DVDs, Blu-ray discs and the like.

[0156] In addition to the memory (22), the processing unit (21) can also be connected to one or more interfaces (11, 12, 31, 32, 33) for displaying, transmitting, and / or receiving information. The interfaces can comprise one or more communication interfaces (32, 33) and / or one or more user interfaces (11, 12, 31). The one or more communication interfaces can be configured to send and / or receive information, e.g., to and / or from an MRI scanner, a CT scanner, an ultrasound camera, other computer systems, networks, data storage devices, or the like. The one or more communication interfaces can be configured to transmit and / or receive information via physical (wired) and / or wireless communication connections.The one or more communication interfaces may include one or more interfaces for connecting to a network, e.g., using technologies such as cellular, Wi-Fi, satellite, cable, DSL, fiber optic, and / or the like. In some examples, the one or more communication interfaces may include one or more short-range communication interfaces configured to connect devices using short-range communication technologies such as NFC, RFID, Bluetooth, Bluetooth LE, ZigBee, infrared (e.g., IrDA), or the like.

[0157] The user interfaces may comprise a display (31). A display (31) may be configured to display information to a user. Suitable examples include a liquid crystal display (LCD), a light-emitting diode display (LED), a plasma display (PDP), or the like. The user input interface(s) (11, 12) may be wired or wireless and may be configured to receive information from a user into the computer system (1), e.g., for processing, storage, and / or display. Suitable examples of user input interfaces include a microphone, an image or video capture device (e.g., a camera), a keyboard or keypad, a joystick, a touch-sensitive surface (separate from or integrated with a touchscreen), or the like.In some examples, the user interfaces may include automatic identification and data capture (AIDC) technology for machine-readable information. This may include barcodes, radio frequency identification (RFID), magnetic stripes, optical character recognition (OCR), integrated circuit cards (ICC), and the like. The user interfaces may further include one or more interfaces for communicating with peripheral devices such as printers and the like.

[0158] One or more computer programs (40) can be stored in the memory (22) and executed by the processing unit (21), which is thereby programmed to perform the functions described in this description. The retrieval, loading, and execution of instructions of the computer program (40) can occur sequentially, such that one instruction is retrieved, loaded, and executed at a time. However, the retrieval, loading, and / or execution can also occur in parallel.

[0159] The machine learning model according to the invention can also be stored in the memory (22).

[0160] The system according to the invention can be designed as a laptop, notebook, netbook and / or tablet PC; it can also be a component of an MRI scanner, a CT scanner or an ultrasound diagnostic device.

Claims

1. Method for training a machine-learning model (MLM) for producing synthetic radiological images based on measured radiological images, wherein the method comprises: - receiving and / or providing training data (TD), the training data (TD) for each examination object of a multiplicity of examination objects comprising a set of input data and target data, wherein each set comprises ∘ at least one input representation (R1, R2) of an examination region of the examination object in a first state as input data and ∘ a target representation (TR) of the examination region of the examination object in a second state and also a transformed target representation (TRT) as target data, wherein the transformed target representation (TRT) represents at least one part of the examination region of the examination object ∘ in frequency space, if the target representation (TR) represents the examination region of the examination object in real space, or ∘ in real space, if the target representation (TR) represents the examination region of the examination object in frequency space, - training a machine-learning model (MLM), the machine-learning model (MLM) being configured to produce a synthetic representation (SR) of the examination region of the examination object on the basis of at least one input representation (R1, R2) and model parameters (MP), with the at least one input representation (R1, R2) representing the examination region in a first time interval before and / or after the application of an amount of a contrast agent, the synthetic representation (SR) representing the examination region in a second time interval after the application of the amount of contrast agent and the second time interval following the first time interval in time, wherein training for each examination object of the multiplicity of examination objects comprises: ∘ supplying the at least one input representation (R1, R2) to the machine-learning model (MLM), ∘ receiving a synthetic representation (SR) of the examination region of the examination object from the machine-learning model (MLM), ∘ producing and / or receiving a transformed synthetic representation (SRT) on the basis of and / or relating to the synthetic representation (SR), the transformed synthetic representation (SRT) representing at least one part of the examination region of the examination object ▪ in frequency space, if the synthetic representation (SR) represents the examination region of the examination object in real space, or ▪ in real space, if the synthetic representation (SR) represents the examination region of the examination object in frequency space, ∘ quantifying the deviations i) between at least one part of the synthetic representation (SR) and at least one part of the target representation (TR) and ii) between at least one part (SRT,P) of the transformed synthetic representation (SRT) and at least one part (TRT,P) of the transformed target representation (TRT) by means of an error function (L), ∘ modifying model parameters (MP) with regards to reduced deviations, - outputting and / or storing the trained machine-learning model (MLMt) and / or model parameters (MP) and / or transmitting the trained machine-learning model (MLMt) and / or model parameters (MP) to a separate computer system.

