Predicting a representation of the test region of a test object in one state of a sequence of states

Abstract

The present invention relates to the technical field of radiology, in particular to the field of assisting radiologists performing radiological examinations by using artificial intelligence methods. The present invention also relates to the training of a machine learning model and the use of the trained model for predicting a representation for an examination region in one or more states of a sequence of states in a radiological examination.

Description

[Technical field]

[0001] The present invention relates to the technical field of radiology, in particular to the field of assisting radiologists in performing radiological examinations by using artificial intelligence methods. The present invention relates to the training of machine learning models and the use of the trained models for predicting representations of examination regions at one or more states of a sequence of states in a radiological examination. [Background technology]

[0002] The use of diagnostic imaging to follow processes over time within the human or animal body plays an important role, inter alia, in the diagnosis and / or treatment of disease.

[0003] One example that may be mentioned is the detection and differential diagnosis of focal liver lesions by dynamic contrast-enhanced magnetic resonance imaging (MRI) using hepatobiliary contrast agents.

[0004] Hepatobiliary contrast agents such as Primovist® can be used to detect tumors in the liver. Although blood is supplied to healthy liver tissue mainly via the hepatic portal vein, most primary tumors are supplied by the hepatic artery. Therefore, after a bolus of contrast agent is injected intravenously, a time delay can be observed between the signal increase in the healthy liver parenchyma and the signal increase in the tumor.

[0005] Besides malignant tumors, benign lesions such as cysts, hemangiomas, and focal nodular hyperplasia (FNH) are commonly found in the liver. These need to be differentiated from malignant tumors in order to plan appropriate treatment. Primovist® can be used to differentiate between benign and malignant focal liver lesions. T1-weighted MRI provides information on the characteristics of the lesions. Differentiation is achieved by exploiting the differences in blood supply to the liver and tumors, and by exploiting the temporal profile of contrast enhancement.

[0006] When contrast enhancement with Primovist® is obtained during the wash-in phase, what is observed is a typical perfusion pattern that provides information about the characterization of the lesion. Visualizing angiogenesis helps to characterize the type of lesion and to determine the spatial relationship of the tumor to the blood vessels.

[0007] In T1-weighted MRI images, Primovist® produces clear signal enhancement in healthy liver parenchyma 10-20 minutes after injection (hepatic contrast phase), whereas lesions containing no or only a few hepatocytes, e.g. metastases or moderately to poorly differentiated hepatocellular carcinomas (HCCs), appear as darker areas.

[0008] Thus, although tracking the diffusion of contrast agent over time provides an excellent method for detecting and differentially diagnosing focal liver lesions, the examination requires a relatively long period of time, during which the patient should be prevented from moving as much as possible in order to minimize motion artifacts in the MRI images, which can be uncomfortable for the patient.

[0009] In order to reduce the time the patient spends in the MRI scanner, International Publication No. 2021 / 052896 proposes that one or more MRI images during the hepatic contrast phase are calculated (predicted) based on MRI images from one or more preceding phases, rather than being generated by measurement.

[0010] The approach described in the previous WO 2021 / 052896 involves training a machine learning model to use MRI images of an examination area before and / or immediately after administration of a contrast agent as a basis for predicting MRI images of the examination area at a later time, such that the model is trained to map a plurality of MRI images as input data onto an MRI image as output data. Summary of the Invention

[0011] Proceeding from the described prior art, the object of the present invention is to improve the predictive quality of machine learning models and / or to create models that learn the dynamics of the diffusion of contrast agents in an examination area, so that they can be used in various ways.

[0012] This object is achieved by the subject matter of the independent claims. Preferred embodiments can be found in the dependent claims and also in the description and the drawings.

[0013] In a first aspect, the present invention provides a computer-implemented method for training a machine learning model, the method comprising: receiving training data; The training data is a representation of the inspection area for a large number of inspection objects. T R1~ T R n where n is an integer equal to or greater than 3; ·Each expression T R i is a set of multiple states Z1 to Z n One state Z of the sequence consisting of iwhere i is an index that runs through the integers 1 to n; -Training a machine learning model, Model represents T Starting from R1, for each test object in the set of test objects, the expression T R2 * ~ T R n * The representation is trained to generate T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index going through the numbers 3 through n, and the expression T R k * is state Z k represents the inspection area in T R k-1 * is state Z k-1 represents the inspection area in state Z k-1 is state Z k It is the direct precursor to Generated Representation T R j * and T R j The difference between is quantified by a loss function, where j is an index that runs from 2 to n, The difference is minimized using an optimization technique by changing the model parameters of the machine learning model; - storing the trained machine learning model and / or utilizing the machine learning model for prediction.

[0014] The present invention further provides a computer-implemented method for predicting one or more representations for an area of ​​interest in an object of interest, the method comprising: - an expression R relating to the inspection area in the inspection object p The purpose of the present invention is to receive ·Expression R p is a set of multiple states Z1 to Z n One state Z of the sequence consisting of p represents the inspection area in where p is an integer less than n, where n is an integer equal to or greater than 3; -Expression R p to a trained machine learning model, A trained machine learning model generates a first representation based on the training data. T Starting from R1, n-1 representations T R2 * ~ T R n * It is trained to generate a sequence of ·1st expression T R1 represents the inspection region in the first state Z1, and each generated representation T R j * is state Z j where j is an index going through the numbers 2 to n, ·Expression T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index running from 3 to n; - one or more representations R for the inspection region from the machine learning model p+q * The purpose of the present invention is to receive One or more expressions R p+q * Each of these is in state Z p+q where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; - one or more expressions R p+q * and outputting and / or storing and / or transmitting the

[0015] The present invention further provides a computer system comprising: An input unit; A control calculation unit; an output unit; The control calculation unit is The input unit is provided with a representation R p where the representation R p is a set of multiple states Z1 to Z n One state Z of the sequence consisting of p where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R p to a trained machine learning model, where the trained machine learning model is configured to provide a first representation based on the training data. T Starting from R1, n-1 representations T R2 * ~ T R n * where the first representation T R1 represents the inspection region in the first state Z1, and each generated representation T R j * is state Z j where j is an index that runs from 2 to n, and T R2* is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index going through the numbers 3 through n, - one or more representations R for the inspection region from the machine learning model p+q * wherein the one or more representations R p+q * Each of these is in state Z p+q where q is an index running through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; - one or more expressions R p+q * to an output unit and / or for storage and / or transmission to another computer system.

[0016] The invention further provides a computer program product comprising a data memory having stored therein a computer program capable of being loaded into the working memory of a computer system, the computer program being adapted to cause the computer system to: - an expression R relating to the inspection area in the inspection object p where R p is a set of multiple states Z1 to Z n One state Z of the sequence consisting of p where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R p to a trained machine learning model, where the machine learning model generates the first representation based on the training data.T Starting from R1, n-1 representations T R2 * ~ T R n * where the first representation T R1 represents the inspection region in the first state Z1, and each generated representation T R j * is state Z j where j is an index that runs from 2 to n, and T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index going through the numbers 3 through n; - one or more representations R for the inspection region from the machine learning model p+q * where one or more representations R p+q * Each of these is in state Z p+q where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; - one or more expressions R p+q * and outputting and / or storing and / or transmitting the

[0017] The present invention further provides the use of a contrast agent in a radiological examination method comprising the steps of: - an expression R relating to the inspection area in the inspection object p where R pis a set of multiple states Z1 to Z n One state Z of the sequence consisting of p represents the examination region before or after administration of a contrast agent, where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R p to a trained machine learning model, where the machine learning model generates the first representation based on the training data. T Starting from R1, n-1 representations T R2 * ~ T R n * where the first representation T R1 represents the inspection region in the first state Z1, and each generated representation T R j * is state Z j where j is an index that runs from 2 to n, and T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index going through the numbers 3 through n; - one or more representations R for the inspection region from the machine learning model p+q * where one or more representations R p+q * Each of these is in state Z p+q where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; - one or more expressions R p+q* and outputting and / or storing and / or transmitting the

[0018] The present invention further provides an imaging agent for use in a radiological examination comprising: - an expression R relating to the inspection area in the inspection object p where R is the representation p is a set of multiple states Z1 to Z n One state Z of the sequence consisting of p represents the examination region before or after administration of a contrast agent, where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R p to a trained machine learning model, where the machine learning model generates the first representation based on the training data. T Starting from R1, n-1 representations T R2 * ~ T R n * where the first representation T R1 represents the inspection region in the first state Z1, and each generated representation T R j * is state Z j where j is an index that runs from 2 to n, and T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index going through the numbers 3 through n; - one or more representations R for the inspection region from the machine learning model p+q* where one or more representations R p+q * Each of these is in state Z p+q where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; - one or more expressions R p+q * and outputting and / or storing and / or transmitting the

[0019] The invention further provides a kit comprising an imaging agent and a computer program product, the computer program product including a computer program capable of being loaded into a working memory of a computer system, the computer program comprising: - an expression R relating to the inspection area in the inspection object p where R p is a set of multiple states Z1 to Z n One state Z of the sequence consisting of p represents the examination region before or after administration of a contrast agent, where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R p to a trained machine learning model, where the machine learning model generates the first representation based on the training data. T Starting from R1, n-1 representations T R2 * ~ T R n * where the first representation T R1 represents the inspection region in the first state Z1, and each generated representation T R j * is state Z jwhere j is an index that runs from 2 to n, and T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * are at least partially equivalent to the previously generated representations T R k-1 * where k is an index going through the numbers 3 through n; - one or more representations R for the inspection region from the machine learning model p+q * where one or more representations R p+q * Each of these is in state Z p+q where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; - one or more expressions R p+q * and outputting and / or storing and / or transmitting the DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] In the following, the present invention will be described more specifically without distinguishing between the subject matter of the present invention (training method, prediction method, computer system, computer program product, use, contrast agent for use, kit). Rather, the following descriptions are intended to apply equally to all subject matter of the present invention, regardless of the context in which they appear (training method, prediction method, computer system, computer program product, use, contrast agent for use, kit).

