MACHINE LEARNING IN CONTRAST-ENHANCED RADIOLOGY

Description

[0001] The present invention relates to the technical field of generating artificial contrast-enhanced radiological images using machine learning methods.

[0002] WO2019 / 074938A1 discloses a method for reducing the amount of contrast agent when generating radiological images using an artificial neural network.

[0003] In a first step, a training dataset is generated. The training dataset includes, for a large number of individuals, for each individual i) a native radiological image ( zero-contrast image ), ii) a radiological image after the application of a small amount of contrast medium ( lowcontrast image ) and iii) a radiological image after the application of a standard amount of contrast medium ( full-contrast image ). The term "plurality" preferably means more than 10, more preferably more than 100.

[0004] In a second step, an artificial neural network is trained to predict an artificial radiological image for each person in the training data set based on the native image and the image after application of a small amount of contrast agent. This image shows an image area after application of the standard amount of contrast agent. The measured radiological image after application of a standard amount of contrast agent serves as a reference during training ( ground truth ).

[0005] In a third step, the trained artificial neural network can be used to predict an artificial radiological image for a new person based on a native image and a radiological image after the application of a small amount of contrast agent, which shows the imaged area as it would look if a standard amount of contrast agent had been applied.

[0006] The method disclosed in WO2019 / 074938A1 has disadvantages.

[0007] Co-registration of the radiological images is required to adjust the individual radiological images so that the pixels / voxels match each other, i.e., so that one pixel / voxel of one radiological image of a person shows the same examination area as the pixel / voxel of another radiological image of the person. If the radiological images do not match, artifacts will appear in the artificially generated radiological images, which can obscure and / or distort and / or simulate small anatomical structures in the image area.

[0008] Furthermore, in the method disclosed in WO2019 / 074938A1, complete radiological images are always used for prediction. In particular, when training the artificial neural network and predicting an artificial radiological image, not only the radiological images mentioned above under points i) and ii) are used, but also additional radiological images - e.g., radiological images after application of varying amounts of contrast agent - the computational effort for generating the artificial radiological image can quickly become very large. It is conceivable that a lot of time is required to calculate artificial radiological images and / or special and / or expensive hardware is required to be able to carry out the calculations (within a reasonable period of time). It is conceivable that radiological images are divided into partial areas ( patches ) and these sections must be processed separately in order not to overload the computer's memory with excessively large radiological images. However, such a procedure can lead to artifacts at the seams when the separately processed sections are reassembled into a complete radiological image ( Stitching -artifacts). The subsequent removal of such Stitching -Artifacts mean additional effort and the risk of errors in the synthetically generated radiological images, which could be misinterpreted by a radiologist (risk of misdiagnosis).

[0009] Based on the described state of the art, the technical task is therefore to create a solution for generating artificial radiological images that is tolerant to errors in co-registration and / or where the prediction requires less computing power and / or where the computing effort can be adapted to the given hardware and / or the available time and / or where the risk of artifacts (in particular Stitching artifacts) can be reduced.

[0010] This object is achieved by the subject matter of the independent claims. Preferred embodiments of the present invention can be found in the dependent claims, in the present description, and in the drawings.

[0011] A first object of the present invention is a computer-implemented method comprising the steps: Receiving a first representation of an examination region of an examination object in the frequency domain, wherein the first representation represents the examination region without contrast agent or after application of a first amount of contrast agent, Receiving a second representation of the examination region of the examination object in the frequency domain, wherein the second representation represents the examination region after application of a second amount of contrast agent, Supplying at least a portion of the first representation and at least a portion of the second representation to a machine learning model, Receiving, from the machine learning model, a third representation of the examination region in the frequency domain, wherein the third representation represents the examination region after application of a third amount of contrast agent,Generating a representation of the examination area in a spatial representation based on the third representation, outputting and / or storing the representation of the examination area in the spatial representation.

[0012] Another object of the present invention 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 two representations of an examination region of an examination object, a first representation and a second representation, wherein the first representation represents the examination region in the frequency domain without contrast agent or after application of a first amount of contrast agent, wherein the second representation represents the examination region in the frequency domain after application of a second amount of contrast agent, wherein the control and computing unit is configured to supply at least a portion of the first representation and at least a portion of the second representation to a machine learning model, wherein the control and computing unit is configured to receive a third representation of the examination region from the machine learning model,wherein the third representation represents the examination area in the frequency domain after application of a third amount of contrast agent, wherein the control and computing unit is configured to generate a representation of the examination area in the spatial domain on the basis of the third representation, wherein the control and computing unit is configured to cause the output unit to output and / or store the representation of the examination area in the spatial domain.

[0013] Another object of the present invention is a computer program that can be loaded into a working memory of a computer system and causes the computer system to carry out the following steps: Receiving a first representation of an examination region of an examination object in the frequency domain, wherein the first representation represents the examination region without contrast agent or after application of a first amount of contrast agent, Receiving a second representation of the examination region of the examination object in the frequency domain, wherein the second representation represents the examination region after application of a second amount of contrast agent, Supplying at least a portion of the first representation and at least a portion of the second representation to a machine learning model, Receiving, from the machine learning model, a third representation of the examination region in the frequency domain, wherein the third representation represents the examination region after application of a third amount of contrast agent,Generating a representation of the examination area in a spatial representation based on the third representation, outputting and / or storing the representation of the examination area in the spatial representation.

[0014] The invention is explained in more detail below, without distinguishing between the subject matter of the invention (method, computer system, computer program). Rather, the following explanations are intended to apply analogously to all subject matter of the invention, regardless of the context in which they occur (method, computer system, computer program).

[0015] 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.

[0016] With the aid of the present invention, an artificial radiological image of an examination area of ​​an examination object can be generated.

[0017] The "object of investigation" is usually a living being, preferably a mammal, most preferably a human.

[0018] The "area of ​​investigation" is a part of the object under investigation, for example an organ or a part of an organ.

[0019] The examination area, also known as the recording volume (English: field of view, A field of view (FOV) is 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.

[0020] The examination area is subjected to a radiological examination.

[0021] "Radiology" is the branch of medicine that deals with the application of 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.

[0022] In a preferred embodiment of the present invention, the radiological examination is a magnetic resonance imaging examination.

[0023] Magnetic resonance imaging, abbreviated MRI or MR (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.

[0024] 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.

[0025] 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 frequency domain and can be transformed into spatial domain (image space) by subsequent inverse Fourier transformation.