2. Method according to Claim 1, wherein receiving and / or providing training data (TD) comprises: - producing a partial transformed target representation (TRT,P), the partial transformed target representation (TRT,P) being reduced to one or more parts of the transformed target representation (TRT), wherein the production and / or reception of a transformed synthetic representation (SRT) on the basis of and / or relating to the synthetic representation (SR) comprises: - producing a partial transformed synthetic representation (SRT,P), the partial transformed synthetic representation (SRT,P) being reduced to one or more parts of the transformed synthetic representation (SRT), wherein the quantification of the deviations between the transformed synthetic representation (SRT) and the transformed target representation (TRT) comprises: - quantifying the deviations between the partial transformed synthetic representation (SRT,P) and the partial transformed target representation (TRT,P).

3. Method according to Claim 1 or 2, wherein training for each examination object of the multiplicity of examination objects comprises: ∘ supplying the at least one input representation (R1, R2) to the machine-learning model (MLM), ∘ receiving the synthetic representation (SR) and a first transformed synthetic representation (SRT) of the examination region of the examination object from the machine-learning model (MLM), the first transformed synthetic representation (SRT) representing at least one part of the examination region of the examination object ▪ in frequency space, if the synthetic representation (SR) represents the examination region of the examination object in real space, or ▪ in real space, if the synthetic representation (SR) represents the examination region of the examination object in frequency space, ∘ producing a second transformed synthetic representation (SRT*) on the basis of the synthetic representation (SR) by means of a transformation (T), the second transformed synthetic representation (SRT*) representing at least one part of the examination region of the examination object ▪ in frequency space, if the synthetic representation (SR) represents the examination region of the examination object in real space, or ▪ in real space, if the synthetic representation (SR) represents the examination region of the examination object in frequency space, - quantifying the deviations i) between at least one part of the synthetic representation (SR) and at least one part of the target representation (TR), ii) between at least one part of the first transformed synthetic representation (SRT) and at least one part of the transformed target representation (TR*), and iii) between at least one part of the first transformed synthetic representation (SRT) and at least one part of the second transformed synthetic representation (SRT*) by means of an error function (L), - modifying model parameters (MP) with regards to reduced deviations.

