Assessment of blood flow and functionally significant stenosis

Abstract

Described herein is a computer implemented method for assessing functionally significant stenosis. Also described, is a method of estimating distribution of tracer mass in a plurality of cardiovasculatures of interest. Also described, is a method of estimating stenosis -induced pressure loss. Systems and non-transitory computer-readable media for executing the methods are also described.

Description

[0001] 02 OCTOBER 2025 (02.10.2025)

[0002] LHS6-0242PCT PATENT

[0003] ASSESSMENT OF BLOOD FLOW AND FUNCTIONALLY SIGNIFICANT STENOSIS

[0004] BACKGROUND OF THE INVENTION

[0005] Field of the Invention

[0006] The present invention relates to dynamic imaging of flow, and more particularly to assessment of a blood flow characteristic in a subject based on dynamic imaging of contrast agent flow through a blood vessel or heart structure.

[0007] Description of the Related Art

[0008] Dynamic contrast-enhanced (DCE) computed tomography (CT) has been used to assess blood flow and flow pressure in blood vessels, for example as described in co-owned International PCT Application No. PCT / CA2019 / 050668 filed 16 May 2019 (published as WO2019 / 218076 on 21 November 2019), incorporated herein by reference. In this disclosure, Indicator Dilution Principle is used for blood flow assessment derived from DCE CT representing average values over the course of many seconds of imaging scans, often greater than 10 seconds, and therefore does not have sufficient temporal resolution to achieve 4D flow imaging, that is, to track a blood flow characteristic in selected image voxels at very fine temporal resolution, for example calculating changes in flow velocity at time intervals of less than 1 second.

[0009] An improvement with respect to fine temporal resolution is described in co-owned International PCT Application No. PCT / CA2021 / 051189 filed 26 August 2021 (published as W02022 / 040806 on 03 March 2022), incorporated herein by reference. In this disclosure, Indicator Dilution Principle and Reynolds Transport Theorem are used to substantially improve the temporal resolution of non-invasive blood flow assessment with DCE CT imaging data.

[0010] These two disclosures improved upon the existing state of CT imaging, and introduced novel extractions and implementations of parameters from a time-enhancement curve at a region of interest, including for example, area-under-curve, time rate of change of tracer mass in the region of interest, or density of tracer in blood in the region of interest.

[0011] As diagnostic imaging is a very active area of clinical workflow, further improvement is welcomed by medical practitioners including for example, improvement in computer efficiency, reduction of radiation dose, or improvement in accuracy.

[0012] Accordingly, there is a continuing need for alternative methods and systems for blood flow imaging based assessment of a blood vessel in a subject. 02 OCTOBER 2025 (02.10.2025)

[0013] SUMMARY OF THE INVENTION

[0014] In an aspect there is provided, a computer implemented method for assessing functionally significant stenosis comprising: obtaining CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; extracting at least one image feature from the CT or MRI image data; providing the at least one image feature and an associated at least one non-image feature to a machine learning model to generate a prediction of a functionally significant stenosis within the cardiovasculature of interest, the machine learning model trained with training inputs of the at least one image feature with the at least one non-image feature, and associated with a binary (YES / NO) assessment of functionally significant stenosis as ground truth.

[0015] In another aspect there is provided, a method of estimating distribution of tracer mass in a plurality of coronary cardiovasculatures of interest.

[0016] In another aspect there is provided, a method of estimating stenosis-induced pressure loss.

[0017] In other aspects, systems and non-transitory computer-readable media for executing the methods are also provided.

[0018] BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 shows a schematic of a blood flow imaging system.

[0020] Figure 2 shows a flow diagram of a blood flow imaging method.

[0021] Figure 3 shows a flow diagram of a pre-scan preparation in the imaging method shown in Figure 2.

[0022] Figure 4 shows a flow diagram of scan data acquisition in the imaging method shown in Figure 2.