[0026] In native MRI, tissue contrasts are generated by the different relaxation times (T1 and T2) and the proton density. T1 relaxation describes the transition of the longitudinal magnetization to its equilibrium state, where T1 is the time required to reach 63.21% of the equilibrium magnetization before resonance excitation. It is also called the longitudinal relaxation time or spin-lattice relaxation time. Similarly, T2 relaxation describes the transition of the transverse magnetization to its equilibrium state.

[0027] Contrast agents are often used in radiological examinations to enhance contrast. "Contrast agents" are substances or mixtures of substances that enhance the visualization of body structures and functions in imaging procedures such as X-rays, magnetic resonance imaging, and sonography.

[0028] In computed tomography, iodine-containing solutions are usually used as contrast agents. In magnetic resonance imaging (MRI), superparamagnetic substances (e.g., iron oxide nanoparticles, superparamagnetic iron-platinum particles (SIPPs)) or paramagnetic substances (e.g., gadolinium chelates, manganese chelates) are usually used as contrast agents. In the case of sonography, liquids containing gas-filled microbubbles ( microbubbles ) enthalten, intravenös verabreicht. Beispiele für Kontrastmittel sind in der Literatur zu finden (siehe z.B. A. S. L. Jascinth et al.: Contrast Agents in computed tomography: A Review, Journal of Applied Dental and Medical Sciences, 2016, Vol. 2, Issue 2, 143 - 149; H. Lusic et al.: X-ray-Computed Tomography Contrast Agents, Chem. Rev. 2013, 113, 3, 1641-1666; https: / / www.radiology.wisc.edu / wpcontent / uploads / 2017 / 10 / contrast-agents-tutorial.pdf, M. R. Nough et al.: Radiographie and magnetic resonances contrast agents: Essentials and tips for safe practices, World J Radiol. 2017 Sep 28; 9(9): 339-349; L. C. Abonyi et al.: Intravascular Contrast Media in Radiography: Historical Development & Review of Risk Factors for Adverse Reactions, South American Journal of Clinical Research, 2016, Vol. 3, Issue 1, 1-10; ACR Manual on Contrast Media, 2020, ISBN: 978-1-55903-012-0; A. Ignee et al.: Ultrasound contrast agents, Endosc Ultrasound. 2016 Nov-Dec; 5(6): 355-362).

[0029] MR contrast agents exert their effect by altering the relaxation times of the structures that absorb the contrast agent. Two groups of substances can be distinguished: paramagnetic and superparamagnetic substances. Both groups of substances contain unpaired electrons that induce a magnetic field around the individual atoms or molecules. Superparamagnetic contrast agents lead to a predominant T2 shortening, while paramagnetic contrast agents essentially lead to a T1 shortening. The effect of these contrast agents is indirect, as the contrast agent itself does not emit a signal but only influences the signal intensity in its surroundings. An example of a superparamagnetic contrast agent is iron oxide nanoparticles (SPIO). superparamagnetic iron oxide Examples of paramagnetic contrast agents are gadolinium chelates such as gadopentetate dimeglumine (trade name: Magnevist ®< among others), gadoteric acid (Dotarem ®< , Dotagita ®< , Cyclolux ®< ), gadodiamide (Omniscan ®< ), gadoteridol (ProHance ®< ) and gadobutrol (Gadovist ®< ).

[0030] With the aid of the present invention, an artificial radiological image can be generated which shows the examination area of ​​an examination subject as it would appear if a specific amount of contrast agent had been administered to the examination subject / examination area, without this specific amount actually having to be administered.

[0031] This involves using a machine learning model (also referred to as a predictive model in this description) that learns from training data how different amounts of contrast agent affect the contrast of a radiological image of an examination area. The trained model can then be used to predict a radiological image with a contrast enhancement that would result after applying a specific amount of contrast agent, without actually applying that amount.

[0032] The present invention can therefore be used, for example, similarly to that described in WO2019 / 074938A1, to reduce the amount of contrast agent applied without having to forego the advantages of a high contrast agent administration (i.e., high contrast enhancement).

[0033] However, the prediction of radiological images is not based on images ( images ), as described in WO2019 / 074938A1. The images used in WO2019 / 074938A1 to predict artificial radiological images ( images ) are representations (depictions) of an area of ​​investigation in the spatial space (also "image space" called).

[0034] According to the invention, the machine learning model is trained with the aid of representations of an examination area of ​​a large number of examination objects in the frequency space, and the prediction is also carried out on the basis of representations of the examination area in the frequency space (also referred to as spatial frequency space or Fourier space or frequency domain or Fourier representation).

[0035] In magnetic resonance imaging, the raw data is typically generated as so-called k-space data due to the measurement method described above. This k-space data is a representation of an examination region in a frequency domain. Such k-space data can be used to train a machine learning model, to validate the model, and to make predictions using the trained model according to the present invention.

[0036] If, on the other hand, representations exist in the spatial space, they can be converted (transformed) into a representation in the frequency space, for example using Fourier transformation.

[0037] If a radiological image of an examination area is available in the form of a two- or three-dimensional image in spatial space, this representation of the examination area can be converted into a two- or three-dimensional representation of the examination area in the frequency space, for example by means of a 2D or 3D Fourier transformation.

[0038] Conversely, representations in the frequency domain can be converted (transformed) into a representation in the position domain by inverse Fourier transformation.

[0039] Fig. 1 shows schematically and exemplarily the relationship between representations of an area of ​​investigation in spatial space and in frequency space. In Fig. 1 A timeline is shown. At three different points in time t 1 , t 2 and t3, measurement-based representations of an examination area are generated. The examination area is the lung of a human. At the time t 1, a first representation is generated. This can be a representation O1 of the examination area (lung) in the spatial domain or a representation F 1 of the examination area (lung) in the frequency domain. The representation O1 of the examination area in the spatial domain can be converted using a Fourier transformation ( FT ) into a representation F1 of the examination area in the frequency domain. The representation F1 of the examination area in the frequency domain can be converted using inverse Fourier transformation ( iFT ) into a representation O1 of the study area in spatial space.

[0040] The representations O1 and F1 can, for example, be representations of the examination area without a contrast agent or after application of an initial amount of contrast agent.

[0041] The two representations O1 and F1 contain the same information about the area of ​​investigation, only in a different representation.

