Flow state normalization for functional CTA

Abstract

A system and related method for medical image processing, comprising an input interface (IN) through which input data can be received. The input data includes a spectral input image of a portion of the vasculature (VS) of the patient (PAT). The spectral input image is based on data that can be reconstructed from projection data acquired by a tomography-type spectral imaging device, where a contrast agent is present in a portion of the vasculature. A predictor module (PM) of the system is configured to predict an increase or decrease in flow caused by the presence of a contrast agent based on the input image. An output interface (u-OUT) provides output data indicating a predicted increase or decrease in traffic.

Description

Technical Field

[0001] The invention relates to a system, an imaging arrangement, an associated method, a computer program element and a computer readable medium for medical image processing. Background Art

[0002] Coronary CTA (computed tomography angiography) imaging has gained importance as a noninvasive first-line diagnostic tool for suspected coronary artery disease.

[0003] Specifically, CT perfusion or CT-derived IMR assessment has attracted recent attention, as these methods can support the diagnosis of microvascular dysfunction and thus address long-standing questions.

[0004] An obstacle to these noninvasive methods is the effect of intravenous contrast injection on coronary blood flow, which is however required to impart adequate image contrast. Intracoronary contrast administration can induce a state of full hyperemia with typical contrast doses. Under intravenous administration, the response of coronary autoregulation is between the normal resting state and the state of full hyperemia. The widening of the vessels is believed to be an autoregulatory response caused by the interfering presence of contrast material in the vessels.

[0005] Therefore, any derived diagnostic measurements / quantities related to blood flow in the coronary arteries are affected by this response of autoregulation. Since contrast agent dose and distribution vary from patient to patient, the autoregulatory response is different, and it is therefore difficult to directly compare patient measurements or derive global thresholds across a patient population. In particular, CTA-derived hemodynamic measurements related to blood flow in the coronary arteries (e.g., perfusion, CFR, IMR) are subject to inter- and intra-patient variability because they are affected by the response of coronary autoregulation to contrast agents. Therefore, the compatibility of these measurements is limited. Summary of the invention

[0006] Therefore, there may be a need for improved image-based measurements of flow-related diagnostic quantities, particularly hemodynamic quantities.

[0007] The objects of the invention are achieved by the subject-matter of the independent claims, wherein further embodiments are included in the dependent claims.It should be noted that the aspects of the invention described below apply equally to the associated method, imaging arrangement, computer program element and computer readable medium.

[0008] According to a first aspect of the present invention, there is provided a system for medical image processing, comprising:

[0009] an input interface through which input data comprising a spectral input image of at least a portion of a patient's vascular system can be received when the system is in use, the spectral input image being based on data reconstructible from projection data acquired by a spectral imaging device of the tomographic type, wherein a contrast agent is present in said portion of the vascular system;

[0010] a predictor module configured to predict, based on input images, an increase or decrease in flow caused by the presence of a contrast agent; and

[0011] An output interface for providing output data indicative of the predicted increase or decrease in flow.

[0012] In an embodiment, the system comprises an intra-image measurement component configured to measure a medical volume of interest based on the input image and said output data.

[0013] In an embodiment, an intra-image measurement component is used to combine information from the input image with information based on said output data.

[0014] In an embodiment, the combining by the intra-image measuring component comprises correcting or normalizing, based on the output data, initial measurements of the medical volume of interest made by the intra-image evaluating component based on the input image.

[0015] In an embodiment, the medical quantity of interest comprises a hemodynamic quantity.

[0016] In an embodiment, the hemodynamic quantity comprises any one or more of the following: perfusion fraction, [virtual] fractional flow reserve FFR, [virtual] instantaneous flow ratio iFR, coronary flow reserve CFR, microvascular resistance index IMR and intraluminal attenuation gradient TAG.

[0017] In an embodiment, the input data also includes one or more vital sign measurements about the patient.

[0018] In an embodiment, the system includes a renderer configured to provide a sensory indication for use by a user based on the output data.

[0019] In an embodiment, the renderer comprises any one or more of: i) a visualizer configured to cause display of the output data on a display device, ii) a transducer configured to generate an alarm signal related to the output data.

[0020] In an embodiment, the system comprises a control interface configured to issue a control signal to control the medical equipment based on the output data.

[0021] In an embodiment, the input image is based on a partial reconstruction from projection data.

[0022] In an embodiment, the input image and / or the predicted flow increase or decrease has time and / or location dependency.

[0023] In an embodiment, the increase or decrease in flow, in particular blood flow in at least a portion of the vasculature, is caused by an autoregulatory response of the vasculature, which is caused by the presence of the contrast agent.

[0024] In an embodiment, the predictor component is configured to correlate a concentration of contrast agent in the portion of the vasculature or in the surrounding tissue to said increase or decrease in contrast agent flow.

[0025] In an embodiment, the predictor module is based on a trained machine learning model.

[0026] In an embodiment, the spectral imaging device is configured for computed tomography angiography.

[0027] In another aspect, there is provided an arrangement comprising a system according to any one of the preceding claims and an imaging device.

[0028] In another aspect, a method for medical image processing is provided, comprising:

[0029] receiving input data comprising a spectral input image of at least a portion of a vasculature of a patient, the spectral input image being based on data reconstructible from projection data acquired by a spectral imaging device of the tomographic type, wherein a contrast agent is present in the portion of the vasculature;

[0030] estimating an increase or decrease in flow caused by the presence of a contrast agent based on the input image; and

[0031] Output data is provided indicating a predicted increase or decrease in flow.

[0032] In another aspect, a method for training a machine learning model based on training data is provided.

[0033] In another aspect, a computer program element is provided which, when executed by at least one processing unit, is adapted to cause the processing unit to perform any one of the methods.

[0034] On the other hand, at least one computer-readable medium is provided, on which the program unit is stored, or on which the machine learning model is stored.

[0035] In yet another aspect, the use of predicted flow increases in calculating diagnostic quantities, such as hemodynamic quantities, is provided. This allows for improved comparability / standardization across patients.

[0036] What is proposed in the embodiments herein is to use spectral 3D imaging (such as CTA) to estimate the response of coronary artery autoregulation due to the presence of contrast agent itself based on this (implicitly or explicitly). Therefore, the presence of contrast agent alone in the vasculature may cause (blood) flow changes due to the autoregulatory response of the vascular system in the presence of contrast agent. Observation of contrast agent concentration can be used for this, as captured in CTA images. The response is mainly one of increased blood flow. Based on the predicted increase in blood flow, the impact on diagnostic measurements can be considered, such as by normalization, correction or other. In an embodiment, the estimated blood flow increase can be monitored locally or globally for a given one or more regions of interest (ROI). Therefore, the impact of coronary artery autoregulation / flow increase on diagnostic measurements can be estimated. The estimation of this impact can be done in a time and / or vascular position resolved manner. Due to the proposed system and method, flow-based or flow-related diagnostic measurements can be performed more accurately and reliably, and can be compared across patient populations. In some alternative embodiments, flow reduction is estimated for some categories / types of contrast agents instead. However, the proposed method was mainly conceived to estimate flow increase, as this effect seems to be more prevalent in patients.

[0037] The predictor module or prediction step may be based at least in part on analytical modeling, wherein the amount of contrast agent is detected due to the excellent accuracy of spectral imaging and (directly) mapped to a corresponding amount of flow increase. However, such (explicit) extraction of the amount of contrast agent is not required in some machine learning based applications that can be implemented end-to-end, wherein the input image data is mapped to flow increases. Hybrid types of implementations using ML and analytical methods are also envisioned herein.