[0023] Figure 5 shows a flow diagram of an example of predicting an area under timeenhancement curve (AUC) in the imaging method shown in Figure 2.

[0024] Figure 6A shows a flow diagram of an example of determining a blood flow characteristic based on the predicted AUC in the imaging method shown in Figure 2. Figure 6B shows a flow diagram of another example of determining a blood flow characteristic based on the predicted AUC in the imaging method shown in Figure 2.

[0025] Figure 7A and Figure 7B show schematic illustrations of two approaches for studying the movement of fluid.

[0026] Figure 8 shows a schematic representation of voxel -defined control volume analysis of flow imaging data; Figure 8A shows a schematic representation of a control volume and its control 02 OCTOBER 2025 (02.10.2025) surfaces; Figure 8B shows a schematic representation of movement of fluid with respect to a control volume.

[0027] Figures 9A and 9B shows schematic illustration for two different approaches of dynamic image acquisition and reconstruction for obtaining AHU / At that is a basis for absolute or relative flow velocity assessment. Figure 9A shows image acquisition and reconstruction at the same cardiac phase (e.g. 75% R-R interval or diastole) over multiple time points. Figure 9B shows image acquisition covering a full cardiac cycle (systole and diastole); multiple image sets corresponding to different cardiac phases are reconstructed.

[0028] Figure 10 shows a schematic illustration of an example of workflow of a cardiac CT test.

[0029] Figure 11 shows a schematic illustration of an example of using machine learning or deep learning to derive blood flow variables from a pseudo truncated dynamic perfusion scan.

[0030] Figure 12 shows a schematic illustration of an example of using machine learning or deep learning to derive blood flow variables from truncated rest dynamic perfusion scan.

[0031] Figure 13 shows a schematic illustration of an example of using machine learning or deep learning to derive blood flow variables from truncated stress dynamic perfusion scan.

[0032] Figure 14 shows a schematic illustration of an example of generation of training image features for training a ML or DL algorithm to learn from and make prediction on the AUC values for future truncated perfusion image data. Top: training path. Bottom: application path.

[0033] Figure 15 shows a schematic illustration of an example of extraction of image features with optional incorporation of non-image features in a coronary artery for training a ML or DL algorithm.

[0034] Figure 16 shows a schematic illustration of an example of extraction of image features with optional incorporation of non-image features in the myocardium for training a ML or DL algorithm.

[0035] Figure 17 shows a simplified diagram to illustrate an example of architecture of a deep neural network (convolution neural network). In this illustration, only one convolution filter is shown, resulting in only one feature map and pooled feature map from each round of convolution and pooling. The grids in each feature map and pooled feature map do not represent the actual dimensions in these maps, and only serve to demonstrate reduced dimensions in these maps after each operation. If an input image has M x M pixels and a convolution filter has C x C pixels, the feature map should have (M-C+l) * (M-C+l) pixels.

[0036] Figure 18 shows an example of an analog of the deep neural network shown in Figure 17, but with three convolution filters applied. 02 OCTOBER 2025 (02.10.2025)

[0037] Figure 19 shows a schematic illustration of an example of deep neural network architecture for handling a dynamic perfusion image set. For visual simplification, only a portion of the architecture is shown.

[0038] Figure 20A shows post-stenotic AUC normalized to pre-stenotic AUC as a function of luminal narrowing in five patients with coronary artery disease (CAD). Figure 20B shows sampling RO Is in the pre-stenotic and post-stenotic segments in a narrowed coronary artery.

[0039] Figure 21 shows stress AUC normalized to rest AUC in pre-stenotic and post-stenotic coronary segments in 5 patients with CAD.

[0040] Figure 22 shows coronary time-enhancement curves obtained in two different patients with different heart rates. Open circles are the original data and solid lines are the corresponding fitted curves. The projected baseline and peak enhancement of each fitted curve are denoted by the blue dashed horizontal and vertical lines, respectively.

[0041] Figure 23 shows time-enhancement curve measured from the left anterior descending (LAD) artery (top row) and the myocardium perfused by the LAD (bottom row) in the same patient.