[0042] At the time t 2, a further representation is generated. This can be a representation O2 of the examination area (lung) in the spatial domain or a representation F2 of the examination area (lung) in the frequency domain. The representation O2 of the examination area in the spatial domain can be converted using Fourier transformation ( FT ) into a representation F2 of the examination area in the frequency domain. The representation F2 of the examination area in the frequency domain can be converted using inverse Fourier transformation ( iFT ) into a representation O2 of the study area in spatial space.

[0043] The representations O2 and F2 can, for example, be representations of the examination area after application of a second amount of contrast agent.

[0044] The two representations O2 and F2 contain the same information about the area under investigation, only in a different representation.

[0045] At the time t 3, a further representation is generated. This can be a representation O3 of the examination area (lung) in the spatial domain or a representation F3 of the examination area (lung) in the frequency domain. The representation O3 of the examination area in the spatial domain can be converted using Fourier transformation ( FT ) into a representation F3 of the examination area in the frequency domain. The representation F3 of the examination area in the frequency domain can be converted using inverse Fourier transformation ( iFT ) into a representation O3 of the study area in spatial space.

[0046] The representations O3 and F3 can, for example, be representations of the examination area after application of a third amount of contrast agent.

[0047] The two representations O3 and F3 contain the same information about the study area, only in a different representation.

[0048] For humans, the representations O1, O2, and O3 of the examination area in spatial space are the most common representations; humans can directly perceive such spatial representations. The representations O1, O2, and O3 demonstrate the influence that different amounts of contrast agent can have on the appearance of the examination area in an MR examination. In the present case, the amount of contrast agent increases from O1 via O2 to O3. The same information is also contained in the representations F1, F2, and F3; however, it is more difficult for humans to perceive it from the representations F1, F2, and F3.

[0049] It is also conceivable to use a transformation other than the Fourier transform to convert position-space representations into frequency-space representations. The three main properties that such a transformation must satisfy are: a) Existence of a unique inverse transformation (unambiguous relationship between spatial and frequency domain representation) b) Locality of the contrast information c) Robustness against insufficient image registration

[0050] Details on the transformation from one representation to another are described in a variety of specialist books and publications (see, for example: W. Burger, MJ Burge: Digital Image Processing: An Algorithmic Introduction Using Java, Texts in Computer Science, 2nd edition, Springer-Verlag, 2016, ISBN: 9781447166849; W. Birkfellner: Applied Medical Image Processing, Second Edition: A Basic Course, Taylor & Francis, 2014, ISBN: 9781466555570; R. Bracewell: Fourier Analysis and Imaging, Springer Science & Business Media, 2004, ISBN: 9780306481871).

[0051] In order for the machine learning model (prediction model) according to the invention to be able to make the predictions described here, it must be configured (trained) accordingly in advance.

[0052] The term "prediction" means that at least one representation of an examination region, which represents the examination region after an application of a specific amount of contrast agent in the frequency domain, is calculated on the basis of at least two representations of the examination region in the frequency domain, wherein the at least two representations represent the examination region after application of different amounts of contrast agent and / or after application of different contrast agents.

[0053] In other words, at least one first representation of an examination area of ​​an examination object and at least one second representation of the examination area of ​​the examination object are used to predict at least one third representation of the examination area of ​​the examination object. All representations, the at least one first, the at least one second, and the at least one third representation, are representations of the examination area in the frequency domain.

[0054] The at least one first representation represents the examination area after application of a first amount of a first contrast agent, whereby this first amount can also be zero (no application of a contrast agent).

[0055] The at least one second representation represents the examination region after application of a second amount of the first contrast agent or a second amount of a second contrast agent. If the first contrast agent is used to generate the second representation, the second amount is typically different from the first amount. If a second contrast agent is used, the second amount may be equal to or different from the first amount.

[0056] The at least one third representation represents the examination region after application of a third amount of the first contrast agent or a third amount of the second contrast agent or a third amount of a third contrast agent. If the same contrast agent is used to generate the first representation, the second representation and the third representation, the third amount is usually not equal to the second amount and not equal to the first amount; for the first amount M1, the second amount M2 and the third amount M3, the following preferably applies: the first amount M1 is greater than or equal to zero, the second amount M2 is greater than the first amount M1 and the third amount M3 is greater than the second amount M2 (0 ≤ M1 < M2 < M3).

[0057] If a different contrast agent is used to generate the third representation than to generate the first and / or second representation, the amount of the other contrast agent may be equal to or different from the amount(s) of contrast agent(s) used to generate the first and / or second representation.

[0058] The prediction model can therefore be trained to learn the influence of different amounts of contrast agent on representations of the examination area in the frequency domain, but it can also be trained to learn the influence of different contrast agents on representations of the examination area in the frequency domain.

[0059] The prediction model is preferably created (configured, trained) using a self-learning algorithm in supervised machine learning. Training data is used for learning. This training data comprises, for each of a large number of examination objects, a plurality of representations of an examination area. The examination area is usually the same for all examination objects (e.g., a part of a human body or an organ or a part of an organ). The representations of the training data set are also referred to as reference representations in this description.

[0060] For each examination object, the training data includes i) at least one first reference representation of the examination region in the frequency domain, representing the examination region without contrast agent or after application of a first amount of contrast agent, ii) at least one second reference representation of the examination region in the frequency domain, representing the examination region after application of a second amount of contrast agent, and iii) at least one third reference representation of the examination region in the frequency domain, representing the examination region after application of a third amount of contrast agent. As described above, the following applies: the first, second, and third amounts of contrast agent and / or the respective contrast agents used differ from one another.

[0061] The prediction model is trained to predict (calculate) for each object of investigation the at least one third reference representation based on the at least one first reference representation and the at least one second reference representation.

[0062] In machine learning, the self-learning algorithm creates a statistical model based on the training data. This means that the algorithm doesn't simply memorize the examples; instead, it "recognizes" patterns and regularities in the training data. This allows the prediction model to evaluate even unknown data. Validation data can be used to test the accuracy of the evaluation of unknown data.

[0063] The prediction model can be trained using supervised learning (SLL). supervised learning ), i.e., the algorithm is presented with triples of data sets (first, second, and third representations) one after the other. The algorithm then learns a relationship between these data sets.

[0064] Self-learning systems that are trained using supervised learning are widely described in the state of the art (see, for example, C. Perez: Machine Learning Techniques: Supervised Learning and Classification, Amazon Digital Services LLC - Kdp Print Us, 2019, ISBN 1096996545, 9781096996545).

[0065] Preferably, the prediction model is or includes an artificial neural network.

[0066] 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.