[0038] The proposed imaging can be considered functional in that it is the functional effects of a condition (such as coronary artery disease) that are studied, rather than just the geometrical / anatomical effects. For example, rather than just measuring lesion diameter or diameter reduction, how the lesion affects blood supply function or other physiological aspects is measured or simulated.

[0039] Although the above content focuses on the inspection of the heart or other blood vessels for stenosis or other lesions, other medical applications are not excluded herein, such as the inspection of other vascular systems, such as lymphatic, urethral or gastric tracts. Non-medical applications of the principles described herein are also not excluded herein, such as the inspection of inaccessible pipes or hydraulic systems, flow / geology / hydrology inspection of water flow and seepage, and other flow-related endeavors.

[0040] A "user" refers to a person who operates the imaging device or oversees the imaging process, such as a medical staff member or other person. In other words, the user is typically not the patient.

[0041] Typically, the term "machine learning" includes a computerized arrangement (or module) that implements a machine learning ("ML") algorithm. Some such ML algorithms operate to adjust a machine learning model that is configured to perform ("learn") a task. Other MLs operate directly on training data without having to use a model such as a model. This adjustment or update of the training data corpus is called "training". Typically, the task execution of an ML module can be measurably improved using training experience. The training experience can include exposure of suitable training data and models to such data. The better the data represents the task to be learned, the better the task performance can be improved. "If the training data well represents the distribution of examples on which the final system performance is measured, the training experience helps to improve performance". Performance can be measured by objective testing based on the output generated by the module in response to feeding the module with test data. Performance can be defined based on a specific error rate to be achieved for given test data. See, for example, "Machine Learning" by TM Mitchell (McGraw-Hill, 1997, Chapter 1.1, page 2, Chapter 1.2.1, page 6. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Exemplary embodiments of the present invention will now be described with reference to the following drawings, which are not to scale unless otherwise specified, and in which:

[0043] Figure 1 An overview block diagram showing a medical imaging arrangement capable of obtaining spectral image data;

[0044] Figure 2 A schematic block diagram of a medical arrangement according to one embodiment is shown in more detail;

[0045] Figure 3 A block diagram of a computerized system for calculating flow increase caused by the presence of a contrast agent in an imaging region based on reconstructed spectral image data is shown;

[0046] Figure 4 According to one embodiment, and as can be seen in Figures 1 to 3 A block diagram of a machine learning model used in a system of any one of the items;

[0047] Figure 5 A block diagram of a training system for training a machine learning model according to one embodiment is shown;

[0048] Figure 6 A flow chart illustrating a method of calculating the flow increase caused by the presence of a contrast agent in an imaging region to facilitate calculation of the associated one or more medical quantities of interest; and

[0049] Figure 7 A flow chart showing a method for training a machine learning model. DETAILED DESCRIPTION

[0050] refer to Figure 1 Schematic overview block diagram of a computer-implemented system SYS for supporting contrast-assisted tomographic imaging operations is presented herein. A region of interest ("ROI") may include a vessel VS, such as a blood vessel through which blood passes. In particular, the vessel VS may specifically include all or part of the cardiac vasculature of a human patient. The proposed system allows processing of volumetric contrast image data V=R(λ) provided by a tomographic imaging device IA, which is operable to acquire a projection image λ of the region of interest (ROI), from which a volumetric image V can be reconstructed.

[0051] Preferably, the X-ray based imaging apparatus IA is configured for spectral imaging such that spectral volume image data V=R(λ) are provided for processing by the system SYS. The spectral volume image data V=R(λ) represent a plurality of cross-sectional (tomographic) images along the imaging axis, wherein due to the spectral processing the image contrast is specifically adjusted to the contrast agent previously applied to the region of interest. A plurality of volumes V=V i∈I (quasi "4D" data), where i is an index in a suitable interval, such as a different moment in a time period. Some volumes in the series can be based on a full scan, others can be based on a partial scan (a scan refers to an acquired set of projection data). Some angular ranges can overlap, others can be disjoint. As an alternative or in addition to time, the index i can relate to space, as will be described in more detail below. The system can also take as input multiple such volumes for multiple ROIs or a single such volume with multiple ROIs.

[0052] As will be described in more detail below, the system SYS is configured to estimate the physiological response of the vasculature due to the presence of a contrast agent in a vessel. In particular, and as primarily contemplated herein, as will be described in more detail below, the system SYS is configured to estimate the blood flow increase Δu due to the presence of a contrast agent in a vessel. It has been found that such undesirable effects of the flow increase adversely interfere with flow-related quantities, such as hemodynamic quantities, which may be calculated from spectral images of the presence of contrast agent, from which other diagnostic quantities may be derived. Similarly, certain contrast agents may result in a decrease in blood flow. However, the following will focus primarily on the flow increase estimation primarily contemplated herein.

[0053] Flow related quantities may describe the flow properties of the blood / contrast agent mixture. For example, such flow characterization quantities allow building a model m(VS) of the vascular system VS of interest, including modeling of the flow resistance Oj. The image measurement component IMC may process the images and flow related quantities in order to calculate a hemodynamic quantity d therefrom, wherein the estimated flow increase Δu is taken into account, for example, by compensation, offset, normalization or other computational operations. The proposed system facilitates obtaining flow related quantities, such as the hemodynamic quantity d, of higher fidelity, better comparability across patients and better accuracy. The proposed system SYS facilitates better diagnostic results. The hemodynamic value d may be processed by the diagnostic system DS, for example to calculate derived quantities, such as the fraction of condition presence, etc.

[0054] Before explaining the operation of the flow increase estimator / predictor system SYS in more detail, reference will now first be made to Figure 2 The block diagram describes the components of the entire imaging arrangement MIA.

[0055] The medical imaging arrangement MIA may comprise an imaging device IA (also referred to herein as "imager"), preferably based on X-rays. The imaging device IA may be of a rotating type. Such a rotating imaging device allows the collection of projection images λ along different projection directions P(α) in a "scan" around a region of interest (such as around a stenosis mentioned in a coronary vessel). The imaging device IA comprises an X-ray source XS and an X-ray sensitive detector XD. In some embodiments, the X-ray source XS and the X-ray sensitive detector XD opposite thereto are preferably arranged on a rotatable gantry G. The gantry G may have a recess in which the examination region is defined during imaging, in which the patient or in particular the region of interest is located. Thus, the rotatable gantry may have a "C", "U" or similar shape, such as in a C-arm imaging device as envisaged herein. A dual-plane imaging setup with two detectors having intersecting (e.g., at right angles) imaging axes is not excluded herein, but is less preferred. Instead of using a C-arm / U-arm system, in some embodiments, CT scanners of any generation, but preferably at least the 3rd generation, are also contemplated, in which the x-ray source XS and the detector D are mounted so that they rotate together around the patient. However, for multi-directional projection data acquisition, higher generation CT settings are also contemplated, in which there is not necessarily mechanical rotation of the source and the detector. Instead, a fixed ring of detectors around the ROI and / or a fixed plurality of X-sources arranged in a ring around the ROI are used. Cone beam imaging or other divergent beam geometries are also contemplated herein, and therefore imaging is performed using a spiral scanning path caused by a circular path of the x-ray source, for example using a translational motion relative to the imaging device IA and the ROI along the longitudinal axis of the patient. In most cases, this translational motion is caused by a patient bed on which the patient lies, which is translated during projection data acquisition. However, scanning paths having geometries other than circular / spiral are not excluded herein.