[0021] When steps are described in a certain order in this specification or claims, this does not necessarily mean that the invention is limited to the described order. Rather, it is envisioned that the steps may be performed in a different order or may be performed in parallel with each other, except when a step is based on another step, in which case the step based on the previous step must be performed later (however, this will be clear in each individual case). Thus, the described order is a preferred embodiment of the invention.

[0022] With the aid of the present invention, a representation for a region of interest in an object of interest can be predicted.

[0023] A "test subject" is typically a living organism, preferably a mammal, and most preferably a human.

[0024] An "examination area" is a part of the subject being examined, for example an organ or part thereof, such as the liver, brain, heart, kidneys, lungs, stomach, intestines or parts of such organs, or multiple organs or other parts of the body.

[0025] In a preferred embodiment of the invention, the examination area is the human liver or a part thereof.

[0026] The examination area, also called the field of view (FOV), is in particular the volume that is imaged in a radiological image. The examination area is typically defined by a radiologist, for example on a localizer image. Of course, the examination area can also alternatively or additionally be defined in an automated manner, for example based on a selected protocol.

[0027] The representation relating to an examination area is preferably a medical image. The representation relating to an examination area is preferably the result of a radiological examination.

[0028] "Radiology" is a branch of medicine that primarily uses electromagnetic radiation and mechanical waves (including, for example, ultrasound) for diagnostic, therapeutic, and / or scientific purposes. Besides X-rays, other ionizing radiations, such as gamma radiation or electrons, are also used. Since imaging is an important application, other imaging methods, such as ultrasound and magnetic resonance imaging (nuclear magnetic resonance imaging), are also counted as radiology, even though they do not use ionizing radiation. Thus, the term "radiology" in the context of the present invention includes, inter alia, imaging methods such as computed tomography, magnetic resonance imaging, and ultrasound.

[0029] In a preferred embodiment of the invention, the radiological examination is a computed tomography examination or a magnetic resonance imaging examination.

[0030] Computed tomography (CT) is an X-ray imaging technique (cross-sectional imaging) that produces cross-sectional images of the human body. Compared to conventional X-ray images, which usually only allow the identification of rough structures and bones, CT images capture details even in soft tissues where the contrast difference is small. The X-ray tube produces a so-called X-ray fan beam, which penetrates the body and is attenuated to different degrees by different structures such as organs and bones. A receiving detector located opposite the X-ray emitter receives signals of different intensities and transfers them to a computer, which compiles a cross-sectional image of the body from the received data. Computed tomography images (CT images) can be viewed in 2D and 3D. In order to make structures in the human body (e.g. blood vessels) more clearly identifiable, a contrast agent can be injected, for example into a vein, before the CT image is produced.

[0031] Magnetic resonance imaging, or MRI for short, is an imaging technique used especially in medical diagnostics to visualize the structure and function of tissues and organs within the human or animal body.

[0032] In MRI, the magnetic moments of protons in the examination volume are aligned in a basic magnetic field, resulting in a macroscopic magnetization along the longitudinal direction. This magnetization is then deflected from its rest position by irradiation with radio frequency (HF) pulses (excitation). The return from the excited state to the rest state (relaxation), i.e. the magnetization dynamics, is then detected as a relaxation signal by one or more radio frequency receiving coils.

[0033] For spatial encoding, rapidly switching magnetic gradient fields are superimposed onto the basic magnetic field. The captured relaxation signals, i.e. the detected MRI data, initially exist as raw data in frequency space and can be transformed into real space (image space) by a subsequent inverse Fourier transformation.

[0034] A representation of an examination region in the context of the present invention may be an MRI image, a computed tomography image, an ultrasound image or the like.

[0035] A representation of an examination region in the context of the present invention can be a representation in real space (image space), in frequency space or some other representation. Preferably, the representation of an examination region exists in a real space representation or exists in a form that can be transformed into a real space representation. A representation in real space is often also called a pictorial representation or image.

[0036] In a representation in real space, also referred to herein as a real space representation or representation, the examination area is typically represented by a number of image elements (pixels or voxels), which may for example be arranged in a raster array, where each image element represents a portion of the examination area and may be assigned a color or gray value. A format widely used in radiology for storing and processing the representation in real space 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.

[0037] A representation in real space can be transformed into a representation in frequency space, for example by a Fourier transform, and conversely a representation in frequency space can be transformed into a representation in real space, for example by an inverse Fourier transform.

[0038] In a representation in frequency space, also referred to herein as a frequency space representation or frequency space representation, the examination region is represented by a superposition of fundamental frequencies. For example, the examination region may be represented by a sum of sine and / or cosine functions with different amplitudes, different frequencies, and different phases. The amplitudes and phases may be plotted as a function of frequency, for example in a two-dimensional or three-dimensional representation. Usually, the lowest frequency (the origin) is located in the center. The further away from this center, the higher the frequency. Each frequency can be assigned an amplitude that represents the frequency in the frequency space representation, and a phase that indicates the degree of shift of the corresponding wave relative to the sine or cosine wave.

[0039] Details regarding real space and frequency space representations, as well as the corresponding conversions between them, are provided in numerous publications, see, for example, https: / / see.stanford.edu / materials / lsoftaee261 / book-fall-07.pdf.

[0040] In the context of the present invention, a representation represents an examination region at one state of a sequence of states.

[0041] The sequence of states is preferably a time series. States that follow each other directly in the time series may always have the same time interval between them, or they may have different time intervals. Mixed formats are also envisaged.

[0042] This will be explained with reference to FIG.

[0043] Figure 1 shows two timelines (t = time), a first timeline (a) and a second timeline (b). Both timelines are labeled with defined time points t1, t2, t3, t4, respectively. At each time point, the inspection area can have different states, i.e. each time point can represent one state for the inspection area.

[0044] Time points t1, t2, t3, and t4 form, for each timeline, the time sequence: t1→t2→t3→t4. Time point t2 follows directly after time point t1, and the time interval between t1 and t2 is t2-t1, time point t3 follows directly after time point t2, and the time interval between t2 and t3 is t3-t2, and time point t4 follows directly after time point t3, and the time interval between t3 and t4 is t4-t3.

[0045] In the case of the first timeline (a), the time interval between immediately adjacent time points is the same for all immediately adjacent time points, i.e. t2-t1=t3-t2=t4-t3.

[0046] In the case of the second timeline (b), the time intervals between directly consecutive points in time are different for all directly consecutive points in time, that is, t2 - t1 (not equal to) t3 - t2 (not equal to) t4 - t3. In this example, for directly consecutive points in time along the second time axis (b), the time intervals are increasing, that is, t2 - t1 < t3 - t2 < t4 - t3. However, it is also assumed that the time intervals may decrease, or that the time intervals may first increase and then decrease, or that the time intervals may first decrease and then increase, or that the time intervals may have some other distribution along the time axis.

[0047] It should be noted that the time axis does not necessarily have to extend in the direction of increasing time. This means that when viewed from time point t2, time point t1 does not necessarily have to be located in the past, that is, it is also assumed that the time axis indicates decreasing time and that time point t1 is located in the future when viewed from time point t2. In other words, when there are a plurality of states forming a time series, the second state directly following the first state may be located in the future or in the past when viewed from the time point of the first state.

[0048] In the first step, a sequence consisting of a plurality of (preferably time-series) states may be defined. The sequence consisting of a plurality of states defines what training data to use for training a machine learning model and what representations the machine learning model can generate (predict). In other words, what is required for training a machine learning model is the representation of the inspection area in a number of inspection objects in a plurality of states defined by a sequence consisting of a plurality of states, and a trained machine learning model can usually generate only the representation of the state that was part of the training (exceptions to this rule are listed later in this specification).

[0049] Preferably, the sequence of states defines different conditions for the examination region before and / or during and / or after single or multiple administrations of contrast agent.

[0050] A "contrast agent" is a substance or mixture of substances that improves the visualization of bodily structures and functions in diagnostic radiological imaging procedures.

[0051] Examples of contrast agents can be found in the literature (e.g. ASL 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 / wp-content / uploads / 2017 / 10 / contrast-agents-tutorial.pdf; MR Nough et al.: Radiographic and magnetic resonances contrast agents: Essentials and tips for safe practices, World J Radiol. 2017 Sep. 28; 9(9): 339-349; LC Abonyi et al.: Intravascular Contrast Media in Radiography: (See 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).