[0067] The input neurons are used to receive first and second representations. The output neurons are used to output at least one third representation for at least one first representation and at least one second representation.

[0068] The processing elements of the layers between the input neurons and the output neurons are connected in a predetermined pattern with predetermined connection weights.

[0069] The artificial neural network is preferably a so-called convolutional neural network (CNN for short).

[0070] A convolutional neural network is capable of processing input data in the form of a matrix. A CNN essentially consists of filters (convolutional layer) and aggregation layers (pooling layer) that alternately repeat, and finally, one or more layers of "normal" fully connected neurons (dense / fully connected layer).

[0071] The neural network can be trained, for example, using a backpropagation method. The goal is to achieve the most reliable mapping possible from given input data to given output data. The quality of the mapping is determined by an error function (English: loss function ). The goal is to minimize the error function. The backpropagation method trains an artificial neural network by changing the connection weights.

[0072] In the trained state, the connection weights between the processing elements contain information regarding the relationship between the at least one first representation and the at least one second representation on the one hand, and the at least one third representation on the other hand. This information can be used to predict at least one third representation based on at least one first and at least one second representation.

[0073] 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 data.

[0074] Further details on setting up and training artificial neural networks can be found in the state of the art (see, for example: S. Khan et al.: A Guide to Convolutional Neural Networks for Computer Vision, Morgan & Claypool Publishers 2018, ISBN 1681730227, 9781681730226, WO2018 / 183044A1, WO2018 / 200493, WO2019 / 074938A1, WO2019 / 204406A1, WO2019 / 241659A1).

[0075] Preferably, the prediction model is a Generative Adversarial Network (GAN) (see e.g.: http: / / 3dgan.csail.mit.edu / ).

[0076] In addition to the representations, further information about the object under investigation, the examination area, examination conditions and / or the radiological examination methods can also be used for training, validation and prediction.

[0077] Examples of information about the subject being examined include: gender, age, weight, height, medical history, type, duration, and amount of medications already taken, blood pressure, central venous pressure, respiratory rate, serum albumin, total bilirubin, blood sugar, iron levels, respiratory capacity, and the like. This information about the subject being examined can also be retrieved from a database or an electronic patient record, for example.

[0078] Examples of information on the area of ​​investigation include: previous illnesses, operations, partial resection, liver transplantation, iron liver, fatty liver and the like.

[0079] Fig. 2 shows schematically and exemplarily how the Fig. 1 The generated representations (F1), (F2), and (F3) of the examination area in frequency space can be used to train a prediction model (PM). The representations (F1), (F2), and (F3) form a set of training data for an examination object. Training is carried out using a large number of training data sets from a large number of examination objects.

[0080] The representations (F1) and (F2) are a first and a second representation of an examination area in the frequency domain, whereby the representations represent the examination area after the application of different amounts of contrast agent (it is also possible that in one case no contrast agent has been applied).

[0081] The representation (F3) is a third representation of the examination area in the frequency domain, which represents the examination area after the application of a third amount of contrast agent.

[0082] The prediction model is Fig. 2 trained to predict the representation (F3) of the examination area in the frequency space from the representations (F1) and (F2) of the examination area in the frequency space.

[0083] The representations (F1) and (F2) are input into the prediction model (PM), and the prediction model calculates a representation (F3*) from the representations (F1) and (F2). The asterisk (*) indicates that the representation (F3*) is a predicted representation. The calculated representation (F3*) is compared with the representation (F3). The deviations between the calculated representation (F3*) and the measured representation (F3) can be used in a backpropagation process to train the prediction model and reduce the deviations to a defined minimum. Once the prediction model has been trained using a large number of training data sets from a large number of test objects and the prediction has achieved a defined accuracy, the trained prediction model can be used for prediction.

[0084] Fig. 3 shows, by way of example and schematically in the form of a flow chart, a preferred embodiment of the method according to the invention for training a machine learning model.

[0085] The method (100) comprises the following steps: (110) Receiving a training data set, wherein the training data set comprises, for each examination object of a plurality of examination objects, i) a first reference representation of an examination region of the examination object in the frequency domain, ii) a second reference representation of the examination region of the examination object in the frequency domain, and iii) a third reference representation of the examination region of the examination object in the frequency domain, wherein the first reference representation represents the examination region after application of a first amount of contrast agent, wherein the first amount can also be zero, wherein the second reference representation represents the examination region after application of a second amount of contrast agent, wherein the third reference representation represents the examination region after application of a third amount of contrast agent,(120) for each examination object: supplying at least a portion of the first reference representation and at least a portion of the second reference representation to the machine learning model, wherein the machine learning model is trained to generate a representation of the examination region in the frequency domain after application of the third amount of contrast agent based on the first reference representation and the second reference representation, wherein the training comprises minimizing an error function, wherein the error function quantifies deviations of the generated representation of the examination region from the third reference representation, (130) outputting and / or storing the trained model and / or supplying the trained model to a method for predicting a representation of the examination region of a new examination object.

[0086] The use of the trained prediction model for prediction is exemplary and schematically shown in Fig. 4 shown. Fig. 4 shows that in Fig. 2 trained prediction model (PM). The prediction model is used to predict at least a third representation of the examination area of ​​the examination object in the frequency domain based on at least a first representation of the examination area of ​​the examination object in the frequency domain and at least a second representation of the examination area of ​​the examination object in the frequency domain, wherein the representations represent the examination area after the application of different amounts of contrast agent.

[0087] In the present example, a first representation (F̃1) and a second representation (F̃2) of the examination area in the frequency domain are input into the prediction model, and the prediction model generates (calculates) a third representation (F̃3*). The tilde (~) signals that the representations are representations of a new examination object, for which no representations are usually available that have been used in the training procedure for training the prediction model. The asterisk (*) signals that the representation (F̃3*) is a predicted representation. The representation (F̃3*) of the examination area in the frequency domain can, for example, be calculated using inverse Fourier transformation. iFT be transformed into a representation (Õ3*) of the study area in spatial space.

[0088] Fig. 5 shows, by way of example and schematically in the form of a flow chart, a preferred embodiment of the method according to the invention for predicting a representation of an examination region of an examination object.