[0042] Figure 24A and Figure 24B show an increase of thickness of the myocardial wall (dotted lines) increases when comparing a maximal vasodilatory stress condition (Fig. 24A) to a resting condition (Fig. 24B). The right coronary artery (RCA; indicated by arrows) also appeared to be more dilated at stress. Figure 24C shows the corresponding time-enhancement curves measured from the RCA at stress and rest.

[0043] Figure 25A shows a default orientation of two CT slices relative to a coronary artery. The red arrows depict the direction of blood flow in the artery. Figure 25B shows slice reformation in one slice location is needed to ensure the slice is perpendicular to the direction of blood flow prior to quantitative blood flow assessment.

[0044] Figure 26A shows a short-axis view and Figure 26B shows an axial view of a contrast- enhanced cardiac image of the same patient. Figure 26C shows a comparison of the coronary timeenhancement curves sampled at the same spatial location in the left anterior descending artery at the short-axis and axial views.

[0045] Figure 27A shows coronary time-enhancement curves sampled at the proximal and distal to a stenosis in the RCA in the stress dynamic perfusion images shown in Figure 27C. Figure 27B shows coronary time-enhancement curve sampled at the same locations across the RCA stenosis in the rest dynamic perfusion images shown in Figure 27D. Note that only the distal sampling location is shown in (Figure 27C and Figure 27D. 02 OCTOBER 2025 (02.10.2025)

[0046] Figure 28 shows AUCs of stress coronary time-enhancement curves versus AUCs of rest coronary time-enhancement curves sampled in different right coronary arteries with different degree of stenosis (percentage in lumen narrowing). All the data points shown in the graph are true measured AUC values.

[0047] Figure 29 shows two sections of Python codes. The top section shows three hidden layers were implemented in the neural network (LI, L2 and L3). No activation was specified in the output layer (last code line in the top section), which means a linear regression was used. By default, a(x) = x, where a is the activation function and x = wlbl+w2b2+ ...., where w are the weights and b are the parameters. The codes shown in the bottom section were used to execute the neural network layers defined above.

[0048] Figure 30A show predicted stress AUC values (represented by crosses) predicted by the trained machine learning model versus the true stress AUC values measured from stress dynamic perfusion images. The training (circle) and test (square) AUC datasets used for the machine learning model are shown in the background for comparison. The substantial overlap between the crosses and training / test datasets showed the good predictions of stress AUC values made by the model. Figure 30B shows the training and test scores and losses associated with the model training.

[0049] Figure 31 shows a schematic illustration of the proposed machine learning approach to predict the area under curve (AUC) of a time-enhancement curve at stress (maximal hyperemia) without actually acquiring the curve. The input AUC can be a coronary or myocardial timeenhancement curve. Likewise, the output AUC can be a coronary or myocardial time-enhancement curve. Because these curves are closely related to each other. Coronary AUC can be used as the input to predict myocardial AUC, or vice versa. These curves can also be used concomitantly as the input training data for the machine learning model.

[0050] Figure 32 shows a schematic illustration of how the trained machine learning model can be integrated into an existing image processing software to facilitate prediction of blood flow parameters corresponding to maximal hyperemia. The asterisk (*) is to note that the algorithms were described in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021.

[0051] Figure 33A shows a truncated perfusion dataset generated by removing data points after peak enhancement. Figure 33B shows AUC (shaded area) computed for the generated truncated data set, which was used for machine learning. 02 OCTOBER 2025 (02.10.2025)

[0052] Figure 34 shows AUCs of full time-enhancement curves sampled from dynamic perfusion images versus AUCs of truncated time-enhancement curves generated from the full curves for different coronary arteries with different degrees of stenosis.