[0052] Preferably, the contrast agent is an MRI contrast agent. MRI contrast agents exert their action by changing the relaxation time of the structure in which they are incorporated. Two groups of substances can be distinguished: paramagnetic and superparamagnetic. Both groups of substances have unpaired electrons that induce a magnetic field around the individual atoms or molecules. Superparamagnetic contrast agents cause a significant T2 shortening, whereas paramagnetic contrast agents mainly cause T1 shortening. The action of contrast agents is indirect, since they do not themselves emit a signal, but merely affect the signal intensity in their vicinity. 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®, and others), gadoteric acid (Dotarem®, Dotagita®, Cyclolux®), gadodiamide (Omniscan®), gadoteridol (ProHance®), and gadobutrol (Gadovist®).

[0053] Preferably, the MRI contrast agent is a hepatobiliary contrast agent. Hepatobiliary contrast agents have the characteristic of being specifically taken up by hepatocytes (liver parenchymal cells), accumulating in functional tissue (parenchymal tissue), and enhancing the contrast of healthy liver tissue. One example of a hepatobiliary contrast agent is disodium salt of gadoxetic acid (Gd-EOB-DTPA disodium), which is described in U.S. Pat. No. 6,039,931A and is commercially available under the trade names Primovist® and Eovist®.

[0054] After intravenous administration of a hepatobiliary contrast agent in the form of a bolus into a vein in the arm of a human, the contrast agent first reaches the liver via the arteries, which are visualized with contrast enhancement in the corresponding MRI images. The phase in which the hepatic arteries are visualized with contrast enhancement in the MRI images is called the "arterial phase."

[0055] The contrast agent then reaches the liver through the hepatic veins. The contrast of the hepatic veins reaches a maximum value, whereas the contrast of the hepatic artery is already reduced. The phase in which the hepatic veins are visualized with contrast enhancement in the MRI image is called the "portal venous phase".

[0056] The portal venous phase is followed by a "transitional phase" in which the contrast of the hepatic artery decreases further, as does the contrast of the hepatic vein. When a hepatobiliary contrast agent is used, the contrast of healthy liver tissue gradually increases during the transitional phase.

[0057] The arterial, portal, and transitional phases are also collectively referred to as the "dynamic phases."

[0058] Hepatobiliary contrast agents produce clear signal enhancement in healthy liver parenchyma 10-20 minutes after injection. This phase is called the "hepatic contrast phase." Because contrast agents are only slowly excreted from hepatocytes, the hepatic contrast phase may last for more than 2 hours.

[0059] The mentioned phases are described in more detail, for example, in the following publications: J. Magn. Reson. Imaging, 2012, 35(3): 492-511, doi:10.1002 / jmri.22833; Clujul Medical, 2015, Vol.88 No.4: 438-448, DOI: 10.15386 / cjmed-414; Journal of Hepatology, 2019, Vol.71: 534-542, http: / / dx.doi.org / 10.1016 / j.jhep.2019.05.005).

[0060] Each state from a sequence of states can be, for example, -Before administration of hepatobiliary contrast agents, - In the arterial phase, - In the portal phase, - in the transition phase and / or - During the hepatography phase, The condition of the liver as the examination area may also be examined.

[0061] It is envisioned that there may be a greater number of states, or that there may be a lesser number of states.

[0062] Each of the described phases will now be explained in more detail with reference to FIG. 2. FIG. 2 shows, in a schematic way, the time profile (t=time) of the signal intensity I induced by a hepatobiliary contrast agent in the hepatic artery A, the hepatic vein V and healthy liver cells L in a dynamic contrast-enhanced MRI study. The signal intensity I is positively correlated to the contrast agent concentration in each of the described regions. After an intravenous bolus injection into the human arm, the concentration of the contrast agent first rises in the hepatic artery A (dashed curve). The concentration passes through a maximum value and then decreases. The concentration in the hepatic vein V rises more slowly than in the hepatic artery and reaches its maximum value at a later time point (dotted curve). The contrast agent concentration in the liver cells L rises more slowly (solid curve) and reaches its maximum value at a much later time point (not shown in FIG. 2). Several characteristic time points can be defined: at time point TP1, the contrast agent is administered intravenously as a bolus. Since the administration of the contrast agent itself requires a certain time span, the time point TP1 preferably defines the time point when the administration is completed, i.e. when the contrast agent is completely introduced into the examination object. At the time point TP2, the signal intensity of the contrast agent in the hepatic artery A reaches its maximum value. At the time point TP3, the curve of the signal intensity for the hepatic artery A and the curve of the signal intensity for the hepatic vein V intersect. At the time point TP4, the signal intensity of the contrast agent in the hepatic vein V passes through its maximum value. At the time point TP5, the curve of the signal intensity for the hepatic artery A and the curve of the signal intensity for the healthy liver cells L intersect. At the time point TP6, the concentration in the hepatic artery A and the concentration in the hepatic vein V have decreased to a level that no longer causes a measurable contrast enhancement.

[0063] Each state in the sequence of states may include a first state existing before time TP1, a second state existing at time TP2, a third state existing at time TP3, a fourth state existing at time TP4, a fifth state existing at time TP5, and / or a sixth state existing at and / or after time TP6.

[0064] In general, the first state may be a state of the examination region at a first time point before administration of a contrast agent, the second state may be a state of the examination region at a second time point after administration of a contrast agent; the third state may be a state of the examination region at a third time point after administration of a contrast agent; the fourth state may be a state of the examination region at a fourth time point after administration of a contrast agent; And so on.

[0065] As mentioned above, all the time points that are directly successive to one another, or some of them, may have a fixed time interval from one another and / or may have a variable time interval from one another.

[0066] According to the present invention, there are at least three states, and preferably the number of states is between 3 and 100. However, the number of states is not limited to 100.

[0067] For each state in a sequence of states, there may be one or more representations that represent the inspection region at the corresponding state.

[0068] Typically, in a first state there is at least one first representation representing the inspection area in the first state, in a second state there is at least one second representation representing the inspection area in the second state, in a third state there is at least one third representation representing the inspection area in the third state, and so on.

[0069] For ease of understanding of this specification, the present invention will be described primarily based on one representation for each state, but this should not be construed as limiting the present invention.

[0070] Returning to the example above, for example, - one or more first representations of the liver prior to administration of a hepatobiliary imaging agent; one or more second representations representative of the liver during the arterial phase; one or more third representations representing the liver during the portal venous phase, one or more fourth representations representative of the liver during the transition phase, and / or There may be one or more fifth representations representing the liver during the hepatography phase.

[0071] With the aid of the present invention, one or more representations for an area of ​​interest in an object of study can be predicted, which represent the area of ​​interest in one state of a sequence of states.

[0072] The predictions are made with the help of machine learning models.

[0073] A "machine learning model" may be understood to mean a computer-implemented data processing architecture. The model may receive input data and provide output data based on the input data and based on model parameters. The model may learn the relationship between the input data and the output data through training. During training, the model parameters may be adjusted to provide a desired output for a particular input.

[0074] When training such a model, the model is presented with training data that it can learn from. A trained machine learning model is the result of the training process. In addition to the input data, the training data contains the correct output data (target data) that the model should generate based on the input data. During training, patterns that map the input data onto the target data are recognized.

[0075] In the training process, input data in the training data are input into the model, and the model generates output data. The output data is compared to target data. Model parameters are changed to reduce the difference between the output data and the target data to a (specified) minimum value.

[0076] The difference can be quantified using a loss function. Using this type of loss function, the loss for a given set of output data and target data can be calculated. The goal of the training process is to change (fit) the machine learning model parameters such that the loss for all pairs of training data sets is reduced to a (prescribed) minimum value. Adjusting the model parameters to reduce the loss can be done with optimization techniques such as gradient methods.

[0077] For example, if the output and target data are numerical, the loss function can be the absolute difference between them, where a large absolute value of the loss means that one or more model parameters need to be significantly changed.

[0078] For output data in vector format, one can choose a difference metric between vectors as the loss function, such as mean squared error, cosine distance, norm of difference vectors such as Euclidean distance or Chebyshev distance, Lp norm of difference vectors, weighted norm, or any other type of difference metric for two vectors.

[0079] In the case of high dimensional outputs, such as two-dimensional, three-dimensional or higher dimensional outputs, for example, an element-wise difference metric can be used. Alternatively or additionally, the output data can be transformed, for example into a one-dimensional vector, before computing the loss.

[0080] In this case, the machine learning model is trained on the training data to generate a representation of the inspection area at one state of a sequence of states based on a representation of the inspection area at a previous state in the sequence of states, and starting from the first state, the model successively (iteratively) generates representations of the inspection area at subsequent states, each representation of the inspection area at each state being generated at least in part based on the (predicted) representation for a previous state.

[0081] This will be explained on the basis of an example of three states, without any intention of limiting the invention to this embodiment, and in this explanation reference is made to FIG.