[0089] The method (200) comprises the following steps: (210) Receiving a first representation of an examination region of an examination object in the frequency domain, wherein the first representation represents the examination region without contrast agent or after application of a first amount of contrast agent, (220) Receiving a second representation of the examination region of the examination object in the frequency domain, wherein the second representation represents the examination region after application of a second amount of contrast agent, (230) Feeding the first representation and the second representation to a machine learning model, (240) Receiving, from the machine learning model, a third representation of the examination region in the frequency domain, wherein the third representation represents the examination region after application of a third amount of contrast agent,(250) Generating a representation of the examination area in a spatial representation based on the third representation, (260) Outputting and / or storing the representation of the examination area in the spatial representation. ,

[0090] As already described, the representations of the study area used for training, validation and prediction are representations of the study area in frequency space (also called spatial frequency space or Fourier space or frequency domain or Fourier representation).

[0091] In magnetic resonance imaging, the raw data is usually generated as so-called k-space data due to the measurement method described above. This k-space data is a representation of an examination region in a frequency domain; i.e., such k-space data can be used for training, validation, and prediction. If representations are present in the spatial domain, such representations in the spatial domain can be converted (transformed), for example, by Fourier transformation, into a representation in the frequency domain; conversely, representations in the frequency domain can be converted (transformed), for example, by inverse Fourier transformation, into a representation in the spatial domain.

[0092] If a radiological image of an examination area is available in the form of a two-dimensional image in spatial space, this representation of the examination area can be converted into a two-dimensional representation of the examination area in frequency space by means of a 2D Fourier transformation.

[0093] A three-dimensional image (volume representation) of an examination area can be treated as a stack of two-dimensional images. It is also conceivable that the three-dimensional image is converted into a three-dimensional representation of the examination area in the frequency domain using a 3D Fourier transform.

[0094] The inventive use of representations of the examination area in the frequency domain has various advantages over the use of representations of the examination area in the spatial domain.

[0095] For example, co-registration of individual representations in frequency domain is less critical than in spatial domain. "Co-registration" (also called "image registration" in the prior art) is an important process in digital image processing and serves to align two or more images of the same scene, or at least similar scenes, as closely as possible. One of the images is designated as the reference image, and the others are called object images. To optimally adapt these to the reference image, a compensating transformation is calculated. The images to be registered differ from one another because they were acquired from different positions, at different times, and / or with different sensors.

[0096] In the case of the present invention, the individual representations were generated at different times; on the other hand, they differ with regard to the contrast agent content in the examination area and / or with regard to the contrast agent used.

[0097] The advantage of using representations of the examination area in frequency space over representations of the examination area in spatial space is that the training, validation, and prediction procedures are more tolerant of image registration errors. In other words, if representations in frequency space are not precisely superimposed, this has less impact than if representations in spatial space are not precisely superimposed. This follows from the properties of the Fourier transform: as already described, the contrast information of Fourier-transformed images is always mapped near the origin of Fourier space. Rotations in image space (spatial space) result in image information (e.g., a visible structure) being localized in a different area of ​​the image after the transformation.However, in Fourier space, these transformations do not change the region in which the contrast information relevant to the present invention is encoded.

[0098] Another advantage of using frequency-space representation is that it allows contrast information, which is important for training and prediction, to be separated from detailed information (fine structure). This makes it possible to focus on the information the predictive model should learn during training, and also focus on the information the predictive model should predict during prediction: contrast information.

[0099] While contrast information in a spatial-space representation of an examination area is typically distributed throughout the entire representation (each pixel / voxel carries contrast information), the contrast information in a frequency-space representation of an examination area is encoded in and around the center of the frequency space. In other words, the low frequencies in a frequency-space representation are responsible for contrast, while the high frequencies contain information about fine structures.

[0100] This makes it possible to separate the contrast information, limit training and prediction to the contrast information, and feed back information about the fine structures after training / prediction.

[0101] For this purpose, a region can be specified in the representations of the study area used for training, validation and prediction.

[0102] The region can be specified, for example, by a user of the computer system according to the invention entering one or more parameters into the computer system according to the invention and / or selecting them from a list that define the shape and / or size of the region. However, it is also conceivable that the specification occurs automatically, for example, by the computer system according to the invention, which has been configured accordingly to select a predefined region in the representations of the examination region.

[0103] The specified range is usually smaller than the frequency space filled by the respective representation, but in any case includes the center of the frequency space.

[0104] A region of the representation encompassing the center of the frequency space (also called the origin or zero point) contains the contrast information relevant for the method according to the invention. If the specified region is smaller than the frequency space filled by the respective representation, the computational effort required for training, validation, and subsequent prediction is reduced. Thus, the size of the region can directly influence the computational effort.

[0105] In principle, it is also possible to specify a range that corresponds to the entire frequency space filled by the respective representation; in such a case, no reduction to a subrange of the frequency space takes place and the computational effort is maximized.

[0106] By specifying a region around the center of the frequency space, the user of the computer system according to the invention can decide for themselves whether they want training, validation, and prediction based on the complete representations of the examination region in the frequency space (which means maximum computational effort) or whether they want to reduce the computational effort by specifying a region that is smaller than the frequency space filled by the representations. The size of the specified region can directly influence the required computational effort.

[0107] The specified region typically has the same dimension as the frequency space: in the case of a 2D representation in a 2D frequency space, the specified region is typically a surface; in the case of a 3D representation in a 3D frequency space, the specified region is typically a volume.

[0108] The specified area can, in principle, have any shape; for example, it can be round and / or square, concave and / or convex. Preferably, the area is cuboid or cube-shaped in the case of a 3D frequency space in a Cartesian coordinate system and rectangular or square in the case of a 2D frequency space in a Cartesian coordinate system. However, it is also conceivable that the specified area is spherical or circular, or has another shape.

[0109] Preferably, the geometric center of gravity of the specified area coincides with the center of the frequency space.

[0110] The representations used for training, validation, and prediction are reduced to the specified range. The term "reducing" means that all parts of a representation that do not lie within the specified range are cut away (discarded) or covered by a mask. Masking involves covering the areas that lie outside the specified range with a mask, leaving only the specified area uncovered. For example, covering with a mask can set the color values ​​of the corresponding pixels / voxels to zero (black).

[0111] The representations obtained in this way are also referred to as reduced representations in this description.

[0112] The reduced representations can be used for training, validation and prediction.