[0056] Furthermore, while digital imaging at the source (such as a flat panel detector) is preferred (where images are acquired locally as digital data), such an arrangement is not necessarily required. More traditional, cartridge-operable analog imaging methods (based on X-ray film) are not excluded herein. The analog image can be digitized after acquisition, for example, by using a digital camera. Imaging based on X-ray image intensifiers, etc., with post-digitization is also contemplated in some embodiments.

[0057] Typically, imaging includes energizing an X-ray source XS so that an X-ray beam XB is emitted from a focal point of the source XS, passes through an examination region (wherein there is a region of interest) and interacts with patient tissue material. The interaction of the X-ray beam XB with the tissue material causes the beam to be modified. The modified beam can then be detected at a detector XD. A conversion circuit (not shown) converts the detected intensity into a digital image, in particular into a projected image. The image can be used by a user in a medical procedure. The patient can reside on a patient support PS (such as a table) during imaging. Soft tissue, such as tissue constituting coronary vessels, typically provides poor contrast in native X-ray images.

[0058] To improve contrast, a contrast agent (sometimes referred to as "CA") (such as iodine solution, gadolinium or other) is delivered to the patient's bloodstream through the entry point AP before or during imaging. The contrast agent can be delivered via a CA delivery device DA (such as an electric pump), but manual delivery is not excluded herein. The delivery device DA can be operated to deliver a defined volume of contrast agent to the patient. The volume of contrast agent so delivered travels with the bloodstream. After a period of time, the contrast agent accumulates in or around an area of ​​interest, such as a stenosis. Image acquisition is preferably synchronized with the arrival of the contrast agent at the ROI. Preferably, a "contrast" image is acquired while the contrast agent is present at the ROI. Optionally, an additional "non-contrast" image is acquired while the contrast agent is not present at the ROI.

[0059] Projection images, in particular contrast, can be acquired in a time series, wherein a varying concentration of a contrast agent is accumulated at a region of interest. Image acquisition can begin before the contrast agent is accumulated at the ROI, can continue throughout the accumulation of the contrast agent at the ROI, and can be terminated after the accumulation of the contrast agent at the ROI. Typically, the contrast agent concentration increases over time after delivery, will stabilize at a certain maximum value, and will then wash out or decrease over time until it disappears completely. Although a single or certain discrete number of still images can be acquired in some cases, it is preferred to acquire streaming images or frames at a suitable frame rate because this allows real-time dynamic observation. Specifically, the acquired projection image stream can be visualized by a visualizer VIZ as a live video feed on a display device DD. Alternatively or additionally, the static images are so visualized. The operation of the proposed system SYS is based on processing such still or streaming images. The images can be stored in an image memory MEM, or can be processed in other ways.

[0060] X-ray imaging is typically based on the attenuation coefficient of the (one or more) materials that make up the region of interest. The attenuation (coefficient) depends on the energy of the X-ray beam. In order to exploit this energy dependency, a spectral imager IA configured for spectral imaging can be used. This can include detector-side or source-side solutions. Detector-side solutions include, for example, photon counting detector hardware or dual energy settings with dual-layer hardware. More than two detector layers can be used, but this is less common. Source-side solutions include multi-source settings, kVp switching, filtering, etc.

[0061] The imaging setup in spectral imaging allows the acquisition of multi-dimensional image data. Such multi-dimensional image data includes projection images at different energy levels, rather than acquiring energy-integrated images along different projection directions α provided by conventional (non-spectral) X-ray imaging setups.

[0062] This spectral CT imaging arrangement includes a spectral image processor SP. The spectral image processor SP can implement spectral image processing algorithms, such as material decomposition or others. A series of such spectral image processing algorithms are contemplated herein. The spectral processor SP processes the multi-energy projection image λ acquired by the source or detector side imaging arrangement just described, and thereby calculates a spectral image different from the original input multi-energy projection image. Therefore, the spectral projection image is computationally derived from the input multi-energy projection image.

[0063] Such spectral images, which may be in the projection domain or the image domain, may specifically include contrast-only images ("VC") as primarily contemplated herein. Such VC images represent an approximation of an image that would be obtained if only contrast agent were present in the field of view of the imager IA, excluding other materials. Thus, image contrast in the VC is exclusively or predominantly focused on the contrast agent.

[0064] The tomographic reconstructor RECON processes the projection data into a tomographic / volume image R(λ)=V, as will be described in more detail below. The volume data V are spectral image data due to a spectral processor SP. The spectral processor SP can operate in the projection domain on the acquired projection data λ, or can operate in the image domain after reconstruction. For simplicity, the symbols "V" or "R(λ)" may be used herein to indicate a spectral volume image in the image domain, regardless of the domain in which the spectral processing SP takes place. It is such spectral volume data V that are processed by the proposed system SYS. The spectral processor SP may be integrated into the reconstructor RECON.

[0065] The spectrum processor SP and the proposed system SYS may be implemented on different computing systems, possibly geographically separated, such as in a distributed "cloud" architecture. However, implementation on a single computing system is also envisaged herein. The imaging envisaged herein is preferably attenuation-based, but this does not exclude other X-ray modalities, such as dark field imaging and / or phase contrast imaging.

[0066] The reconstructor RECON is configured to reconstruct a volumetric image V=R(λ) from acquired projection data λ acquired when at least some contrast agent was present in the field of view of the imager IA and thus in a region of interest including at least a portion of the vasculature of interest.

[0067] It is mainly envisaged herein that the system SYS operates on spectral image volume data R(λ). However, it is not excluded herein that some embodiments of the system SYS can operate in the projection domain itself, but this is not preferred. The projection domain is the space located at the detector that detects the projection data. The tomographic reconstruction implemented by the reconstructor RECON is a transformation operation R according to which the volume data is calculated. The reconstruction transformation operation R is a transformation from the projection domain to the image domain where the volume data is located. The image domain is the part of the space defined between the x-ray source XD and the detector D, in which the region of interest (through which the contrast agent passes) resides during imaging. Therefore, the region of interest specifically includes the part of the vascular system VS to be imaged, in which contrast agent is sometimes present. Conceptually, the image domain can be considered to consist of a 3D grid of spatial points ("voxels"). The reconstruction operation R consists of calculating the image value at each grid point (voxel) to obtain the volume data R(λ).

[0068] Reference now Figure 3 , which provides more details of the components of a system SYS configured to calculate the flow increase due to the presence of contrast agent.

[0069] At one or more input ports / interfaces IN, input data comprising spectral volume data V=R(λ) is received. As described above, the output data may comprise a single volume or a series of volumes acquired over a time period I.

[0070] The predictor / estimator module PM calculates an estimate Δu of the flow increase caused by the presence of a contrast agent based on the input data. For example, the blood flow increase Δu may be expressed as a percentage value (p%), or by any other suitable quantification. The output data u-OUT provided by the predictor / estimator module PM may include the estimate Δu, which may be output at an output port OUT for further processing, storage for later viewing / processing, etc. Such processing may include the operation of a visualizer VIZ for displaying the output data (or data derived from the output data) on a display device DD. In particular, such data processing may include processing the output data by an intra-image measurement component IMC.