4. Method according to any of Claims 1 to 3, wherein the machine-learning model (MLM) comprises a first machine-learning model (MLM1) and a second machine-learning model (MLM2), wherein the first machine-learning model (MLM1) is configured to produce a synthetic representation (SR) of the examination region of the examination object on the basis of at least one input representation (R1) and model parameters (MP1), wherein the second machine-learning model (MLM2) is configured to reconstruct the input representation (R1) on the basis of the synthetic representation (SR) of the examination region of the examination object, wherein training for each examination object of the multiplicity of examination objects comprises: ∘ producing a transformed input representation (R2) on the basis of the input representation (R1) by means of a transformation (T), the transformed input representation (R2) representing at least one part of the examination region of the examination object ▪ in frequency space, if the input representation (R1) represents the examination region of the examination object in real space, or ▪ in real space, if the input representation (R1) represents the examination region of the examination object in frequency space, ∘ supplying the at least one input representation (R1) to the first machine-learning model (MLM1), ∘ receiving a synthetic representation (SR) of the examination region of the examination object from the first machine-learning model (MLM1), ∘ producing and / or receiving a transformed synthetic representation (SRT) on the basis of and / or relating to the synthetic representation (SR), the transformed synthetic representation (SRT) representing at least one part of the examination region of the examination object ▪ in frequency space, if the synthetic representation (SR) represents the examination region of the examination object in real space, or ▪ in real space, if the synthetic representation (SR) represents the examination region of the examination object in frequency space, ∘ supplying the synthetic representation (SR) and / or the transformed synthetic representation (SRT) to the second machine-learning model (MLM2), ∘ receiving a predicted input representation (R1*) from the second machine-learning model (MLM2), ∘ producing and / or receiving a transformed predicted input representation (R2*) on the basis of and / or relating to the predicted input representation (R1*), the transformed predicted input representation (R2*) representing at least one part of the examination region of the examination object ▪ in frequency space, if the predicted input representation (R1*) represents the examination region of the examination object in real space, or ▪ in real space, if the predicted input representation (R1*) represents the examination region of the examination object in frequency space, ∘ quantifying the deviations i) between at least one part of the synthetic representation (SR) and at least one part of the target representation (TR), ii) between at least one part of the transformed synthetic representation (SRT) and at least one part of the transformed target representation (TRT), iii) between at least one part of the input representation (R1) and at least one part of the predicted input representation (R1*) and iv) between at least one part of the transformed input representation (R2) and at least one part of the transformed predicted input representation (R2*) by means of an error function, ∘ modifying model parameters (MP1, MP2) with regards to reduced deviations.

5. Method according to any of Claims 1 to 4, wherein the examination region is a liver or a part of a liver of a human being.

6. Method according to any of Claims 1 to 5, wherein each of the at least one input representation (R1, R2) is a representation of the examination region in real space, the target representation (TR) is a representation of the examination region in real space and the synthetic representation (SR) is a representation of the examination region in real space.

7. Method according to any of Claims 2 to 5, wherein the partial transformed synthetic representation (SRT,P) represents the examination region in frequency space, the partial transformed synthetic representation (SRT,P) being reduced to a frequency range of the transformed synthetic representation (SRT), with contrast information being encoded in the frequency range.

8. Method according to any of Claims 2 to 5, wherein the partial transformed synthetic representation (SRT,P) represents the examination region in frequency space, the partial transformed synthetic representation (SRT,P) being reduced to a frequency range of the transformed synthetic representation (SRT), with information relating to fine structures being encoded in the frequency range.

9. Computer-implemented method for producing a synthetic representation of an examination region of an examination object, wherein the method comprises: - receiving at least one input representation (R1,2*) of the examination region of the examination object in a first state, the at least one input representation (R1,2*) representing the examination region in a first time interval before and / or after the application of an amount of a contrast agent, - providing a trained machine-learning model (MLMt), • wherein the trained machine-learning model (MLMt) has been trained on the basis of training data (TD) to produce a synthetic representation (SR) of the examination region in a second state on the basis of at least one input representation (R1, R2) of an examination region of an examination object in a first state, • wherein the training data (TD) for each examination object of a multiplicity of examination objects comprise i) an input representation (R1) of the examination region, ii) a target representation (TR) and iii) a transformed target representation (TRT), • wherein the transformed target representation (TRT) represents at least one part of the examination region of the examination object in frequency space, if the target representation (TR) represents the examination region of the examination object in real space, or in real space, if the target representation (TR) represents the examination region of the examination object in frequency space, • wherein the machine-learning model (MLMt) has been trained for each examination object to minimize deviations between i) at least one part of the synthetic representation (SR) and at least one part of the target representation (TR) and ii) between at least one part of a transformed synthetic representation (SRT) and at least one part of the transformed target representation (TRT), - inputting the at least one received input representation (R1,2*) of the examination region of the examination object into the trained machine-learning model (MLMt), - receiving a synthetic representation (SR*) of the examination region of the examination object in the second state from the machine-learning model (MLMt), the synthetic representation (SR*) representing the examination region in a second time interval after the application of the amount of contrast agent, with the second time interval following the first time interval in time, - outputting and / or storing the synthetic representation (SR*) and / or transmitting the synthetic representation (SR*) to a separate computer system.