[0113] If a third representation of an area of ​​investigation is to be predicted for an object under investigation based on a first and a second representation, the reduced first representation and the reduced second representation can be fed into the prediction model, and the prediction model then generates a reduced third representation. In a further step, the detailed information that was cut out or covered during the reduction can be added back to the predicted representation. For example, the parts of the first representation that lie outside the specified area can be added to the reduced third representation. It is also conceivable that the parts of the second representation that lie outside the specified area are added to the reduced third representation.It is also conceivable that both parts of the first representation that lie outside the specified range and parts of the second representation that lie outside the specified range are added to the reduced third representation. In other words, the third representation is supplemented by parts of the first and / or second representation that were cut away / covered with a mask during the reduction process. The result is a supplemented third representation. A position-space representation can then be obtained from the supplemented third representation using a transformation (e.g., an inverse Fourier transform).

[0114] The resulting signal strengths for different location coordinates can be converted into gray values ​​or color values ​​in a further step in order to have a digital image in a common image format (e.g. DICOM).

[0115] The representation of the study area in the spatial representation can be displayed on a screen, printed out and / or stored in a data storage device.

[0116] The procedure described is explained below using the example of Fig. 6 and Fig. 7 explained in more detail.

[0117] Fig. 6 shows, by way of example and schematically, a step in training a prediction model according to a preferred embodiment of the present invention.

[0118] As in the case of Fig. 2 , a first representation (F1), a second representation (F2) and a third representation (F3) of an examination area of ​​an examination object in the frequency domain are received. In the representations (F1), (F2) and (F3) each area A of the same size and shape. The area Aencompasses the center of the frequency space and, in this case, has a square shape, with the geometric center of gravity of the square coinciding with the center of the frequency space. The representations (F1), (F2), and (F3) are mapped to the respective specified range. A reduced: this results in three reduced representations (F1 red ), (F2 red ) and (F3 red ). The reduced representations are used for training. The prediction model is trained to predict the reduced representation (F3 red ) from the reduced representations (F1 red ) and (F2 red ). The reduced representations (F1 red ) and (F2 red ) are fed to the prediction model (PM), and the prediction model calculates a reduced representation (F3* red ) that should be as close as possible to the reduced representation (F3 red ).

[0119] Fig. 7 shows an example and schematically how the Fig. 6 trained prediction model can be used for prediction.

[0120] In the present example, a first representation (F̃1) and a second representation (F̃2) of the examination area are received in the frequency domain and each mapped to a specified area Areduced. This results in two reduced representations (F̃1 red ) and (F̃2 red ). The reduced first representation (F̃1 red ) and the reduced second representation (F̃2 red ) are fed to the trained prediction model (PM). The trained prediction model (PM) calculates a reduced third representation (F̃3 red *) from the reduced representations (F̃1 red ) and (F̃2 red ). In a further step, the reduced third representation (F̃3 red *) is supplemented by that area of ​​the received first representation (F̃1) that was discarded when the received first representation (F̃1) was reduced. In other words, those parts of the received first representation (F̃1) that lie outside (not within) the specified area are supplemented to form the reduced third representation (F̃3 red *).As described, instead of or in addition to parts of the received first representation (F̃1), parts of the received second representation (F̃2) can also be added to the reduced third representation (F̃3 red *).

[0121] From the supplemented representation (F̃3 red *) + (F̃1 DI ) a representation of the investigation area in the position space (Õ3*) can be generated by inverse Fourier transformation.

[0122] It should be noted that other methods can also be used to transform a frequency-space representation into a position-space representation, such as iterative reconstruction methods.

[0123] Fig. 8 shows schematically in the form of a flowchart a preferred embodiment of the inventive method for training a machine learning model.

[0124] The method (300) comprises the steps: (310) Receiving a training data set, wherein the training data set comprises, for each examination object of a plurality of examination objects, i) a first reference representation of an examination region of the examination object in the frequency domain, ii) a second reference representation of the examination region of the examination object in the frequency domain, and iii) a third reference representation of the examination region of the examination object in the frequency domain, wherein the first reference representation represents the examination region after application of a first amount of contrast agent, wherein the first amount can also be zero, wherein the second reference representation represents the examination region after application of a second amount of contrast agent, wherein the third reference representation represents the examination region after application of a third amount of contrast agent,(320) Specifying a region in the reference representations, wherein the specified region comprises the center of the frequency space, (330) Reducing the representations to the specified region, wherein a reduced first reference representation, a reduced second reference representation, and a reduced third reference representation are obtained for each examination object, (340) For each examination object: Feeding the reduced first reference representation and the reduced second reference representation to the machine learning model, wherein the machine learning model is trained to generate a reduced representation of the examination region in the frequency space after application of the third amount of contrast agent based on the reduced first reference representation and the reduced second reference representation, wherein the training comprises minimizing an error function,wherein the error function quantifies deviations of the generated reduced representation of the examination area from the reduced third reference representation, (350) outputting and / or storing the trained model and / or feeding the trained model to a method for predicting a representation of the examination area of ​​a new examination object. ,

[0125] Fig. 9 shows schematically in the form of a flow chart a preferred embodiment of the method according to the invention for predicting a representation of an examination region of an examination object.

[0126] The method (400) comprises the steps: (410) Receiving a first representation of an examination region of an examination object in the frequency domain, wherein the first representation represents the examination region without contrast agent or after application of a first amount of contrast agent, (420) Receiving a second representation of the examination region of the examination object in the frequency domain, wherein the second representation represents the examination region after application of a second amount of contrast agent, (430) Specifying a region in the first representation and in the second representation, wherein the specified region comprises the center of the frequency domain, (440) Reducing the first representation and the second representation to the specified region, wherein a reduced first representation and a reduced second representation are obtained,(450) Feeding the reduced first representation and the reduced second representation to a machine learning model, (460) Receiving, from the machine learning model, a third representation of the examination region in the frequency domain, wherein the third representation represents the examination region after application of a third amount of contrast agent, (470) Supplementing the third representation with those parts of the received first and / or the received second representation that do not lie in the specified area, (480) Transforming the supplemented third representation obtained after supplementing the third representation into a representation of the examination region in a spatial representation, (490) Outputting and / or storing the representation of the examination region in the spatial representation.

[0127] Fig. 10 shows schematically in the form of a flow chart a further preferred embodiment of the method according to the invention for predicting a representation of an examination region of an examination object.