[0053] Figure 35A shows AUCs (depicted by crosses) of full coronary time-enhancement curves predicted by the trained machine learning model in comparison to the AUCs of full timeenhancement curves measured from full dynamic perfusion scans. The training (circle) and test (square) datasets are provided in the background. A substantial overlap between the crosses and the training / test datasets showed that accurate prediction about the full AUCs was made by the trained model. Figure35B shows train loss and test loss of the machine learning model as a function of epoch (number of complete passes through the entire training datasets). The train and test losses in the flat portion of the plots were low and close to each other, suggesting that the training data was not overfitted by the model. Figure 35C shows the train and test loss and score returned by the model.

[0054] Figure 36 illustrates a machine learning scheme discussed in Example 6.

[0055] Figure 37 illustrates how the trained machine learning as trained according to the scheme shown in Figure 36, can be used for clinical hemodynamic assessment. The asterisk (*) is to note that the algorithms were described in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in coowned PCT / CA2021 / 051189 filed 26 August 2021.

[0056] Figure 38A shows measured short-axis AUC values plotted against measured axial AUC values (circle dots). Figure 38B shows predicted short-axis AUC values (crosses) in comparison to the measured AUC values used for training (circles) and test (squares). Figure 38C shows Train / Test losses and scores associated with the model training.

[0057] Figure 39A shows feature-labeled contrast-enhanced cardiac CT images in a 2D view and Figure 39B shows a corresponding 3D view. Figure 39C shows MATLAB codes used for implementing a three-dimensional U-NET architecture for segmenting the cardiac CT volumetric image sets shown in Figure 39A and Figure 39B. The settings used for training the U-NET are also shown.

[0058] Figure 40 illustrates how the training results from a deep learning model can be used as the input labels to train a machine learning model described in Examples 5-7.

[0059] Figure 41 shows a flow chart of a supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and real rest / stress image data. 02 OCTOBER 2025 (02.10.2025)

[0060] Figure 42 shows a flow chart of a dual supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and virtual stress image data.

[0061] Figure 43 shows a flow chart of a dual supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and virtual image data corresponding to a virtual tomographic orientation.

[0062] Figure 44 shows a flow chart of a dual supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and image data corresponding to virtual full (untruncated) time-enhancement curves.

[0063] Figure 45 shows a flow chart of a supervised machine learning algorithm that predicts the presence or absence of downstream microvascular dysfunction in the presence of functionally significant lumen narrowing in the upstream artery.

[0064] Figure 46 shows a flow chart of a dual supervised machine learning algorithm that predicts which image set of a dynamic contrast-enhanced imaging sequence corresponds to peak arterial enhancement, and using selected images for subsequent anatomical inspection and prediction of functional significance of arterial lumen narrowing.

[0065] Figure 47 illustrates location of a prelesion slice, a postlesion slice, an entry slice and an exit slice relative to a stenosis within an artery.

[0066] Figure 48 illustrates location of radius measurements for tracer percentage assignments.

[0067] Figure 49 shows a part of a computer programmable code written in Python to execute a supervised machine learning for binary classification of functionally significant coronary artery disease based on the image and non-image variables input to a neural network.

[0068] Figure 50 shows a dataset used to train a neural network for binary classification of functionally significant coronary artery disease using the Python program illustrated in Figure 49.

[0069] Figure 51 shows images slices and corresponding time-enhancement curves from two cases from the training dataset shown in Figure 50.

[0070] Figure 52 shows the training score and loss corresponding to the inputs “entry AUC” and “exit AUC” returned by the Python program illustrated in Figure 49.

[0071] Figure 53 shows a comparison of prediction of functional significant coronary artery disease made by the trained neural network (Figure 53B) versus the ground truth (Figure 53A) obtained from the dataset shown in Figure 50.

[0072] Figure 54 shows part of computer programmable code written in MATLAB to execute a supervised machine learning for segmenting the coronary arteries in contrast-enhanced CT images. 02 OCTOBER 2025 (02.10.2025)

[0073] Figure 55 shows an example of contrast-enhanced CT images used for supervised machine learning for coronary artery segmentation.