[0082] It should be noted that in this description, expressions relating to the inspection area are identified by the letter R. The letter R can be followed by an index, for example R1, R2, R3, or R i and so on. The index added as a suffix indicates what is represented by the corresponding expression. Expression R1 represents the inspection area of ​​the object to be inspected at a first state Z1 of the sequence of states, expression R2 represents the inspection area of ​​the object to be inspected at a second state Z2 of the sequence of states, expression R3 represents the inspection area of ​​the object to be inspected at a third state Z3 of the sequence of states, and so on. In general, expression R i is a state Z of a sequence of states. i represents an inspection region in an inspection object in, where i is an index that passes through the integers i to n, where n is an integer equal to or greater than 3. The expression "index i passes through the integers a to b" means that i takes on the values ​​a to b in succession, that is, i first takes on the value a (i=a), then the value a+1 (i=a+1), and so on until i reaches the value b (i=b). "Each expression R i * However, there are multiple states Z1 to Z n One state Z of the sequence consisting of iThe expression "inspection area in state Z1 to Z2" means n If (in the mathematical sense) forms a sequence, then the representation R1 * represents the inspection region in state Z1, and the representation R2 * represents the search region at state Z2, etc. In particular, in the claims and drawings, the representations used for training and the representations generated in training may be, for example, T R1, T R2, and T R k-1 * It is added as a prefix, as in T In particular, in the claims and drawings, the representations generated by the machine learning model are identified by R1 * , R2 * , and T R k-1 * A superscript asterisk, as in * The labels set forth herein serve merely for the sake of clarity, and in particular to avoid clarity-related objections in the patent granting process.

[0083] FIG. 3 shows three representations: T R1 and the second representation T R2 and the third representation T R3 and the first expression T R1 represents the inspection area of ​​the inspection object in the first state Z1, and the second representation T R2 represents the inspection area in the second state Z2, and the third representation T R3 represents the inspection area in the third state Z3. The three states form a multi-state sequence, namely, first state (Z1) → second state (Z2) → third state (Z3).

[0084] Three Expressions T R1, T R2, TR3 forms a data set in the training data TD. The training data TD contains a multiplicity of such data sets. The term "multiplicity" preferably means more than 100. Each data set usually (although not necessarily) contains one representation for each state, and for three states, three representations of the test area. The test area is usually always the same, and each data set contains a representation of the test area in three states, which are also usually the same for all data sets, namely the first state, the second state and the third state. Only the test object may differ. Each data set is usually obtained from a different test object. The description in this paragraph is not limited to the example shown in FIG. 3, but is generally applicable.

[0085] In the first step (A), the first representation T R1 is fed to a machine learning model M, which is a first representation T Based on R1, and based on the model parameters MP, the output T R2 * (Step (B)) is configured to generate the output T R2 * is the second representation T The intention is to approximate R2 as closely as possible, and ideally the output T R2 * is the second representation T In other words, the output T R2 * is the predicted second representation. Output T R2 * is the (actual) second representation T The loss value LV2 is then compared to R2 and the difference is quantified using the loss function LF2. When multiple test objects are present, the loss value LV2 is calculated by the loss function LF2. T R2 * and the second representation T For each pair of R2, it can be calculated.

[0086] Examples of loss functions that can be commonly used to implement the present invention (not limited to the example of FIG. 3) are L1 loss function, L2 loss function, Lp loss function, Structural Similarity Index Measure (SSIM), VGG loss function, perceptual loss function, or combinations of the above functions with each other or with other loss functions. Further details on loss functions can be found, for example, in scientific literature (see, for example, R. Mechrez et al.: The Contextual Loss 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).

[0087] output T R2 * is re-fed to the machine learning model M in step (C). Although FIG. 3 may suggest otherwise, in step (A) the first representation T A machine learning model M fed with R1 and the generated representations T R2 * is the same model as the machine learning model M provided in step (C), i.e. the machine learning model M is configured not only to generate a predicted second representation based on the first representation, but also to derive a third representation from the (predicted and / or actual) second representation. T Output that approximates R3 as closely as possible T R3 * It is also configured to generate the output T R3 * is the predicted third representation. Output T R3 * is the third expression T The output is compared to R3 using the loss function LF3. T R3 * and the third expression TThe difference from R3 can be quantified. In the case of multiple inspection objects, the loss value LV3 can be expressed as a third representation T R3 and the generated representation T R3 * For each pair consisting of , it can be determined.

[0088] Preferably, the machine learning model is trained end-to-end, i.e., the machine learning model is trained simultaneously to generate a predicted second representation based on the first representation and to generate a predicted third representation based on the predicted second representation. Preferably, this is achieved by using a loss function that takes into account both the difference between the predicted second representation and the second representation and the difference between the predicted third representation and the third representation.

[0089] Loss function LF2 may be used to quantify the difference between the second representation and the predicted second representation, and loss function LF3 may be used to quantify the difference between the third representation and the predicted third representation. The loss function LF taking into account both differences may for example be the sum of the individual loss functions, i.e. LF=LF2+LF3. Also, the components formed by the individual loss functions LF2 and LF3 in the overall loss function LF may be weighted differently, i.e. LF=w2·LF2+w3·LF3, where w2 and w3 are weighting factors that may for example take values ​​between 0 and 1. For example, in case of missing representations in the dataset, the value zero may be used for the weighting factors (further details can be found herein below).

[0090] The training method shown in FIG. 3 is a preferred embodiment of the present disclosure, receiving training data, The training data is a representation of the inspection area for a large number of inspection objects. T R1~ T R3 included, ·Expression T R1 represents the inspection area in the first state Z1, and is represented by TR2 represents the inspection area in the second state Z2, and is expressed as T R3 represents an inspection region in a third state Z3, and states Z1, Z2, and Z3 form a sequence of states; - training a machine learning model, the training for each test object of the multiple test objects being ·Expression T R1 to a machine learning model, where the machine learning model is, at least in part, T Based on R1 and based on model parameters, the expression T R2 * and ·Expression T R2 * to a machine learning model, where the machine learning model is, at least in part, T R2 * Based on,and,based on the model parameters, the expression T R3 * and further configured to generate ·Expression T R2 and the generated representation T R2 * Quantify the difference and express T R3 and the generated representation T R3 * Quantifying the difference between Minimizing the difference between them by modifying model parameters; - storing the trained machine learning model and / or using the machine learning model for prediction.

[0091] When the trained machine learning model is (later) used for predicting new representations, it may predict the third representation based on the first representation. For this purpose, in principle, it is possible to train the model to generate the third representation directly based on the first representation. However, according to the invention, the machine learning model is trained to generate the third representation in two steps, rather than in a single step based on the first representation, such that the second representation is predicted in a first step and then the third representation is predicted based on the second representation. The advantage of the approach according to the invention compared to the above-mentioned "training directly" case is, among other things, that a more accurate prediction can be achieved since additional training data (a second representation representing the inspection area in the second state) can be used. Furthermore, instead of learning how one representation is mapped onto another representation (or how several representations are mapped onto one representation, as in the case of WO 2021 / 052896), the model learns the dynamic behavior of the inspection area, i.e. the dynamics of the passing through states. The more states covered by the training data, the more accurately the model can learn the dynamics from one state to a subsequent state. Furthermore, a model trained in this manner can be used to predict representations for states for which training data is unavailable, by using the model multiple times (iteratively) to predict representations for further subsequent states. This extrapolation to states not considered during training is described in more detail below in this specification.

[0092] When the machine learning model shown in FIG. 3 has been trained, it can be used for prediction. This is illustrated in FIG. 4. In a first step (A), a first representation R1 is provided to the trained machine learning model M. The first representation R1 represents the inspection region of a new inspection object in a first state Z1. The term "new" means that no representation of the inspection object was used when training the machine learning model M. The model M, in step (B), calculates a second representation R2 based on the first representation R1 and on the model parameters MP. * Generate the second representation R2 * represents the inspection area of ​​the inspection object in the second state Z2. In the third step (C), the generated second representation R2 * is fed to the machine learning model M. The model M is fed in step (D) to the second representation R2 * Based on, and based on the model parameters MP, the third representation R3 * Generate the third representation R3 * represents the inspection area of ​​the inspection object in the third state Z3.

[0093] The prediction method shown in FIG. 4 and based on the training method shown in FIG. 3 are preferred embodiments of the present disclosure, - receiving a representation R1 of an examination area, where the representation R1 represents the examination area in a first state Z1 of a sequence of states Z1, Z2, Z3; - using the machine learning model to generate a representation R2 regarding the inspection region based at least in part on the representation R1 * where R2 is the representation * represents an inspection region in a second state Z2, the second state Z2 being the direct successor to the first state Z1 in the sequence of states; - Use a machine learning model to at least partially express R2 * Based on the expression R3 regarding the inspection area * where R3 is a representation of *represents an inspection region at a third state Z3, the third state Z3 being the direct succession in the sequence of states to the second state Z2; -Representation R3 * and / or the step of outputting the expression R3 * and / or the step of storing the representation R3 * and transmitting, where the machine learning model is trained based on training data, the training data including a number of datasets, each dataset including a representation. T R1 and the expression T R2 and expression T R3, where the expression T R1 represents the inspection region at the first state Z1 of the sequence of states, and is represented by T R2 represents the inspection region at the second state Z2 of the sequence of states, and is represented by T R3 represents an inspection region at a third state Z3 of the sequence of states, where the machine learning model at least partially represents, for each inspection object of the multiple inspection objects, T Representation based on R1 T R2 * and further at least partially expressing T R2 * Based on the expression T R3 * where the training is the representation T R2 and Expression T R2 * Difference between, and expression T R3 and Expression T R3 * and modifying the model parameters to minimize the difference.

[0094] In general, the machine learning model according to the present invention has a number of states Z1 to Z n One state Z of the sequence consisting of iThe method may be trained to predict an expression for an examination region in an examination object in x, where n is an integer greater than or equal to 3, and i is an index running through the numbers 2 to n.