[0128] The method (500) comprises the steps: (510) Providing a trained machine learning model, wherein the model has been trained according to the method (100) or (300) described above to learn the influence of the amount of contrast agent on representations of the examination area in the frequency space, (520) Receiving a first representation of the examination area of ​​the examination object in the frequency space, wherein the first representation represents the examination area after application of a first amount of contrast agent, wherein the first amount can also be zero, and receiving a second representation of the examination area of ​​the examination object in the frequency space, wherein the second representation represents the examination area after application of a second amount of contrast agent, (530) Optionally: Specifying a region in the first representation and in the second representation, wherein the specified region comprises the center of the frequency space,(540) optionally: reducing the first representation and the second representation to the specified area, obtaining a reduced first representation and a reduced second representation, (550) supplying the received first representation and the received second representation or, if present, the reduced first representation and the reduced second representation to the machine learning model, (560) receiving, from the machine learning model, a third representation of the examination area in the frequency domain, wherein the third representation represents the examination area after application of a third amount of the contrast agent, (570) optionally: supplementing the third representation with those parts of the received first and / or the received second representation that do not lie in the specified area, obtaining a supplemented third representation,(580) Generating a representation of the examination area in a spatial representation from the third representation, or, if present, the supplemented third representation, (590) Outputting the representation of the examination area in the spatial representation. ,

[0129] Figur 11 shows schematically and by way of example an embodiment of the computer system according to the invention. The computer system (10) comprises a receiving unit (11), a control and computing unit (12), and an output unit (13).

[0130] A "computer system" is an electronic data processing system that processes data using programmable computing instructions. Such a system typically includes a control and processing unit, often referred to as a "computer," which includes a processor for performing logical operations and a RAM for loading a computer program, as well as peripherals.

[0131] 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 (screens), printers, scanners, mice, keyboards, joysticks, drives, cameras, microphones, speakers, etc. Internal connectors and expansion cards are also considered peripherals in computer technology.

[0132] Today's computer systems are often divided into desktop PCs, portable PCs, laptops, notebooks, netbooks and tablet PCs and so-called handhelds (e.g., smartphones); all of these systems can be used to implement the invention.

[0133] Inputs to the computer system (for example, for control by a user) are made via input devices such as a keyboard, a mouse, a microphone, a touch-sensitive display, and / or the like. Outputs are made via the output unit (13), which can in particular be a monitor (screen), a printer, and / or a data storage device.

[0134] The computer system (10) according to the invention is configured to predict, from at least two representations of an examination region in the frequency domain, which represent the examination region after application of different amounts of contrast agent, a representation of the examination region that shows the examination region after application of a specific amount of contrast agent, without this specific amount actually having to be administered.

[0135] The control and computing unit (12) serves to control the receiving unit (11) and the output unit (13), to coordinate the data and signal flows between the various units, to process representations of the examination area, and to generate artificial radiological images. It is conceivable that several control and computing units are present.

[0136] The receiving unit (11) serves to receive representations of an examination area. The representations can, for example, be transmitted from a magnetic resonance tomograph or a computed tomograph or read out from a data memory. The magnetic resonance tomograph or the computed tomograph can be a component of the computer system according to the invention. However, it is also conceivable that the computer system according to the invention is a component of a magnetic resonance tomograph or a computed tomograph. The transmission of representations can take place via a network connection or a direct connection. The transmission of representations can take place via a radio connection (WLAN, Bluetooth, mobile radio and / or the like) and / or via a cable. It is conceivable that several receiving units are present. The data memory can also be a component of the computer system according to the invention or can be linked to it, for example.be connected via a network. It is conceivable that multiple data storage devices exist.

[0137] The receiving unit receives the representations and, if applicable, other data (such as information on the object under examination, recording parameters and / or the like) and transmits them to the control and computing unit.

[0138] The control and computing unit is configured to generate artificial radiological images based on the received data.

[0139] The output unit (13) can be used to display the artificial radiological images (e.g., on a monitor), output them (e.g., via a printer), and / or store them in a data storage device. Multiple output units are conceivable.

[0140] As already described, the invention can be used to reduce the amount of contrast agent in a radiological examination. The prediction model can be trained to predict a third representation of the examination region of the examination region of the examination object based on a first representation of an examination region of an examination object, which represents the examination region without contrast agent or after application of a first amount of contrast agent, and based on a second representation of the examination region of the examination object, which represents the examination region after application of a second amount of contrast agent, wherein the third representation represents the examination region after application of a third amount of contrast agent, wherein the third amount is greater than the first amount and the second amount.Therefore, only a first amount (which can also be zero) and a second amount of contrast agent need to be applied to produce radiological images that look as if a larger third amount of contrast agent had been applied.

[0141] The invention can be used to generate artificial radiological images that show an examination area of ​​an examination subject after application of a contrast agent, even though a different contrast agent has been applied.

[0142] The invention can be used to generate an artificial radiological image based on a radiological image of a radiological examination (e.g., a magnetic resonance imaging scan) that shows the result of another radiological examination (e.g., the result of a computed tomography scan).

[0143] Further applications are conceivable.

Claims

1. Computer-implemented method comprising the steps of: - receiving a first representation (F̃1) of an examination region of an examination object, the first representation (F̃1) representing the examination region without contrast agent or after administration of a first amount of contrast agent, - receiving a second representation (F̃2) of the examination region of the examination object, the second representation (F̃2) representing the examination region after an administration of a second amount of contrast agent, - feeding at least part of the first representation (F̃1) and at least part of the second representation (F̃2) to a machine learning model (PM), - receiving from the machine learning model (PM) a third representation (F̃3*) of the examination region, the third representation (F̃3*) representing the examination region after administration of a third amount of contrast agent, - outputting and / or storing a representation (Õ3*) of the examination region in a real-space depiction, characterized in that the first representation (F̃1), the second representation (F̃2) and the third representation (F̃3*) represent the examination region of the examination object in frequency space, the method further comprising: - generating the representation (Õ3*) of the examination region in the real-space depiction on the basis of the third representation (F̃3*).

2. Computer-implemented method according to Claim 1, wherein the step of feeding at least part of the first representation (F̃1) and at least part of the second representation (F̃2) to a machine learning model (PM) comprises: - specifying a region (A) in the received first representation (F̃1) and in the received second representation (F̃2), the specified region (A) comprising the centre of the frequency space, - reducing the first representation (F̃1) and the second representation (F̃2) to the specified region (A), a reduced first representation (F̃1red) and a reduced second representation (F2red) being obtained, - feeding the reduced first representation (F̃1red) and the reduced second representation (F2red) to the machine learning model (PM).

3. Computer-implemented method according to Claim 2, wherein the step of generating a representation (Õ3*) of the examination region in a real-space depiction on the basis of the third representation (F̃3*) comprises: - supplementing the third representation (F̃3*red) by those parts (F̃1DI) of the received first representation (F̃1) and / or the received second representation (F̃2) that do not lie in the specified region (A), - transforming the supplemented third representation, obtained after the third representation (F̃3*red) has been supplemented, into the representation (Õ3*) of the examination region in the real-space depiction.