[0074] Figure 56 shows results of detection of the coronary arteries in a contrast-enhanced CT image with supervised machine learning including the original CT image (Figure 56A) for comparison, the original CT image with labels of the coronary arteries and background (Figure 56B), and the test result of prediction of the coronary arteries made by the trained neural network (Figure 56C).

[0075] Figure 57 shows a comparison of visualization and lumen diameter measurement of the right coronary artery (RCA) with a standard CCTA image set (Figure 57(A-C)) versus a dynamic angiographic image set at peak contrast enhancement (Figure 57(D-F)).

[0076] Figure 58 shows comparison of an artery having functionally significant stenosis without downstream microvascular dysfunction (Figure 58A) versus an artery having functionally significant stenosis with downstream microvascular dysfunction (Figure 58B).

[0077] Figure 59 shows a comparison of prelesion and postlesion coronary time-enhancement curves sampled in the patient in Figure 58A (Figure 59(A-D)) versus the patient in Figure 58B (Figure 59(E-H)).

[0078] Figure 60 shows a comparison of myocardial time-enhancement curves sampled in the myocardium supplied by the artery shown in Figure 58A (Figure 60(A and B) versus the artery shown in Figure 58B (Figure 60(C and D), with best-fit linear upslope for both curves plotted in Figure 60E.

[0079] Figure 61 shows a comparison of the X-ray tube current (mA) and exposure time (s) settings between a clinical CCTA scan (Figure 61A) versus a DCE scan (Figure 61B) on the same patient.

[0080] Figure 62. (A)-(B) Assessment of FFR across the prelesion and postlesion locations in the RCA artery (see the cross in each image). (C) Calculation of the residual tracer percentage at different locations along the RCA. The residual tracer percentages are shown in column J in the Excel screenshot. There were six branches between the two locations shown in (A) and (B). (D) Plot of residual tracer percentage in the RCA after passing through each branch. (E)-(J) The timeenhancement curves sampled from the prelesion and postlesion locations used to compute the prelesion and postlesion flow velocities, and to compute the pressure loss arising from the stenosis.

[0081] Figure 63. (A) Computer code written in MATLAB for executing a convolution neural network to predict the time point corresponding to the peak arterial enhancement in a dynamic CT perfusion image set. (B)-(C) Zoomed version of (A). 02 OCTOBER 2025 (02.10.2025)

[0082] Figure 64. Graphical representation of the convolution neural network as shown in Figure

[0083] 63.

[0084] Figure 65. Computer code for preparing the image sets to train the convolution neural network shown in Figure 63.

[0085] Figure 66. Partial list of the data used to train the neural network shown in Figures 63 and

[0086] 64.

[0087] Figure 67. Computer code for loading the image sets processed with the code in Figure 65 and running the trained neural network (code shown in Figure 63) to predict a time point for peak enhancement.

[0088] Figure 68A. Plot generated by running the code shown in Figure 67. Figure 68B. A magnification of the training results shown in Figure 68A.

[0089] DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

[0090] With reference to the drawings, a system and method for blood flow imaging is described. The system and method compare favourably with current blood flow imaging techniques.

[0091] Figure 1 shows a computer implemented imaging system 2, incorporating a computed tomography (CT) scanner 4. The CT scanner 4 may be any mutli-row or multi-slice CT scanner typically comprising a radiation source and a radiation detector disposed in a gantry and an adjustable, often motorized, support or table for maintaining a subject in a desired position (for example, a prone or supine position) in an open central chamber formed in the gantry during a scan procedure. The radiation source generates radiation that traverses one or more predetermined sampling sites targeting a blood vessel of interest in the subject in synchronization with a contrast agent (also referred to as a tracer) administered to the subject. The radiation detector, often configured as a panel of rotating detectors, receives radiation that traverses the subject at the predetermined sampling site(s) providing projection data (also referred to as scan data) over a time range that encompasses the increase phase and also optionally the decrease phase of contrast agent flowing through the blood vessel of interest.