[0095] The machine learning model is a first representation of the inspection region in state Z1. T Starting from R1, states Z2 to Z n A set of expressions for the inspection area in T R1 * ~ T R i * may be trained to generate successive representations, where each representation for a state is generated based, at least in part, on a corresponding representation previously generated for the immediately preceding state.

[0096] This will be explained in more detail using the example of Figure 5. Figure 5 can be understood as an extension of the scheme shown in Figure 3 from 3 states to n states, where n is an integer equal to or greater than 3. Preferably, the number n is in the range 3-100.

[0097] FIG. 5 shows a machine learning model M. Although model M is specified three times, it is always the same model. The machine learning model M is trained to predict n-1 representations in succession, where n is an integer greater than or equal to 4. The model is given input data: T R1 is provided. First Representation T R1 represents an inspection region in the inspection object at the first state Z1. The machine learning model M is based, at least in part, on the first representation T Based on R1 and based on the model parameters MP, the output T R2 * It is configured to generate the output T R2 * is a predicted second representation that represents the inspection region at the second state Z2. The model is based, at least in part, on the output T R2 *Based on and based on the model parameters MP, the output T R3 * The output is further configured to generate T R3 * is a predicted third representation that represents the inspection region at the third state Z3. The model is based, at least in part, on the output T R3 * and based on the model parameters MP, generating a further output indicative of a predicted representation for the examination region at a state subsequent to state Z3. T R n * This continues until the output T R n * is the nth state Z n is the predicted n-th representation of the test region in

[0098] State Z i (where i = 1 to n) is a sequence of states (Z1 → Z2 → Z3 → ... → Z n ) is formed.

[0099] In the example shown in FIG. T Not only R1, but also state Z j Further (real, measurement-generated) expressions for the test region, where j = 2 to n T R j (where j=2~n) also exists, and these expressions T R j can be used as ground truth data for training machine learning models. For example, the output T R j * (This is state Z j is the predicted j-th representation of the test region in T R j The difference between the loss function LF j That is, the loss function LF2 may be quantified using the expression T R2 and the generated representation T R2* The loss function LF3 can quantify the difference between the T R3 and the generated representation T R3 * and so on.

[0100] In end-to-end training, a total loss function LF that takes into account all individual differences may be used, e.g., it may be the sum of the individual differences.

[0101]

number

[0102] Alternatively, the sum may be a weighted sum.

[0103]

number

[0104] Here, the weighting factor w j can take on values ​​between 0 and 1, for example.

[0105] The advantage of weighting is that one state Z n Expression for expressing the inspection area in T R n * When generating state Z n One or more (generated) representations that represent the previous state of the model are given more or less weight compared to other representations. For example, the weighting factor w j , w2, w3, ..., w n By increasing the order of n A weighting factor may be given to the representation associated with the state closer to w2, w3, ..., w4. The weighting factor may be logarithmic, linear, squared, cubic, or exponential, or some other weighting factor. It is also contemplated to give less weight to subsequent states, in which case the weighting factors may be w2, w3, ..., w4.n Since it decreases in the order of, a larger weight is given to the representation related to the state closer to the initial state Z1. Even in such a case of decrease, the decrease may be logarithmic, linear, quadratic, cubic, or exponential, or may be some other decrease.

[0106] Also, it is assumed that the training data includes an incomplete data set. This means that some data sets regarding the inspection target do not include all the representations T R1~ T R n As will be described later in this specification, the incomplete data set may also be used for training the machine learning model. When the representation T R p is missing, the weight coefficient w T R p that weights the difference between the generated representation R p * and R may be set to zero, where p is an integer that can take values from 2 to n. p

[0107] Also, it is assumed that the training includes considering random numbers 1 <= j < k <= n and related partial sequences Z j 、Z j+1 、...、Z k-1 、Z k consisting of a plurality of states. The above-mentioned learning problems may be solved based on these partial sequences (optionally, changing from moment to moment). For example, it is assumed that a random initial state Z j (1 <= j <= n - 2) is determined, and the model always synthesizes representations regarding the subsequent two states Z j+1 、Z j+2 based on such an initial state.

[0108] In FIG. 5, also, the representation j representing the inspection area in the state Z T R j * ​However, the preceding state Z j-1 Expression for expressing the inspection area in T R j-1 * As well as the further preceding state Z j-2 , Z j-3 Further (generated) representations of the inspection regions in , ..., Z1 T R j-2 * and / or T R j-3 * And / or... T R1 * It is also shown diagrammatically that the expression T R3 * is the expression T R2 * Not just an expression T Even R1 can be generated by feeding it to the machine learning model M. T R4 * is the expression T R3 * Not just an expression T R2 * and / or representation T Even R1 can be generated by feeding it to the machine learning model M.

[0109] Additional information may also be incorporated into the machine learning model to generate the representation, such as information about the state in which the inspection region is located, and the representation for that state is used to generate a representation for a subsequent state. T R2 * is the expression T Not just R1, but expression T It can be generated by using information about which state the examination region represented by R1 is located in. T R3 * Also, expression T R2 *and based on information about the second state Z2.

[0110] If information about the states corresponding to the representations is provided during training and / or when using a trained machine learning model for prediction, then the machine learning model "knows" where each one is located in the sequence of states.

[0111] For example, a state may be represented by a numerical value or a vector. For example, states may be numbered consecutively, such that a first state is represented by the number 1, a second state is represented by the number 2, and so on.

[0112] Instead of or in addition to information about the respective corresponding states, times that may be assigned to the respective corresponding states may also be used in generating the representations. For example, a first state in a sequence of states may be assigned a time t=0. The time t=0 may be input into the machine learning model together with the first representation to predict the second representation. The second state may be assigned a number of seconds, minutes, or some other unit of time difference that indicates how much time has passed between the first and second states. This time difference may be provided to the machine learning model together with the predicted second representation (and optionally the first representation) to predict the third representation. The third state may also be assigned a time difference that indicates how much time has passed between the first and third states. This time difference may be provided to the machine learning model together with the predicted third representation (and optionally the first representation and / or the predicted second representation) to predict the third representation. And so on.

[0113] The machine learning model according to the present invention may also be understood as meaning a transformation that may be applied to a representation of an inspection area at a state of a sequence of states to predict a representation of the inspection area at a subsequent state of the sequence of states. The transformation may be applied singly (once) to predict a representation of the inspection area at a subsequent state, or the transformation may be applied multiple times (multiple times, iteratively) to predict a representation of the inspection area at a further downstream state in the sequence of states.

[0114] n states Z1 to Z n If there exists a sequence consisting of i In state Z i The expression R expresses the inspection area in i If exists, state Z 1+q A predicted representation R representing the inspection region in 1+q * To generate, a machine learning model M may be applied q times, starting from representation R1, where q can take values ​​from 1 to n-1. M q (R1)=R 1+q * q=1:M(R1)=R2 * q = 2:M(R2 * )=M(M(R1))=M 2 (R1)=R3 * q=3:M(R3 * )=M(M(R2 * ))=M(M(M(R1)))=M 3 (R1)=R4 * ... q=n-1:M(R n-1 * )=M(M(R n-2 * ))=...=M n-1 (R1)=R n *

[0115] This applies equally to the training method, where the transformation M over q times may be described by the following equation: M q ( T R1) = T R 1+q *

[0116] Generated Representation T R q * and the actual (measurement-generated) representation T R q A loss function that quantifies all differences between and may be expressed, for example, by the following equation: LV = w2 d(M( T R1), T R2) + w3 d(M 2 ( T R1), T R3)+...+w n d(M n-1 ( T R1), T R n )

[0117] where LV is the (actual, measured) representation T R1, T R2, ..., T R n d is the loss value generated for a dataset containing the predicted representation M( T R q-1 ) and T R q As previously described herein, this may be, for example, one of the following loss functions: an L1 loss function, an L2 loss function, an Lp loss function, a structural similarity index measure (SSIM), a VGG loss function, a perceptual loss function, or a combination thereof.

[0118] w2, w3, ..., w n are weighting factors also previously described herein.

[0119] n indicates the number of states.

[0120] It is also assumed that the loss value is the maximum difference calculated over the dataset, i.e. LV = max(w2·d(M( T R1), T R2);w3·d(M 2 ( T R1), T R3);...;w n d(M n-1 ( T R1), T R n )) Again, weighting factors can be used in this formula to give different weights to individual differences.

[0121] As already shown herein, it should be noted that in training a machine learning model, each dataset of training data does not need to contain a representation for all states that the model should learn. In principle, two representations per dataset are sufficient. This will be illustrated using an example. Assume that a machine learning model is trained to predict a representation for an inspection region at a state in a sequence of six states Z1 to Z6. Assume that training data including 10 datasets for 10 inspection objects is sufficient to train the machine learning model. Each dataset includes, for example, a representation for the inspection region at a different state. Dataset 1: T R1, T R3, T R4, T R5, T R6 Dataset 2: T R1, T R2, T R4, T R6 Dataset 3: T R1, T R2, T R3, T R4, T R5 Dataset 4: T R1, T R2, T R3, T R5, T R6 Dataset 5: T R2, T R3, T R5, T R6 Dataset 6: T R2, T R3, T R4, T R5, T R6 Dataset 7: T R2, T R3, T R5, T R6 Dataset 8: T R3, T R5, T R6 Dataset 9: T R3, T R4, T R5, T R6 Dataset 10: T R3, T R4, T R6

[0122] In our example, there is no single "perfect" dataset; that is, no dataset contains all possible representations. T R1, T R2, T R3, T R4, T R5, and T It does not include R6. Nevertheless, a machine learning model can be trained based on such training data to predict the expression at each state. This is another advantage of the present invention over the "direct training" method described above.