[0071] The intra-image measurement component IMC analyses voxel data in the volume image V, such as grayscale values, in particular its contrast distribution due to the contrast agent, etc., and calculates the mentioned flow-related quantities therefrom. The hemodynamic quantities are preferably configured to describe the blood flow or blood-contrast agent mixture flow through the vascular system of interest VS.

[0072] From such flow related quantities, other suitable medical quantities of interest may be derived, such as other diagnostic quantities, such as a score for the presence of a disease. The later score or other diagnostic quantity may be calculated by the diagnostic system DS.

[0073] The hemodynamic quantities or other flow-related quantities that can be calculated by the intra-image measurement component IMC may include one or more of the following: FFR, IFR, CFR, IR, TAG (intraluminal attenuation gradient), perfusion fraction and / or others. TAG can be determined using the average contrast opacity (e.g., measured in Hounsfield units) in a series of regions of interest along the vessel of interest. For example, the ROI can be evenly spaced perpendicular to the centerline of the vessel (such as the coronary artery). The TAG process can be configured to measure the gradient (rate of decline) of the contrast opacity along the vessel. For example, TAG can be defined as the average value of opacity and how the average value changes along the vessel. For example, TAG can be calculated as the regression coefficient of a line that fits a curve of the average value of such contrast opacity relative to the distance from a reference point (e.g., the coronary artery ostium). Such TAG type processing is described by AJ Einstein in Journal of the American College of Cardiology (Volume 61, Issue 12, 2013, Pages 1280-1282).

[0074] If the image measurement component IMC calculates such quantities on the reconstructed image alone as is, the flow increase caused by the presence of contrast agent (and as estimated by the flow increase predictor PC) may compromise the accuracy of the quantity (if not taken into account). Therefore, it is proposed herein that the intra-image measurement component IMC uses the estimate of the flow increase calculated by the predictor module PM to provide improved (such as more accurate and / or with better intra-patient / inter-patient comparability) hemodynamic quantities. Therefore, the intra-image measurement component IMC calculates hemodynamic quantities (hemodynamic, flow-based, etc.) based on input data (including image data) and the flow increase estimate. Specifically, in some embodiments, it is proposed herein that the intra-image measurement component IMC uses the estimate of the flow increase calculated by the predictor module PM to compensate for such flow increase by correcting, normalizing or otherwise modifying the initial measurement of the intra-image measurement component IMC, thereby providing improved (such as more accurate) hemodynamic quantities. Therefore, there may be explicit and different such compensation steps. This compensation may be handled by the intra-image measurement component IMC itself, or by a downstream, internal or external compensator component / service CC. The quantities thus improved may then be stored and may also be further processed by the diagnostic system DS as mentioned above or displayed as required. Alternatively, the flow increase estimate is already used by the intra-image measurement component IMC when calculating the hemodynamic quantities. Therefore, no initial version of the quantities which need to be corrected in the second step is needed. The flow estimate is included in the calculation of the hemodynamic quantities from the outset. Therefore, the flow increase is implicitly taken into account in the calculation. The operation of taking into account the estimated increased flow when calculating the hemodynamic quantity d may also be referred to herein as Δu compensation (operation).

[0075] Δu compensation can be configured as a linear or nonlinear function C that maps Δu (such as a percentage) to a corresponding compensation effect ω, such as an offset to be applied to the hemodynamic quantity d. The Δu compensation operation can be implemented using a threshold-based strategy. For example, if the estimated flow increase Δu is p% and exceeds p0%, the same percentage value p% (or an appropriately scaled version β*p%) can be used to mark down (or up) the diagnostic value initially calculated / measured based solely on the spectral image data Vt. For example, if the increase in flow has been found to be p%>p0, the calculation of the FFR value can be increased or decreased by p%. In all embodiments, the threshold method is not required in this article, and a floating upward or downward marking compensator strategy can be applied instead. Whether the estimated increase Δu attracts a relevant upward or downward marking of the image-based hemodynamic quantity d may depend on the type or nature of the hemodynamic quantity d, the underlying modeling and medical basic principles, etc.

[0076] Δu operations (such as correction, normalization, offset down marking, etc.) based on the calculated flow Δu ​​can be done by simply subtracting the corresponding percentage value, or can be done by forming a ratio of the initial diagnostic value based on a normalization / scaling factor calculated according to the increased flow, etc. The Δu operation can take any suitable form, wherein the Δu value is appropriately combined with the (initial) diagnostic value(s) and / or image value(s) of the input spectral image V or series thereof for one or more ROIs. For example, in one exemplary embodiment, the resting flow rate measurement (Δu is constructed as an increase %) can be corrected to d 校正 =d 测量 / (c+Δu) or similar, or any function thereof. The constant c may be a non-zero constant, such as c=1. As another example, in one exemplary embodiment, the microvascular resistance IMR (again, Δu is constructed as an increase in %) may be corrected to

[0077] d 校正 =d 校正 *(c+Δu). In general, a Δu operation (such as a correction or the like) can thus be constructed as a function of i) an applicable measured value of the hemodynamic quantity d and ii) a factor 1 / (c+Δu) or (c+Δu).

[0078] If contrast agent is not present, an estimated increased flow can be obtained relative to the normal flow expected in the vessel. This normal baseline flow can be estimated based on medical knowledge, or can be calculated, for example, during the ramp-up phase or washout phase when very little contrast agent is present, but just enough to allow a rough estimate of normal baseline blood flow. For such low concentrations of contrast agent, the autoregulatory response of the vessel caused by the presence of contrast agent is negligible. However, such knowledge of the baseline flow is optional and may not be required for practicing the principles described herein for obtaining more accurate / reliable image-based diagnostic values ​​d.

[0079] The flow Δu ​​calculated herein can be understood as a volume flow rate or flow velocity. The volume flow rate or flux describes the amount (volume or mass) of the contrast agent / blood mixture passing through a unit area in a unit time. Both the volume flow rate and the flow velocity can be calculated as needed, or only one of the two can be calculated. The exact form and expression of Δu can depend on the nature of the amount d that people are considering improving. Therefore, in some cases, the volume flow rate can be called, and in some other cases, the flow velocity can be called.

[0080] The flow increase estimate provided by the predictor PC may be a single value Δu=u0, or may be a time curve Δu(t) of such an estimate, for example when the input data V=V tWhen having such a time series character. There can be one (e.g. exactly one) such estimate Δu(t) per frame V(t), or more than one. Higher dimensional applications are also envisaged, where there is also a position s dependency for the spatial resolution: Δu(t,s). In some embodiments, it is specifically and preferably envisaged to calculate such a space-time resolved flow increase compensation. A dependency Δu(s) that varies only spatially is also envisaged. In the presence of a time series Δu(t) of flow estimates, multiple such Δu values ​​can be used for the Δu-compensated hemodynamic quantity d, or a single such Δu value can be used for each hemodynamic quantity d.

[0081] For example, one "scan" (projection data set) can be used to reconstruct a time series such as multiple 3D images Vt for different time points t, for example if the CT IA is kept rotating multiple times. Therefore, Δu values ​​can be calculated at different time points to provide a time-resolved curve of flow increase estimates Δu(t) as mentioned above. Such a time-resolved Δu curve can then be used to correct a single hemodynamic quantity describing the entire time range, or multiple such single hemodynamic quantities, one or more per time point t. For example, in perfusion imaging, peak time contrast can be measured. There are 3D images Vt for multiple time points before the peak appears. Each 3D image Vt can result in a different corresponding flow increase prediction Δu(t). And the peak time estimate may be affected by some or all previous flow states. As another option, and more generally, the time-dependent flow increase estimates Δu(t) together with their matching time-dependent hemodynamic quantities d(t) can be used to extrapolate corrected hemodynamic quantities under specific flow increase states (e.g., Δu=0) that have never been actually measured.