[0092] The imaging system 2 includes a data acquisition component 6 incorporating a data acquisition scheme or data acquisition computer code that receives, organizes and stores projection data from the radiation detector of the CT scanner. The projection data is sent to an image reconstruction component 8 incorporating an image reconstruction computer code. The projection data can then be processed using the image reconstruction computer code resulting in image data including multiple images of the predetermined sampling site(s) spanning the increase phase and 02 OCTOBER 2025 (02.10.2025) also optionally the decrease phase of contrast agent flowing through the blood vessel of interest. The image reconstruction computer code can easily be varied to accommodate any available CT imaging technique. The image data can then be processed by an image analysis component 10 incorporating image analysis computer code that generates a time-enhancement curve of the contrast signal from the image data. The time-enhancement curve data can then be processed by a blood flow estimation component 12 incorporating a blood flow estimation computer code to determine a blood flow characteristic of the blood vessel of interest from the time-enhancement curve data. The imaging system 2 is controlled by a computer 16 with data and operational commands communicated through bus 14. The imaging system 2 may include any additional component as desired to assess a blood vessel of interest including multiplexers, digital / analog conversion boards, microcontrollers, physical computer interface devices, input / output devices, display devices, data storage devices and the like. The imaging system 2 may include controllers dedicated to different components of the CT scanner 4, such as a radiation source controller to provide power and timing signals to control the radiation source, a gantry controller to provide power and timing signals to a gantry motor to control rotation of the gantry and thereby control rotation of the radiation source and detector, and a table controller to provide power and timing signals to a table motor to control table position and thereby control position of a subject in the gantry by moving the subject along a z-axis through an opening of the gantry communicative with the interior open chamber of the gantry. The imaging system 2 is shown with a CT scanner as an illustrative example only, and the system may be modified to include other imaging modalities, including for example, non-CT X-ray imaging or MRI.

[0093] Figure 2 shows a computer implemented method 20 for blood flow imaging. The method 20 comprises a pre-scan preparation 30 and positioning of a subject for CT scanning of a desired sampling site. Once the subject is prepared and positioned within a CT scanner, the subject is injected 40 with a contrast agent solution, with CT scanning 50 synchronized with the injection of the contrast agent solution to acquire projection data (also referred to as scan data) over a time range that includes flow of the contrast agent through a blood vessel at the sampling site. The projection data is processed to reconstruct 60 image data from the projection data. The image data is analyzed to predict 70 an area under a time-enhancement curve of a contrast signal parameter, such as contrast signal intensity, extracted from the image data. A blood flow value is calculated 80 based on the time-enhancement curve.

[0094] Figure 3 shows an example of a pre-scan preparation 30 of a subject for CT scanning. The pre-scan preparation 30 includes identifying a region of interest 32 in the subject. For example, the 02 OCTOBER 2025 (02.10.2025) region of interest may be a portion of a blood vessel targeted for assessment of blood flow in the blood vessel. Once a region of interest is established, sampling site(s) for CT scan slices are identified 34 at or near the region of interest. Based on the predetermined sampling site(s), the subject is positioned 36 in the CT scanner in an alignment that allows for a radiation source of the CT scanner to direct radiation at the sampling site(s). Prior to scanning, the subject optionally holds breath 38 and maintains a breath-hold throughout scanning. As a further option, a hyperemic condition can be induced in the subject, for example by administering a vasodilator to the subject.