[0123] After a machine learning model has been trained, the model can be fed a new (i.e., unused in training) representation of the test region in a (new) test object at a state in a sequence of states, and the model can predict (generate) one or more representations for the test region at a subsequent state, or at multiple subsequent states, in the sequence of states.

[0124] This is illustrated, by way of example, diagrammatically in FIG. 6. In the example shown in FIG. p The expression R expresses the inspection area in p Starting from state Z, p+1 , Z p+2 , Z p+3 , Z p+4 The expression R expresses the inspection area in p+1 * , R p+2 * , R p+3 * , R p+4 * The state Z is a sequence of p+1 , Z p+2 , Z p+3 , Z p+4 forms a sequence of states, i.e., Z p+1 →Z p+2 →Z p+3 →Z p+4 It is formed.

[0125] In step (A), we assign a representation R to a machine learning model M. p is provided. Representation R p In addition to the above, for the machine learning model M, and as described herein, the state Z p and / or additional / other information may be provided.

[0126] In step (B), the machine learning model M is p+1 The expression R expresses the inspection area in p+1 * Generate.

[0127] In step (C), we apply the previously generated representation R to the machine learning model M. p+1 * is provided. Representation R p+1 * In addition to the above, for the machine learning model M, and for the state Z p+1 may be provided with information about the representation R and / or additional information. p But, and / or state Z p Information regarding the

[0128] Preferably, the machine learning model M comprises a memory S for storing the input data (preferably also the output data), so that the input data once entered and / or the output data generated are already available to the machine learning model without the need to receive or enter them again. This applies not only for the use of a trained machine learning model for prediction as described in this example, but also for the training of a machine learning model according to the invention.

[0129] In step (D), the machine learning model M changes the state Z based on the input data provided. p+2 The expression R expresses the inspection area in p+2 * Generate.

[0130] In step (E), we apply the previously generated representations R to the machine learning model M. p+2 * is provided. Representation R p+2 * In addition to the above, for a machine learning model M, we also have the state Z p+2 In addition, for a machine learning model M, information about the representation R may also be provided. p and / or expression R p+1 * But, and / or state Z p and / or State Z p+1Information regarding the

[0131] In step (F), the machine learning model M changes the state Z based on the input data provided. p+3 The expression R expresses the inspection area in p+3 * Generate.

[0132] In step (G), we apply the previously generated representations R to the machine learning model M. p+3 * is provided. Representation R p+3 * In addition to the above, for a machine learning model M, we also have the state Z p+3 In addition, for a machine learning model M, information about the representation R may also be provided. p and / or expression R p+1 * and / or expression R p+2 * But, and / or state Z p and / or State Z p+1 and / or State Z p+2 Information regarding the

[0133] In step (H), the machine learning model M creates a state Z based on the input data provided. p+4 The expression R expresses the inspection area in p+4 * Generate.

[0134] The generated representation R p+1 * , R p+2 * , R p+3 * , and / or R p+4 * may be output (e.g., displayed on a monitor and / or printed by a printer) and / or may be stored in a data memory and / or may be transmitted to a (separate) computer system.

[0135] The machine learning model generates a first representation that represents the inspection region in the first state Z1. T Starting from R1, the sequence of expressions T R2 * , ..., T R n * Here, each representation is T R j * is one state Z j represents the inspection region in, j is an index that runs from 2 to n, and for such a model, p A new expression R representing the inspection region in p The machine learning model can be fed into the expression R p+1 * , R p+2 * , ..., R n * where p is a number that can take on values ​​from 2 to n.

[0136] This means that the trained model does not necessarily need to be fed with a new representation R1 that represents the inspection region in the first state Z1. Furthermore, the trained machine learning model does not necessarily need to be fed with R p+1 * From R n * It is not necessary to generate all representations up to state Z1. Instead, n You can "drop in" and "drop out" anywhere in the sequence consisting of p Based on the result, one or more representations may be generated that represent the inspection area at one or more subsequent states.

[0137] It is even possible to generate representations that represent states that were never encountered in training. Thus, the representation R n * Instead of stopping at R n+1 * , R n+2 *Predictions may also be made for representations of states such as, etc. Thus, by using a trained machine learning model, the learned dynamics can be continued and representations for states that are never generated by measurements can be computed. In this regard, the trained machine learning model may be used to extrapolate to new states.

[0138] Moreover, prediction is not limited to a subsequent state. It is also possible to predict the representation of the inspection area at a previous state of a sequence of states. First, as already described herein, a machine learning model can essentially be trained in both directions, toward a subsequent state and toward a previous state. Second, the machine learning model performs a transformation on the input representation that is, in principle, reversible. By analyzing the mathematical function of the model that transforms the input representation into an output representation, it is possible to determine an inverse function that reverses the process and returns the previous output representation to the previous input representation. The inverse function may then be used to predict the representation of the previous state, even if the model was trained to predict the representation of the subsequent state, and vice versa.

[0139] FIG. 7 shows an extension of the scheme shown in FIG. 6. p The expression R expresses the inspection area in p Based on state Z p+1 , Z p+2 , Z p+3 , Z p+4 The expression R expresses the inspection area in p+1 * , R p+2 * , R p+3 * , R p+4 * 7 shows how state Z p The expression R expresses the inspection area in p Based on the expression R p+1 * ~R p+q *, where q is an integer equal to or greater than 2. The iterative nature of the machine learning model M is particularly evident in FIG. 7. The machine learning model M is a representation R p It is applied q times starting from x, and the output data is fed back to the machine learning model M((q-1)x) (q-1) times.

[0140] It should be noted that the machine learning model (trained or untrained) does not have to be applied to the entire radiological image (e.g., MRI image, or CT scan, or the like). The machine learning model can be applied to only a portion of the radiological image. For example, the radiological image can be first segmented to identify / select regions of interest. The model can then be applied, for example, to only the regions of interest.

[0141] Furthermore, the application of the machine learning model may include one or more pre-processing steps and / or one or more post-processing steps. For example, it is envisaged to first subject the received representation of the examination area to one or more transformations, such as motion compensation, color space conversion, normalization, segmentation, Fourier transformation (e.g., for conversion from an image space representation to a frequency space representation), inverse Fourier transformation (e.g., for conversion from a frequency space representation to an image space representation), and / or the like. In a further step, the transformed representation may be fed to a machine learning model, which then goes through a series of iterations (cycles) to generate, starting from the transformed representation, a series of further (subsequent) representations of the examination area in a series of further (subsequent) states (as shown diagrammatically in FIG. 7).

[0142] A machine learning model according to the present invention may, for example, be or include an artificial neural network.

[0143] An artificial neural network includes at least three layers of processing elements: a first layer having input neurons (input nodes), an Nth layer having at least one output neuron (output node), and N-2 hidden layers, where N is a natural number greater than or equal to 3.

[0144] The input neurons serve to receive the representation. For example, there may be one input neuron for each pixel or voxel of the representation when the representation is a real space representation in the form of a raster graphic, or one input neuron for each frequency present in the representation when the representation is a frequency space representation. Additional input neurons may be provided for additional input values ​​(e.g., information about the examination area, information about the examination subject, information about the state when the representation was generated, information about the state represented by the representation, and / or information about the time when the representation was generated or information about the period during which the representation was generated).

[0145] The output neuron may function to output the generated representation that represents the examination region at a subsequent state.

[0146] The processing elements in the layer between the input neurons and the output neurons are connected to each other in a predetermined pattern with predetermined connection weights.

[0147] Preferably, the artificial neural network is or comprises a so-called Convolutional Neural Network (abbreviated as CNN).

[0148] A CNN typically consists essentially of alternating filters (convolutional layers) and aggregation (pooling) layers, terminated with one or more layers of fully connected neurons (densely / fully connected layers).

[0149] The artificial neural network may be trained, for example, by backpropagation. The goal for the network is to predict as reliably as possible the dynamics of the test domain from a state, through at least one intermediate state, to a final state. The quality of the prediction is described by a loss function. The goal is to minimize the loss function. In backpropagation, the artificial neural network is taught by modifying the connection weights.

[0150] In the trained state, the connection weights between the processing elements contain information about the dynamics of state changes, and this information may be used to predict one or more representations representing the inspection area in one or more subsequent states based on a first representation representing the inspection area in a first state.

[0151] Cross-validation may be used to classify data into a training data set and a validation data set, which are used in backpropagation training of the network weights, and a validation data set to check the prediction accuracy of the trained network when applied to unknown data.

[0152] The artificial neural network may have an autoencoder architecture, for example, the artificial neural network may have an architecture such as U-Net (see, for example, 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).

[0153] The artificial neural network may be a generative adversarial network (GAN) (see, e.g., M.-Y. Liu et al.: Generative Adversarial Networks for Image and Video Synthesis: Algorithms and Applications, arXiv:2008.02793; J. Henry et al.: Pix2Pix GAN for Image-to-Image Translation, DOI: 10.13140 / RG.2.2.32286.66887).