[0082] Calculating TAG as a hemodynamic quantity is one example of the use of the proposed system with multiple ROIs (such as the ROIs described above acquired along blood vessels), where the measurements (blur) are compensated by Δu.

[0083] The vital sign measurement device MD may collect vital sign measurements of the patient, such as temperature, blood pressure, etc. These vital sign measurements may be combined with corrected medical quantities to calculate a more advanced hemodynamic model of the vasculature VS of interest In simpler embodiments, the Δu values ​​and / or the hemodynamic quantities d derived therefrom may be sufficient to construct such a hemodynamic model.

[0084] The estimated increased flow Δu ​​data provided by the predictor module PM may be used by the plotter RD to provide a sensory representation that can be interpreted or otherwise used / utilized by a human user (such as medical personnel). For example, a transducer TR may be used that provides an alarm signal, such as visual, auditory or tactile or other form, if the estimated flow increase exceeds a certain threshold, at which point the accuracy of the hemodynamic quantities calculated by the IMC may be seriously compromised and countermeasures are advisable. For example, a user may choose to wait until the contrast agent concentration has been sufficiently reduced so that the autoregulatory response to open the vasculature or part thereof subsides, and then base the calculation of the hemodynamic quantities on volume data available at a later time.

[0085] Thus, a properly presented calculated value Δu may provide a quality control feature that allows a user to better assess the accuracy, authenticity or reliability of the calculated diagnostic value d based solely on the image information. The user may then manually initiate a Δu compensation operation, rather than having it done automatically. Even if the Δu compensation operation is done automatically, a threshold method may still be used as described above to initiate a Δu compensation operation with respect to the d value only when a certain threshold value p0 is exceeded. Thus, the Δu value may be used for monitoring purposes, such as for quality control of the diagnostic value output by the intra-image measurement component IMC.

[0086] The renderer RD may use the visualizer VIZ to generate a graphical display that can be displayed on the display device DD. Such a graphical display may include, for example, the current image and additional data informing the user of the estimated flow increase Δu. For example, a bar symbol of a different color and / or length may be indicated in correspondence with the estimated Δu value. Thus, the graphical display may be updated in real time and / or dynamically as new image data is processed to calculate the corresponding Δu value, since these values ​​are mapped to a bar symbol or any other symbol / widget that allows the representation of a changing value of Δu. A time curve may be used instead of, or in addition to, such a GUI widget to represent the changing value.

[0087] The hemodynamic quantities improved / calculated based on the Δu values ​​can be processed by the diagnostic system DS which can calculate other parameters, such as scores for certain diseases. Such scores can support or inform downstream therapeutic measures.

[0088] The improved / calculated hemodynamic quantities based on the Δu data, or indeed the Δu data itself, may be provided to the control interface IC. In addition to or in lieu of any one or more of the above-mentioned systems (diagnostic system DS, transducer TR, etc.), the control interface IC may provide control signals or parameters m-OUT to control other medical equipment or procedures. For example, the control signals or parameters m-OUT may be used to support an ongoing imaging procedure by (re)setting parameters of the imager IA and / or by controlling a contrast agent delivery device DA (such as a pump), etc.

[0089] The predictor module PM can be implemented by analytical modeling using, for example, a threshold-based approach or the like. In some such embodiments, the amount of contrast agent that causes the increase in flow, such as the mass or volume of contrast agent present, is extracted from the spectral input image by a segmentation operation. Since the contrast in the 3D spectral image is tuned to the contrast agent material, an accurate voxel-based thresholding method can be performed to identify the contribution of the contrast agent, thereby facilitating a good estimate of the contrast dose. Based on the extracted contrast dose, the increase in blood flow caused by the presence of the contrast agent can be calculated. Therefore, the extracted contrast dose can be mapped to a flow increase Δu value, such as using a linear function. The linear function maps the extracted contrast dose to an increase in blood flow. A nonlinear function may be used instead. The explicit model function, such as its coefficients and form, etc., may be based on medical knowledge. For example, the underlying physiological aspects of the autoregulatory response may be explicitly modeled in the analytical model function.

[0090] However, such explicit analytical modeling of the relationship of the spectral image information with respect to the flow increase is not necessarily necessary in this article, because instead of or in addition to such analytical modeling (hybrid modeling), a general machine learning approach is also envisaged. Therefore, the predictor model PM can be implemented by a trained machine learning model M, which is trained on previous training data including, for example, images that can be found in medical databases. A regression model or a classifier model can be used. A neural network model can be used.

[0091] The increase in flow is considered to be the result of a potentially complex physiological interaction of various factors. As described above, for example, the displacement of blood in the blood vessels by contrast agents over a certain period of time triggers a kind of "fight and flight" response of the body, in which vasodilation occurs, resulting in the observed increase in flow. However, the precise nature and onset of such vasodilation may vary from patient to patient depending on the patient's physiological characteristics (such as body mass index, age, gender, patient medical history, etc.). Therefore, it is assumed that there is a potential relationship between contrast agent concentration / hemodilution and the associated increase in flow. But this relationship may be complex: for example, CA-induced vasodilation may not necessarily be set exactly when the maximum amount of contrast agent has been accumulated. Instead, there may be a certain delay δ after the maximum CA concentration, after which the blood vessels expand and the flow increases. And there may be another delay δ' during which the blood begins to refill this expanded part of the blood vessel. It may be challenging to model the applicable physiological characteristics and / or the mentioned delay factors δ, δ' analytically. Moreover, this autoregulatory vascular response may vary from patient to patient. Therefore, to better account for these physiological effects, such as delays Δ, Δ', etc., in a preferred embodiment, the predictor PM uses machine learning ("ML"), where such relationships can be modeled using a general model trained on clinical training data or experimentally generated, and / or synthetically generated training data through simulation, etc. Typically, the predictor PM (whether implemented in ML or analytically) is configured to correlate input spectral images with flow changes (such as flow increases or, in other embodiments, flow decreases).

[0092] Reference now Figure 4 , which shows a block diagram of a machine learning implementation of a predictor module PM according to one embodiment.

[0093] Specifically, the machine learning model M can be arranged as an artificial neural network ("NN"), preferably as a convolutional type of NN, called a "CNN". CNN is useful when processing spatially related data (such as image data), as explained above, the processing of image data is contemplated herein.

[0094] The network M may comprise a plurality of computational nodes arranged in layers in a cascaded manner, wherein the data flow proceeds from left to right and thus from layer to layer. Thus, the model may be of feed-forward type. Recurrent networks are not excluded herein, in particular if the input data is a time series of volumes Vt.

[0095] The model network M can be said to have a deep architecture because it has more than one hidden layer. In a feedforward network, the "depth" is the number of hidden layers Lj (1≤j≤N) between the input layer IL and the output layer OL, while in a recurrent network, the depth is the number of hidden layers multiplied by the number of passes.