4. Computer-implemented method according to any of Claims 1 to 3, wherein the first representation (F̃1) represents the examination region without contrast agent or after administration of the first amount of a contrast agent, wherein the second representation (F̃2) represents the examination region after an administration of a second amount of the contrast agent, the second amount being unequal to the first amount, wherein the third representation (F̃3*) represents the examination region after administration of a third amount of the contrast agent, the third amount being unequal to the first amount and / or unequal to the second amount.

5. Computer-implemented method according to any of Claims 1 to 4, wherein the first amount of contrast agent is greater than or equal to zero, wherein the second amount of contrast agent is greater than the first amount of contrast agent and wherein the third amount of contrast agent is greater than the second amount of contrast agent.

6. Computer-implemented method according to any of Claims 1 to 5, wherein the first representation (F̃1) represents the examination region without contrast agent or after administration of a first amount of a first contrast agent, wherein the second representation (F̃2) represents the examination region after an administration of a second amount of a second contrast agent, wherein the third representation (F̃3*) represents the examination region after administration of a third amount of a third contrast agent, the third contrast agent being different from the first contrast agent and / or different from the second contrast agent or the second contrast agent being different from the first contrast agent.

7. Computer-implemented method according to any of Claims 1 to 6, wherein the first representation (F̃1) and the second representation (F̃2) are the result of a radiological examination.

8. Computer-implemented method according to Claim 7, wherein the radiological examination is a magnetic resonance imaging examination, a computed tomography examination or an ultrasound examination.

9. Computer-implemented method according to any of Claims 1 to 8, wherein the first representation (F̃1) and the second representation (F̃2) are k-space data of a magnetic resonance imaging examination.

10. Computer-implemented method according to any of Claims 1 to 9, wherein the first representation (F̃1) and the second representation (F̃2) are Fourier-transformed real-space depictions.

11. Computer-implemented method according to any of Claims 1 to 7, further comprising the steps of: - receiving a first real-space representation (O1) of the examination region of the examination object, the first real-space representation (O1) representing the examination region without contrast agent or after administration of a first amount of contrast agent, - receiving a second real-space representation (O2) of the examination region of the examination object, the second real-space representation (O2) representing the examination region after an administration of a second amount of contrast agent, - generating the first representation (F̃1) of the examination region of the examination object in frequency space from the first real-space representation (O1), preferably by means of Fourier transform (FT). - generating the second representation (F̃2) of the examination region of the examination object in frequency space from the second real-space representation (O2), preferably by means of Fourier transform (FT).

12. Computer-implemented method according to any of Claims 1 to 11, further comprising the step of: training the machine learning model, the training comprising the following sub-steps: - receiving a training data set, the training data set comprising, for a multiplicity of examination objects, in each case i) a first reference representation (F1) of an examination region of the examination object in frequency space, ii) a second reference representation (F2) of the examination region of the examination object in frequency space and iii) a third reference representation (F3) of the examination region of the examination object in frequency space, the first reference representation (F1) representing the examination region after administration of a first amount of a contrast agent, it also being possible for the first amount to be zero, the second reference representation (F2) representing the examination region after an administration of a second amount of the contrast agent, the third reference representation (F3) representing the examination region after administration of a third amount of the contrast agent, - specifying a region (A) in the reference representations, the specified region (A) comprising the centre of the frequency space, - reducing the representations (F1, F2, F3) to the specified region (A), a reduced first reference representation (F1red), a reduced second reference representation (F2red) and a reduced third reference representation (F3red) being obtained for each examination object, - for each examination object: feeding the reduced first reference representation (F1red) and the reduced second reference representation (F2red) to the machine learning model (PM), the machine learning model (PM) being trained to generate a reduced representation (F3*red) of the examination region in frequency space after administration of the third amount of contrast agent on the basis of the reduced first reference representation (F1red) and the reduced second reference representation (F2red), the training comprising the minimization of a loss function, the loss function quantifying deviations of the generated reduced representation (F3*red) of the examination region from the reduced third reference representation (F3red).

13. 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 two representations of an examination region of an examination object, a first representation (F̃1) and a second representation (F̃2), the first representation (F̃1) representing the examination region without contrast agent or after administration of a first amount of contrast agent, the second representation (F̃2) representing the examination region after an administration of a second amount of contrast agent, - wherein the control and calculation unit (12) is configured to feed at least part of the first representation (F̃1) and at least part of the second representation (F̃2) to a machine learning model (PM), - wherein the control and calculation unit (12) is configured to receive from the machine learning model (PM) a third representation (F̃3*) of the examination region, the third representation (F̃3*) representing the examination region after administration of a third amount of contrast agent, - wherein the control and calculation unit (12) is configured to generate on the basis of the third representation (F̃3*) a representation (Õ3*) of the examination region in real space, - wherein the control and calculation unit (12) is configured to cause the output unit (13) to output and / or store a representation (Õ3*) of the examination region in real space, characterized in that the first representation (F̃1), the second representation (F̃2) and the third representation (F̃3*) represent the examination region of the examination object in frequency space, wherein the control and calculation unit (12) is further configured to generate on the basis of the third representation (F̃3*) the representation (Õ3*) of the examination region in real space.

14. Computer program that can be loaded into a working memory of a computer system (10), where it causes the computer system (10) to execute the following steps: - receiving a first representation (F̃1) of an examination region of an examination object, the first representation (F̃1) representing the examination region without contrast agent or after administration of a first amount of contrast agent, - receiving a second representation (F̃2) of the examination region of the examination object, the second representation (F̃2) representing the examination region after an administration of a second amount of contrast agent, - feeding at least part of the first representation (F̃1) and at least part of the second representation (F̃2) to a machine learning model (PM), - receiving from the machine learning model (PM) a third representation (F̃3*) of the examination region, the third representation (F̃3*) representing the examination region after administration of a third amount of contrast agent, - outputting and / or storing a representation (Õ3*) of the examination region in a real-space depiction, characterized in that the first representation (F̃1), the second representation (F̃2) and the third representation (F̃3*) represent the examination region of the examination object in frequency space, wherein the computer program also causes the computer system (10) to execute the following step: - generating the representation (Õ3*) of the examination region in the real-space depiction on the basis of the third representation (F̃3*).