[0154] The artificial neural network may be or may include a recurrent neural network. A recurrent or feedback neural network, as opposed to a feedforward network, refers to a neural network in which the neurons of a layer are distinguished by their connections with neurons of the same layer or with neurons of the previous layer. The artificial neural network may include, for example, a long short-term memory (LSTM) (see, for example, Y. Gao et al.: Fully convolutional structured LSTM networks for joint 4D medical image segmentation, DOI: 10.1109 / ISBI.2018.8363764).

[0155] The artificial neural network may be a transformer network (see, for example, D. Karimi et al.: Convolution-Free Medical Image Segmentation using Transformers, arXiv:2102.13645 [eess.IV]).

[0156] FIG. 8 illustrates, by way of example, a schematic diagram of a computer system according to the present disclosure.

[0157] A "computer system" is an electronic data processing system that processes data by means of programmable computational rules. Such systems typically include a "computer," which is a unit that includes a processor and peripheral devices for performing logical operations.

[0158] In computer technology, the term "peripheral" refers to any device that is connected to a computer and that is used to control the computer and / or is used as an input / output device. Examples of peripherals are monitors (screens), printers, scanners, mice, keyboards, drives, cameras, microphones, speakers, etc. Internal ports and expansion cards are also considered peripherals in computer technology.

[0159] The computer system (1) shown in FIG. 8 includes an input unit (10), a control calculation unit (20), and an output unit (30).

[0160] The control and calculation unit (20) serves to control the computer system (1) and to coordinate the data flow between the units of the computer system (1), and also to perform the calculations.

[0161] The control calculation unit (20) - via an input unit (10), a number of states Z1 to Z n One state Z of the sequence consisting of p An expression R representing the inspection area in the inspection object in p where n is an integer greater than or equal to 3 and p is an integer less than n; - Received Representation R p to a trained machine learning model, where the machine learning model is configured to provide a first representation based on the training data. T Starting from R1, n-1 representations TR2 * ~ T R n * where the first representation T R1 represents the inspection region in state Z1, and each generated representation T R j * is state Z j represents the inspection area in, j is an index that runs through the numbers 2 to n, and T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * is, at least in part, the generated representation T R k-1 * where k is an index going through the numbers 3 through n, - configured to receive one or more representations of the inspection region from the machine learning model, where each of the one or more representations represents a state Z p represents the inspection area in one state following - configured to output one or more representations received from the machine learning model to an output unit (30), and / or to store and / or transmit to another computer system.

[0162] FIG. 9 shows diagrammatically, by way of example, a further embodiment of a computer system according to the invention.

[0163] The computer system (1) includes a processing unit (21) connected to a memory (22). The processing unit (21) and the memory (22) form a control and computing unit, as shown in FIG.

[0164] The processing unit (21) may include one or more processors, either alone or in combination with one or more memories. The processing unit (21) may be a conventional computer hardware capable of processing information such as digital images (e.g., representations of an area of ​​examination), computer programs, and / or other digital information. The processing unit (21) typically consists of an arrangement of electronic circuits, some of which may be designed as an integrated circuit or as multiple integrated circuits connected together (integrated circuits are sometimes also referred to as "chips"). The processing unit (21) may be configured to execute computer programs, which may be stored in a working memory of the processing unit (21) or in a memory (22) of the same or a different computer system.

[0165] The memory 22 may be conventional computer hardware and may store information, such as digital images (e.g., representations of an area under examination), data, computer programs, and / or other digital information, temporarily and / or permanently. The memory 22 may include volatile and / or non-volatile memory and may be non-removable or removable. Examples of suitable memory are RAM (random access memory), ROM (read only memory), hard disks, flash memory, interchangeable computer floppy disks, optical disks, magnetic tapes, or combinations thereof. Optical disks may include compact disks with read-only memory (CD-ROM), compact disks with read / write capability (CD-R / W), DVDs, Blu-ray disks, and the like.

[0166] The processing unit (21) may be connected to the memory (22) as well as to one or more interfaces (11, 12, 31, 32, 33) for displaying, transmitting, and receiving information. The interfaces may include one or more communication interfaces (32, 33) and / or one or more user interfaces (11, 12, 31). The one or more communication interfaces (32, 33) may be configured to transmit and / or receive information, for example, to an MRI scanner, a CT scanner, an ultrasound camera, another computer system, a network, a data memory, or the like. The one or more communication interfaces (32, 33) may be configured to transmit and / or receive information via a physical (wired) communication connection and / or a wireless communication connection. The one or more communication interfaces (32, 33) may include one or more interfaces for connecting to a network using technologies such as, for example, cellular, Wi-Fi, satellite, cable, DSL, fiber optic, and / or the like. In some examples, the one or more communication interfaces (32, 33) may include one or more short-range communication interfaces configured to connect devices having short-range communication technologies, such as NFC, RFID, Bluetooth, Bluetooth LE, ZigBee, infrared (e.g., IrDA), or the like.

[0167] The user interface (11, 12, 31) may include a display (31). The display (31) may be configured to display information to a user. Suitable examples of the display are a liquid crystal display (LCD), a light emitting diode display (LED), a plasma display panel (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, for example, for processing, storage, and / or display within the computer system (1). Suitable examples of the user input interface are a microphone, an image or video recording device (e.g., a camera), a keyboard or keypad, a joystick, a touch-sensitive surface (separate from or integrated with a touch screen), or the like. In some examples, the user interface (11, 12, 31) may contain automatic identification data capture technology (AIDC) for machine-readable information. This may include bar codes, radio frequency identification (RFID), magnetic strips, optical character recognition (OCR), integrated circuit cards (ICC), and the like. The user interface (11, 12, 31) may further include one or more interfaces for communicating with peripheral devices such as printers and the like.

[0168] One or more computer programs (40) may be stored in memory (22) and such computer programs may be executed by processing unit (21), where processing unit (21) is programmed by such computer programs to perform the functions described herein. The retrieval, loading, and execution of instructions of computer program (40) may be performed sequentially, such that instructions are retrieved, loaded, and executed, respectively. However, the retrieval, loading, and / or execution may also be performed in parallel.

[0169] The machine learning models of the present disclosure may also be stored in memory (22).

[0170] The computer system of the present disclosure may be designed as a laptop, notebook, netbook, and / or tablet PC, and may be a component of an MRI scanner, a CT scanner, or an ultrasound diagnostic device.

[0171] FIG. 10 illustrates generally in the form of a flow chart one embodiment of a method for training a machine learning model.

[0172] The method (100) comprises: (110) receiving training data; The training data is a representation of the inspection area for a large number of inspection objects. T R1~ T R n where n is an integer equal to or greater than 3; ·Each expression T R i is a set of multiple states Z1 to Z n One state Z of the sequence consisting of i where i is an index that runs through the integers 1 to n; (120) Training a machine learning model, Model represents T Starting from R1, the expression T R2 * ~ T R n * are trained to generate a sequence of ·Expression T R2 * is, at least in part, represented T Each further representation is generated based on R1. T R k * is, at least in part, represented T R k-1 * where k is an index going through the numbers 3 through n, ·Expression T R k * is state Z k represents the inspection area in T R k-1 * is state Z k-1 represents the inspection area in state Z k-1 is state Z k It is the direct precursor to Generated Representation T R j * and T R j The difference between is quantified by a loss function, where j is an index that runs from 2 to n, The difference is minimized by changing the model parameters of the machine learning model; and (130) Storing the trained machine learning model and / or utilizing the machine learning model for prediction.

[0173] FIG. 11 illustrates generally in flow chart form one embodiment of a computer-implemented method for generating a representation of a state.

[0174] The method (200) comprises: (210) Expression R for the inspection region in the inspection object p The purpose of the present invention is to receive ·Expression R p is a set of multiple states Z1 to Z n One state Z of the sequence consisting of p where p is an integer smaller than n, and n is an integer equal to or greater than 3; (220) Expression R p to a trained machine learning model, The machine learning model is trained according to the method (100); and (230) From the machine learning model, one or more representations R p+q *The purpose of the present invention is to receive One or more expressions R p+q * Each of these has one state Z p+q represents the inspection area in · q is an index that runs from 1 to m, m is an integer less than, equal to, or greater than n-1; and (240) One or more expressions R p+q * and outputting and / or storing and / or transmitting the

Claims

1. 1. A computer-implemented method for training a machine learning model, comprising: receiving training data, The training data includes a plurality of representations of test areas in a large number of test objects. Ｔ R １ ~ Ｔ R ｎ wherein n is an integer of 3 or greater; ・Each expression Ｔ R ｉ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｉ represents the inspection area in, i is an index running through the integers 1 to n, Each test subject is preferably a human being, and the test area is preferably a part of the human being. Receiving and training said machine learning model, The model represents Ｔ R １ Starting from the plurality of test objects, for each test object in the plurality of test objects, Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ are trained to generate a series of ・Expression Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running from 3 to n, ・Expression Ｔ R ｋ ＊ is state Z ｋ and represents the inspection area in Ｔ R ｋ－１ ＊ is state Z ｋ－１ represents the inspection area at state Z ｋ－１ is state Z ｋ is the direct precursor to To train and - storing the trained machine learning model and / or using the machine learning model for prediction; 10. A computer-implemented method comprising:

2. - each generated representation Ｔ R ｊ ＊ Regarding the received representation Ｔ R ｊ and the generated expression Ｔ R ｊ ＊ and calculating a loss value using a loss function for pairs consisting of - calculating a total loss value using a total loss function, said total loss function being a function of said loss value; - minimizing the total loss value by modifying the parameters of the machine learning model; The method of claim 1 further comprising:

3. The total loss function is expressed by the following formula: [Equation 1] and LV is the total loss value, LF ｊ is the received representation Ｔ R ｊ and the generated expression Ｔ R ｊ ＊ is a loss function for calculating the loss value for the difference between ｊ is a weighting coefficient, The method of claim 2.