[0096] The machine learning model includes certain parameters that are adjusted during the training phase, such as the weights of the convolutional filters CV, which will be explained in more detail below. Once trained, i.e., once the parameters are appropriately adjusted, an input x is input at the input layer IL during deployment or testing and receives the input x, including the input contrast image V that may be enriched by context data c. Therefore, the input x is V or (V,κ). The input data x in deployment or testing propagates through the layers of the model and appears at the output layer OL as the expected result M(X)=Δu. Therefore, at the output layer OL, the expected result Δu is produced. Therefore, the output can be regressed to a scalar value Δu, and in some embodiments, there is one such scalar for a given time t and position s.

[0097] Thus, the output layer OL may be a regression layer, where the input data, including the image V or (V,c), is regressed into a single scalar value Δu, possibly with (s,t) dependencies. Alternatively, a classifier layer OL may be used, where the input is classified into Δu value levels, as such a qualitative output may be sufficient in some cases. For example, in some embodiments, the output may be sufficient to simply indicate which range / interval the estimated flow increase falls into, such as Δu∈[5%,10%]. In this classifier case, the output may be a vector whose entries represent the categories of the Δu value levels.

[0098] In the case of classification results, the output layer OL can be configured as a softmax function layer or similar computational node, where the feature maps from (one or more) previous layers are combined into normalized counts to represent the classification probability or score for each category. In the regression setting, OL can be a fully connected layer.

[0099] Preferably, some or all of the hidden layers Lj (1≤j≤N) and optionally the input layer are convolutional layers, i.e., include one or more convolutional filters CV that process the input feature maps from earlier layers into layer outputs, sometimes called logit. For example, an optional bias term can be applied by adding. The activation layer processes the logit into the next generation feature map in a nonlinear manner, which is then output and passed as input to the next layer, and so on. The activation layer can be implemented as a linear rectification function RELU as shown, or as a soft-max function, a sigmoid function, a hyperbolic tangent function, or any other suitable nonlinear function. In contrast to the convolutional layer, the output layer can be a fully connected layer. A hybrid network is also contemplated herein, comprising fully connected layers and convolutional layers.

[0100] Additional optional layers such as dropout layers, pooling layers P, etc. can be used in the CNN model M. The pooling layer reduces the size of the output while the dropout layer acts as a connection between nodes from different layers.

[0101] The trained model may be stored in one or more data storage devices MEM'.

[0102] Preferably, in order to achieve good throughput, the computing device PU comprises one or more processors (CPUs) supporting parallel computing, such as those of multi-core design. In one embodiment, (one or more) GPUs (Graphics Processing Units) are used.

[0103] However, the NN model M is not the only model contemplated: support vector machines (SVM), decision trees, random forests, linear regression or other models may still be used beneficially in some cases. However, due to the spatially correlated nature of the input images V, NN type models have been found to produce good results. Recurrent models such as LSTM are beneficial for time series data.

[0104] Reference now Figure 5 , which shows a training system TS, which can be used to train such a model M based on training data (x, y) such as can be stored in a training data storage TD. The training data (x, y) can be obtained from existing medical data, such as can be found in a medical database (such as a PACS, HIS or other database or storage system). Such training data may include historical image data or other health records of patients who have undergone similar examinations, in which contrast images of the relevant ROI were acquired. For example, the historical images can be viewed by a human expert. The expert can view the contrast images of historical patients and can select historical images corresponding to a certain flow increase or contrast agent concentration based on their medical knowledge. The human expert can view the patient health records to obtain annotations about the corresponding patients to draw their conclusions. Therefore, the human expert can assign corresponding scores y (or labels) to different corresponding samples of the training input image x. The scores yε[ab] are on a predefined scale [a, b]. The score y represents the expert's estimate of the corresponding flow increase associated with the corresponding input image x. Information about the contrast agent concentration can be used by the user to estimate the flow increase based on his medical experience and expertise.

[0105] Experimental options to obtain the labels y are also envisaged, for example based on a phantom. Moreover, training data can be built up over time in an intervention where flow measurements are made and recorded. General relationships can be obtained from previous invasive angiography and can be converted and used for training on CT.

[0106] Corresponding (historical) context data κ, such as historical flow-related measurements or notes, can be used to assist the user in assigning its score y. The context κ can be found in the health records of historical patients. In some cases, the health record context κ can be used to infer indications for increased flow. Specifically, medical records associated with historical images can be collected by human experts or by automatic text analyzer tools (such as NLP pipelines or string matcher tools, such as grep, etc.) to extract corresponding flow-related states, because this information can be mentioned in corresponding reports or records referring to specific images. In this way, the collection process can be fully or partially automated, allowing appropriately labeled training data construction to be obtained more quickly. The context κ can also include patient biological characteristics, such as the patient's gender, weight, height, BMI, etc. Such context data can be processed together with the image training data input to improve learning.

[0107] The training managed by the training system TS is the process of adapting the parameters of the model based on the training data. An explicit model is not necessarily needed, since in some examples it is the training data itself that constitutes the model, such as in clustering techniques or k-nearest neighbors, etc. In explicit modeling, such as in NN-based methods and many other methods, the model may include a system of model functions / computation nodes, where their inputs and / or outputs are at least partially interconnected. The model functions or nodes are associated with the parameters θ that are adapted in the training. The model functions may include the above in conjunction with the NN-type model in Figure 4 The convolution operator and / or weights of the nonlinear unit (such as RELU) mentioned above. In the case of NN, the parameters θ may include the weights of the convolution kernel of the operator CV and / or the nonlinear unit.

[0108] The parameterized model can be formally written as M θ . Parameter adaptation can be implemented by a numerical optimization process. The optimization can be iterative. The objective function f can be used to guide or control the optimization process. The parameters are adapted or updated by the updater UP so that the objective function is improved. i Applied to the model. The model responds to produce the training data output M(x i )=y i The objective function maps from the parameter space to a set of numbers. The objective function f measures the output y of the training data. i With the corresponding target y i The parameters are iteratively adjusted to reduce the combined deviation until the stop condition preset by the user or designer is met. The objective function can use the distance metric ||.|| to quantify the deviation.

[0109] In some embodiments, but not all embodiments, the combined bias may be implemented as a function of the training data instance / pair (xi ,y i ) i The sum of some or all residuals of , and the optimization problem in terms of the objective function can be formulated as:

[0110] argmin θ f=∑ k ||M θ (x k ),y k || (1)

[0111] In setting (1), the optimization is formulated as the minimization of a cost function f, but this is not limiting here, as the dual formulation of maximizing a utility function may be used instead. Sum over the training data instances i.

[0112] The cost function f may be pixel / voxel based, such as an L1 or L2 norm cost function. For example, in least squares or similar methods, the (squared) Euclidean distance type cost function in (1) may be used for the above regression task. When the model is configured as a classifier, the sum in (1) is instead formulated as one of the cross entropy or Kullback-Leiber divergence or the like. The Huber cost function may be used.

[0113] The exact functional composition of the updater UP depends on the optimization procedure implemented. For example, an optimization scheme such as back / forward propagation or other gradient-based methods can be used to adapt the parameters θ of the model M in order to reduce all or a subset of the training pairs (x k ,y k ). Such subsets are sometimes called batches, and optimization can be performed in batches until all of the training dataset has been exhausted, or until a predetermined number of training data instances have been processed.

[0114] Training can be a one-time operation, or can be repeated as new training data becomes available.