10. Method according to Claim 9, wherein the trained machine-learning model (MLMt) has been trained in a method according to any of Claims 1 to 8.

11. Computer system (10) comprising • a receiving unit (11), • a control and calculation unit (12) and • an output unit (13), - wherein the control and calculation unit (12) is configured to cause the receiving unit (11) to receive at least one input representation (R1,2*) of an examination region of a new examination object in a first state, the at least one input representation (R1,2*) representing the examination region in a first time interval before and / or after the application of an amount of a contrast agent, - wherein the control and calculation unit (12) is configured to input the received at least one input representation (R1,2*) into a trained machine-learning model (MLMt), ∘ wherein the trained machine-learning model (MLMt) has been trained on the basis of training data (TD) to produce a synthetic representation (SR) of the examination region in a second state on the basis of at least one input representation (R1, R2) of an examination region of an examination object in a first state, ∘ wherein the training data (TD) for each examination object of a multiplicity of examination objects comprise i) an input representation (R1) of the examination region, ii) a target representation (TR) and iii) a transformed target representation (TRT), ∘ wherein the transformed target representation (TRT) represents at least one part of the examination region of the examination object in frequency space, if the target representation (TR) represents the examination region of the examination object in real space, or in real space, if the target representation (TR) represents the examination region of the examination object in frequency space, ∘ wherein the machine-learning model (MLMt) has been trained for each examination object to minimize deviations between i) at least one part of the synthetic representation (SR) and at least one part of the target representation (TR) and ii) between at least one part of a transformed synthetic representation (SRT) and at least one part of the transformed target representation (TRT), - wherein the control and calculation unit (12) is configured to receive from the machine-learning model (MLMt) a synthetic representation (SR*) of the examination region of the new examination object in the second state, the synthetic representation (SR*) representing the examination region in a second time interval after the application of the amount of contrast agent, the second time interval following the first time interval in time, - wherein the control and calculation unit (12) is configured to cause the output unit (13) to output and / or store the synthetic representation (SR*) and / or transmit the latter to a separate computer system.

12. Computer program product comprising a computer program (40) that can be loaded into a working memory (22) of a computer system (1), where it causes the computer system (1) to execute the following steps: - receiving at least one input representation (R1,2*) of the examination region of the examination object in a first state, the at least one input representation (R1,2*) representing the examination region in a first time interval before and / or after the application of an amount of a contrast agent, - providing a trained machine-learning model (MLMt), • wherein the trained machine-learning model (MLMt) has been trained on the basis of training data (TD) to produce a synthetic representation (SR) of the examination region in a second state on the basis of at least one input representation (R1, R2) of an examination region of an examination object in a first state, • wherein the training data (TD) for each examination object of a multiplicity of examination objects comprise i) an input representation (R1) of the examination region, ii) a target representation (TR) and iii) a transformed target representation (TRT), • wherein the transformed target representation (TRT) represents at least one part of the examination region of the examination object in frequency space, if the target representation (TR) represents the examination region of the examination object in real space, or in real space, if the target representation (TR) represents the examination region of the examination object in frequency space, • wherein the machine-learning model (MLMt) has been trained for each examination object to minimize deviations between i) at least one part of the synthetic representation (SR) and at least one part of the target representation (TR) and ii) between at least one part of a transformed synthetic representation (SRT) and at least one part of the transformed target representation (TRT), - inputting the at least one received input representation (R1,2*) of the examination region of the examination object into the trained machine-learning model (MLMt), - receiving a synthetic representation (SR*) of the examination region of the examination object in the second state from the machine-learning model (MLMt), the synthetic representation (SR*) representing the examination region in a second time interval after the application of the amount of contrast agent, with the second time interval following the first time interval in time, - outputting and / or storing the synthetic representation (SR*) and / or transmitting the synthetic representation (SR*) to a separate computer system.