4. 1. A computer-implemented method for predicting one or more representations for an examination region in an examination object, comprising: - a representation R relating to the examination area ｐ and ・Expression R ｐ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｐ where p is an integer smaller than n, and n is an integer greater than or equal to 3; The test subject is preferably a human being, and the test area is preferably a part of the human being. Receiving and -Expression R ｐ to a trained machine learning model, The trained machine learning model generates a first representation based on training data. Ｔ R １ Starting from Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ It is trained to generate a series of ・First expression Ｔ R １ is the first state Z １ and each generated representation represents the inspection area in Ｔ R ｊ ＊ is state Z ｊ where j is an index running from 2 to n, ・Expression Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running from 3 to n, the training data are results of radiological examinations; To supply and - one or more representations R for the examination region from the machine learning model ｐ＋ｑ ＊ and said one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; Receiving and said one or more representations R ｐ＋ｑ ＊ outputting and / or storing and / or transmitting the 10. A computer-implemented method comprising:

5. The method comprises: - a representation R relating to the examination area ｐ and receiving the expression R ｐ is the plurality of states Z １ ~Z ｎ A state Z in the sequence consisting of ｐ where p is an integer less than n; -Expression R ｐ to the trained machine learning model, wherein the trained machine learning model is trained by the method of claim 1; and - from the trained machine learning model, the one or more representations R for the examination region ｐ＋ｑ ＊ and receiving one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ ＊ where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; said one or more representations R ｐ＋ｑ ＊ outputting and / or storing and / or transmitting the The method of claim 4, comprising:

6. Each expression R １＋ｑ ＊ is expressed as follows: M ｑ (R １ )=R １＋ｑ ＊ is calculated according to M represents a transformation applied by the machine learning model to the input data of the machine learning model, and M ｑ means that the transformation is applied q times, and in a first step, the transformation is applied to a first representation R １ and in a second step, the transformation is applied to the result of the first step, which is then fed back into the machine learning model, and the procedure is repeated for each corresponding further transformation result until the transformation has been applied a total of q times, where q is an integer that can range from 1 to m, and m is an integer less than, equal to, or greater than n-1. The method of claim 1.

7. The sequence of states comprises: a first state of the examination region at a first time point before administration of a contrast agent, a second state of the examination region at a second time point after administration of the contrast agent, a third state of the examination region at a third time point after administration of the contrast agent, a fourth state of the examination region at a fourth time point after administration of the contrast agent, a fifth state of the examination region at a fifth time point after administration of the contrast agent, and one or more of the following conditions: The method of claim 1 , wherein the first time point, the second time point, the third time point, the fourth time point, and the fifth time point form a time series.

8. The method of claim 1 , wherein the examination area is a liver of the human or a portion of the liver of the human.

9. The sequence of states comprises: the liver or a part of the liver before administration of a hepatobiliary imaging agent, the liver or a part of the liver during the arterial phase after administration of the hepatobiliary imaging agent, the liver or a part of the liver during the portal phase after administration of the hepatobiliary imaging agent, the liver or a part of the liver during the transition phase after administration of the hepatobiliary imaging agent, the liver or a part of the liver during the hepatic imaging phase after administration of the hepatobiliary imaging agent, The method of claim 8 , comprising one or more of the following conditions:

10. Each representation R generated by the machine learning model ｐ＋ｑ ＊ and / or Ｔ R ｐ＋ｑ ＊ is the expression R ｐ＋ｑ＋１ ＊ and / or Ｔ R ｐ＋ｑ＋１ ＊ The method of claim 1 , wherein the metric is fed back to the trained machine learning model to generate:

11. 5. The method of claim 4, wherein m lies in the range from n to n+2.

12. The method of claim 1 , wherein the received representation is a CT image, an MRI image, or an ultrasound image.

13. 1. A computer system comprising: an input unit; a control calculation unit; an output unit; Including, The control calculation unit - the input unit is provided with a representation R ｐ and configured to receive a representation R ｐ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｐ where p is an integer smaller than n, and n is an integer equal to or greater than 3, the test subject is preferably a human, and the test area is preferably a part of the human; -Expression R ｐ to a trained machine learning model, the machine learning model being configured to generate a first representation based on the training data. Ｔ R １ Starting from Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ The first representation is trained to generate Ｔ R １ is the first state Z １ and each generated representation represents the inspection area in Ｔ R ｊ ＊ is state Z ｊ where j is an index going through the numbers 2 to n, and Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running from 3 to n, - one or more representations R for the examination region from the machine learning model ｐ＋ｑ ＊ and configured to receive one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ where q is an index running through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; said one or more representations R ｐ＋ｑ ＊ to the output unit and / or to store and / or transmit to another computer system. Computer system.

14. 1. A computer program product comprising a computer program that can be loaded into the working memory of a computer system, The computer program is configured to: -Representation R for the inspection area in the inspection object ｐ receiving a representation R ｐ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｐ where p is an integer smaller than n, and n is an integer equal to or greater than 3, the test subject is preferably a human, and the test area is preferably a part of the human; -Expression R ｐ to a trained machine learning model, wherein the machine learning model generates a first representation based on the training data. Ｔ R １ Starting from Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ The first representation is trained to generate Ｔ R １ is the first state Z １ and each generated representation represents the inspection area in Ｔ R ｊ ＊ is state Z ｊ where j is an index going through the numbers 2 to n, and Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running through the numbers 3 to n; - one or more representations R for the examination region from the machine learning model ｐ＋ｑ ＊ receiving one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; said one or more representations R ｐ＋ｑ ＊ outputting and / or storing and / or transmitting the A computer program product that causes the execution of

15. 1. Use of a contrast agent in a radiological examination, comprising: The radiological examination method includes: -Representation R for the inspection area in the inspection object ｐ and receiving the representation R ｐ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｐ representing the examination region before or after administration of the contrast agent, where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R ｐ to a trained machine learning model, which generates a first representation based on the training data. Ｔ R １ Starting from Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ The first representation is trained to generate Ｔ R １ is the first state Z １ and each generated representation represents the inspection area in Ｔ R ｊ ＊ is state Z ｊ where j is an index going through the numbers 2 to n, and Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running from 3 to n; - one or more representations R for the examination region from the machine learning model ｐ＋ｑ ＊ and receiving one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; said one or more representations R ｐ＋ｑ ＊ outputting and / or storing and / or transmitting the Use of contrast agents including

16. 1. A contrast agent for use in radiological examinations, comprising: The radiological examination method includes: -Representation R for the inspection area in the inspection object ｐ and generating a representation R ｐ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｐ representing the examination region before or after administration of the contrast agent, where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R ｐ to a trained machine learning model, which generates a first representation based on the training data. Ｔ R １ Starting from Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ The first representation is trained to generate Ｔ R １ is the first state Z １ and each generated representation represents the inspection area in Ｔ R ｊ ＊ is state Z ｊ where j is an index going through the numbers 2 to n, and Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running from 3 to n; - one or more representations R for the examination region from the machine learning model ｐ＋ｑ ＊ and receiving one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; said one or more representations R ｐ＋ｑ ＊ outputting and / or storing and / or transmitting the A contrast agent comprising:

17. A kit comprising an imaging agent and a computer program product, The computer program product includes a computer program that can be loaded into a working memory of a computer system, the computer program causing the computer system to: -Representation R for the inspection area in the inspection object ｐ receiving a representation R ｐ is a set of multiple states Z １ ~Z ｎ One state Z of the sequence ｐ representing the examination region before or after administration of the contrast agent, where p is an integer less than n, and n is an integer greater than or equal to 3; -Expression R ｐ to a trained machine learning model, wherein the machine learning model generates a first representation based on the training data. Ｔ R １ Starting from Ｔ R ２ ＊ ~ Ｔ R ｎ ＊ The first representation is trained to generate Ｔ R １ is the first state Z １ and each generated representation represents the inspection area in Ｔ R ｊ ＊ is state Z ｊ where j is an index going through the numbers 2 to n, and Ｔ R ２ ＊ is, at least in part, the representation Ｔ R １ Each further representation is generated based on Ｔ R ｋ ＊ are, at least in part, the representations previously generated by Ｔ R ｋ－１ ＊ where k is an index running through the numbers 3 to n; one or more representations R of said examination area ｐ＋ｑ ＊ receiving one or more representations R ｐ＋ｑ ＊ Each of these is in state Z ｐ＋ｑ ＊ where q is an index going through the numbers 1 to m, and m is an integer less than, equal to, or greater than n-1; said one or more representations R ｐ＋ｑ ＊ outputting and / or storing and / or transmitting the A kit that allows you to perform the following.

Images