[0115] Optionally, one or more batch normalization operators ("BN", not shown) may be used. The batch normalization operator may be integrated into the model M, for example coupled to one or more convolution operators CV in a layer. The BN operator allows mitigation of the vanishing gradient effect, a gradual decrease in the magnitude of the gradient in repeated forward and backward passes experienced during a gradient-based learning algorithm in the learning phase of the model M.

[0116] For all learning scenarios (especially supervised ones), consider Figure 5 The training system shown. In alternative embodiments, unsupervised learning schemes can also be envisioned herein. A GPU can be used to implement the training system TS.

[0117] Figure 6 A flow chart of a computer-implemented method of estimating the blood flow increase due to the autoregulatory response of a considered vessel caused by the presence of a contrast agent, and in which the contrast agent is present, is shown. The calculated increase in blood flow / flux can be used to obtain more accurate diagnostic values, such as hemodynamic quantities or other such quantities describing or related to blood flow.

[0118] In step S610, volume spectral image data is received. The volume image data V=R(λ) is based on tomographic data reconstructed from projection data λ acquired by a spectral imaging device ("CTA" or angiographic CT (computed tomography)). The projection data is acquired when at least part of the contrast agent is at least sometimes present in the vascular system of interest (ROI) in the field of view. The projection data or the volume data is processed by a spectral processing algorithm to obtain a spectral image, such as contrast-only image data. In such spectral tomographic images as are primarily contemplated herein for input data, the contrast is primarily or exclusively a function of the amount of contrast. The contributing attenuation effects of intervening structures and / or other materials through which the acquisition X-ray beam is projected are largely eliminated. The symbol "V" is used herein to designate such an image volume obtained after applying such a spectral processing algorithm. The input spectral image volume V may include a series of such volumes, and there may be more than one ROI.

[0119] In step S620, the blood flow increase caused by the presence of the contrast agent is calculated based on the input spectral volume data.

[0120] At step S630, the estimated flow increase Δu is output as output data. The output data Δu may be a single flow increase value or a time series of these values. The output data may be provided as a percentage or as a flux value, a velocity value, etc. as required. The estimated flow increase Δu may be globally associated with the ROI as a whole, or may be locally associated with a subset of the ROI. The output may be spatially resolved to be associated with multiple ROIs.

[0121] In step S640, a hemodynamic quantity d, such as a hemodynamic volume, is calculated based on the spectral volume input image.

[0122] At step S650, the calculated amount is compensated for the effect of the flow increase, such as by correction, normalization or by other modification, based on the increased flow Δu ​​calculated at step S620. Thus, in an embodiment of the "explicit" type, an initial version d is calculated at step S640 and then improved by using the increased flow Δu ​​according to step S630. In other embodiments of the implicit type, steps S640 and S650 may alternatively be combined into a single processing step. If the Δu compensation step S650 is applied separately from the step S640 of calculating the diagnostic value (rather than integrated), the order of steps S620, S630 and S640 may be reversed.

[0123] The Δu compensation step S650, whether explicit or implicit, may be automatically and / or always applied according to an unconditional strategy, or only when certain conditions are met, such as when the estimated flow increase exceeds a threshold. Other conditions besides this threshold approach may alternatively be formulated.

[0124] Therefore, the hemodynamic quantities obtained herein are of better quality, such as being more reliable, more accurate and providing better cross-patient comparisons, as opposed to the quantities calculated without the Δu compensation step S650. The proposed method based on the Δu compensation step S650 allows eliminating or at least reducing interfering effects caused by an increase in flow / flux.

[0125] The corrected diagnostic value(s) and / or indeed the Δu data itself may be provided for processing, such as for calculating a score representing the presence of disease, or the value may be stored in a memory device, displayed, may be used for control purposes or may be otherwise processed as desired, etc.

[0126] Some of the above steps are now described in more detail.

[0127] For example, predicting / estimating the flow increase Δu at step S630 may be done analytically and / or using machine learning.

[0128] In the analysis+ML hybrid setting of step S630, the following sub-steps may be performed:

[0129] processing the image data to segment coronary arteries or other vessels of interest, and optionally perfusion regions;

[0130] Analyzing image values ​​(such as HU values) at voxels to estimate blood iodine concentration in the segmented region. The analysis may include a thresholding method; and

[0131] Estimate the blood flow increase Δu in the segmented region caused by the presence of diluted blood / contrast agent. For example, this can be done directly by estimating the dilution. For example, a linear model function can be used to relate contrast agent mass / volume / dilution / concentration to flow increase. For example and for simplicity, a uniform scaling factor β=1 can be used. In such a model M, a p% iodine concentration will result in a p% flow increase. Such a model can be based on a regression model trained on a clinical training data set derived from a medical database. Non-linear models are also envisioned.

[0132] The above-described mixing steps can be supplemented by incorporating hemodynamic modeling of the coronary system. Examples of such hemodynamic modeling can include a lumped resistor model (LRM). In such an LRM, the blood vessels can be modeled as a network of resistor elements, and boundary conditions can be used to model either or both of the pumping heart and the microvasculature with its autoregulation. This allows for better consideration of local and temporal variations in a particular vascular system. Again, linear types of hemodynamic modeling are contemplated herein, but nonlinear approaches are also contemplated.

[0133] However, step S630 may be implemented in a purely analytical manner as described above, or purely in ML. In the latter case, no explicit segmentation or other extraction may be required: the ML model may be trained end-to-end to map directly from the image data V to an output representing the desired Δu estimate(s).

[0134] The estimation step S630 of the flow increase of vascular autoregulation due to the presence of contrast agent may be further supported by the patient's live vital signs, such as blood pressure and heart rate, etc. Therefore, the input may also include such vital sign parameters as context data κ. Such context data κ is particularly useful for ML.

[0135] To achieve more accurate temporal information of contrast agent concentration and associated flow increases, a series of partial scan reconstructions from the CT data can be performed and used as input, rather than a single volume V (although this scenario is also envisioned herein). For example, using a time series of volume spectral data such as a series of partial reconstructions can be used to normalize the perfusion analysis over longer time scales.

[0136] The proposed method and system can also be applied in determining functional parameters to quantify peripheral vascular disease.

[0137] Reference now Figure 7 , which shows a flowchart of a method for training a machine learning model that can be used in step S620 above.

[0138] In step S710, for example, k ,yk ) form. As defined above, each pair includes a training input x k and the associated target y k .x k .

[0139] In step S720, the training input x k Applied to the initialized machine learning model M to produce training output.

[0140] In step S730, the training output M(x k ) and the associated target y k In step S740, one or more parameters of the model are adapted in one or more iterations in the inner loop to improve the cost function. For example, the model parameters are adapted to reduce the residual measured by the cost function. In the case of using a convolutional model M, the parameters particularly include the weights of the convolution operator.

[0141] The training method then returns in an outer loop to step S710 where the next pair of training data is fed. In step S740 the parameters of the model are adapted such that the aggregate residual of all pairs considered is reduced, in particular minimized. A cost function quantifies the aggregate residual. Forward-backward propagation or similar gradient-based techniques may be used in the inner loop. Instead of looping over individual pairs at step S710, a collection or batch of pairs of training data is looped over and these are fed at once in steps S720 and S730.

[0142] Examples of gradient-based optimization may include gradient descent, stochastic gradient, conjugate gradient, maximum likelihood method, EM maximization, Gauss-Newton, etc. Methods other than gradient-based methods are also contemplated, such as Nelder-Mead, Bayesian optimization, simulated annealing, genetic algorithms, Monte Carlo methods, etc.

[0143] More generally, the parameters of the model M are adjusted to improve an objective function F which is a cost function or utility function. In an embodiment, the cost function is configured to measure an aggregated residual. In an embodiment, the aggregation of the residuals is achieved by summing all or some of the residuals of all pairs considered. In particular, in a preferred embodiment, the outer summation is performed in batches (subsets of training instances), whose summed residuals are considered simultaneously when adjusting the parameters in the inner loop. The outer loop then proceeds to the next batch, and so on, until the necessary number of training data instances has been processed.

[0144] It will be appreciated that the above embodiment may be modified: instead of predicting / estimating a flow increase, a decrease in flow may instead be estimated. This may be caused by some types of contrast agents, and this effect likewise interferes with the calculation of the standardized diagnostic value d as intended herein. For example, some contrast agents may cause spasms in microvascular muscle cells. However, the estimation of the flow increase as described above is preferred herein.

[0145] The components of the system SYS may be implemented as one or more software modules, running on one or more general purpose processing units PU, such as a workstation associated with an imager IA, or on a server computer associated with a group of imagers.

[0146] Alternatively, some or all components of the system SYS may be arranged in hardware, such as a suitably programmed microcontroller or microprocessor, such as an FPGA (field programmable gate array) or as a hardwired IC chip, an application specific integrated circuit (ASIC) integrated into the imaging system IA. In yet another embodiment, the system SYS may be implemented partly in software and partly in hardware.

[0147] The different components of the system SYS may be implemented on a single data processing unit PU. Alternatively, some or more components are implemented on different processing units PU, possibly arranged remotely in a distributed architecture and connectable in a suitable communication network, such as in a cloud setup or a client-server setup or the like.

[0148] One or more features described herein can be configured or implemented as or with circuits encoded in a computer-readable medium, and / or combinations thereof. The circuits may include discrete and / or integrated circuits, systems on a chip (SOCs) and combinations thereof, machines, computer systems, processors and memories, computer programs.

[0149] In a further exemplary embodiment of the present invention, a computer program or a computer program element is provided, which is characterized by being adapted to execute the method steps of the method according to one of the preceding embodiments on a suitable system.

[0150] Therefore, the computer program element can be stored on a computer unit, which can also be a part of an embodiment of the present invention. The computing unit can be suitable for executing or inducing the steps of the method described above. In addition, it can be suitable for operating the components of the device described above. The computing unit can be suitable for automatic operation and / or execution of user's commands. The computer program can be loaded into the working memory of a data processor. Therefore, a data processor can be equipped to perform the method of the present invention.

[0151] This exemplary embodiment of the invention covers both a computer program which uses the invention from scratch and a computer program which by an update turns an existing program into a program which uses the invention.

[0152] Furthermore, the computer program element may be able to provide all necessary steps to complete the procedures of an exemplary embodiment of the method as described above.

[0153] According to a further exemplary embodiment of the present invention, a computer-readable medium, such as a CD-ROM, is proposed, wherein the computer-readable medium has a computer program element stored thereon, the computer program element being described by the preceding section.

[0154] The computer program may be stored and / or distributed on a suitable medium (particularly, but not necessarily, a non-transitory medium) such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.

[0155] However, the computer program may also be presented over a network like the World Wide Web and can be downloaded from such a network into a working memory of a data processor. According to a further exemplary embodiment of the invention, a medium for making a computer program element available for downloading is provided, the computer program element being arranged to perform a method according to one of the previously described embodiments of the invention.

[0156] It must be noted that embodiments of the present invention are described with reference to different subject matters. In particular, some embodiments are described with reference to method type claims, while other embodiments are described with reference to device type claims. However, a person skilled in the art will appreciate from the above and following descriptions that, unless otherwise stated, any combination of features relating to different subject matters, in addition to any combination of features belonging to one type of subject matter, is also considered to be disclosed by the present application. However, all features can be combined to provide a synergistic effect that is more than the simple sum of the features.

[0157] Although the present invention has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description are to be considered illustrative or exemplary rather than restrictive. The present invention is not limited to the disclosed embodiments. Other variations of the disclosed embodiments can be understood and implemented by those skilled in the art in practicing the claimed invention by studying the drawings, the disclosure, and the dependent claims.

[0158] In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope. Such reference signs may include numbers, letters or any alphanumeric combination.

Claims

1. A system for medical image processing, comprising: an input interface (IN) through which input data comprising a spectral input image of at least a portion of the vascular system (V) of a patient (PAT) can be received when the system is in use, said spectral input image being based on data reconstructible from projection data acquired by a spectral imaging device of the tomographic type, wherein a contrast agent is present in said portion of said vascular system; a predictor module (PM) configured to predict, based on the input images, an increase or decrease in flow caused by the presence of a contrast agent; and An output interface (u-OUT) for providing output data indicative of a predicted increase or decrease in flow.

2. The system according to claim 1, comprising an intra-image measurement component (IMC) configured to measure a medical quantity of interest based on the input image and the output data.

3. The system according to claim 2, wherein: The intra-image measurement component (IMC) combines information from the input image with information based on the output data.

4. The system according to claim 3, wherein: The combining by the intra-image measurement component (IMC) includes correcting or normalizing, based on the output data, initial measurements of the medical volume of interest made by the intra-image evaluation component (IMC) based on the input image.

5. A system according to any one of the preceding claims, wherein: The medical quantities of interest include hemodynamic quantities such as any one or more of the following: perfusion fraction, fractional flow reserve FFR, instantaneous flow ratio iFR, coronary flow reserve CFR, microvascular resistance index IMR, and intraluminal attenuation gradient TAG.

6. A system according to any of the preceding systems, comprising a renderer (RD) configured to provide a sensory indication for a user based on the output data, wherein: The renderer (RD) comprises any one or more of: i) a visualizer (VIZ) configured to cause display of the output data on a display device (DD); ii) a transducer (TR) configured to generate an alarm signal associated with the output data.

7. The system according to any of the preceding claims, comprising a control interface (IC) configured to issue control signals to control medical equipment based on the output data.

8. A system according to any one of the preceding claims, wherein: The increase or decrease in flow is caused by an autoregulatory response of the vasculature caused by the presence of contrast agent.

9. A system according to any one of the preceding claims, wherein: The predictor module (PM) is based on a trained machine learning model (M).

10. A system according to any one of the preceding claims, wherein: The spectral imaging device is configured for computed tomography angiography.

11. An arrangement (MIA) comprising a system according to any one of the preceding claims and the imaging device.

12. A method for medical image processing, comprising: receiving (S610) input data comprising a spectral input image of at least a portion of a vascular system (VS) of a patient (PAT), said spectral input image being based on data reconstructible from projection data acquired by a spectral imaging device of the tomographic type, wherein a contrast agent is present in said portion of said vascular system; estimating (S620) a flow increase or decrease caused by the presence of a contrast agent based on the input image; and Output data indicative of the predicted increase or decrease in flow is provided (S630).

13. A method for training a machine learning model according to claim 9 based on training data.

14. A computer program element, which, when executed by at least one processing unit, is adapted to cause the processing unit (PU) to perform the method according to any of claims 12-13.

15. At least one computer-readable medium having stored thereon the program unit of claim 14, or having stored thereon the machine learning model of claim 9 or 13.

Images