[0095] Figure 4 shows an example of CT scanning 50 synchronized to injection of the contrast agent. The synchronized CT scanning 50 includes initiating a dynamic CT scan at a desired time based on an injection of the contrast agent. Optionally, the CT scanning can be synchronized to an electrocardiogram (ECG), such as provided by prospectively electrocardiogram (ECG) gated contrast-enhanced dynamic CT imaging. When ECG gating is deployed retrospective ECG-gating or prospective ECG-gating may be used. The dynamic CT scan includes acquiring of projection data prior to entry 54 of contrast agent at the sampling site(s) to set a baseline, as well as acquiring projection data during an increase phase 56 of the contrast agent at the sampling site(s) and acquiring projection data during a decline phase 58 of the contrast agent at the sampling site. An increase phase refers to an increase of mass of contrast agent at the sampling site as time advances subsequent to initial entry of the contrast agent into the sampling site, while a decline phase or decrease phase refers to a decrease of mass of contrast agent at the sampling site as time advances prior to substantially complete clearance of the contrast agent from the sampling site. Peak (maximum value) mass of contrast agent at the sampling site occurs during progression from the increase phase to the decline phase. Time elapsed from entry to clearance of contrast agent at the sampling site may be referred to as a transit time of the contrast agent. The duration of CT scanning is not limited by a requirement to capture a complete transit time of contrast agent at the sampling site provided that at least a portion of both increase and decrease phases are captured.

[0096] Figure 5 shows an example of image analysis to predict 70 an area under a timeenhancement curve with a machine learning model. Prediction of an area under a timeenhancement curve can include identifying 72 a voxel or region of interest within a plurality of corresponding images at the sampling site. Contrast agent signal data is extracted 74, for example contrast agent signal intensity, from an area defined by the voxel or region of interest from each of the plurality of corresponding images. An area under a time-enhancement curve is predicted by a computer-mplemented learning model 74 based on the contrast agent signal data during the 02 OCTOBER 2025 (02.10.2025) increase phase, the decrease phase, or both the increase phase and the decrease phase at the sampling site.

[0097] Figure 6A shows an example of estimating blood flow 80 in a voxel or region of interest based on the time-enhancement curve. The estimation of blood flow can be achieved by determining a flow velocity value based on the time-enhancement curve. The determination of a flow velocity value can include selecting first and second time points 82 from the timeenhancement curve. A rate of change of enhancement (AHU / At) 84 can be determined by subtraction of the CT image data at the selected first and second time points. Calculating a rate of change of tracer mass (dm / dt) 85 based on the rate of change of enhancement (AHU / At) and the fractional volume of tracer solution that passes through the target voxel per unit time, for example using Equation 17 presented in co-owned PCT / CA2021 / 051189 filed 26 August 2021. The tracer density can be determined 86 from the predicted area under the time-enhancement curve from step 76, for example using Equations 13, 14 and 15 presented in co-owned PCT / CA2021 / 051189 filed 26 August 2021. Flow velocity through the target voxel is determined 88 based on the rate of change of tracer mass (dm / dt) and tracer density (p), for example using Equation 1 IB derived from Reynolds Transport Theorem presented in co-owned PCT / CA2021 / 051189 filed 26 August 2021. The determined flow velocity value can be communicated or displayed to a technician / operator or other end-user through any conventional computer or display device.

[0098] Equation (1 IB) states that the magnitude of the net flow velocity of the fluid in a given image voxel (control volume) can be estimated if the time rate of change of the mass of fluid in the control volume and the density of the fluid are known. Both pieces of information can be obtained from dynamic contrast-enhanced CT imaging. A depends on the area of the selected ROI, and therefore is it an independent variable that can be selected / controlled by the operator.

[0099] The p in Equation (1 IB) refers to the density of iodine in a mixture of blood and contrast solution. To estimate the density of iodine in a control volume (image voxel), we need to know: i) the total mass of contrast agent injected into the patient body, and, ii) the total volume of blood mixed with the contrast solution.

[0100] The total mass of contrast agent injected into the patient body is given by the following equation: 02 OCTOBER 2025 (02.10.2025) where Mi is the mass of iodine in unit of milligram (mg), Co is the original concentration of iodine-based contrast solution in unit of mg per millilitre (mg / mL), D is the contrast dilution factor ranges from 0 to 1, and Vt is the total injected contrast volume in unit of millilitre (mL).

[0101] The volume of blood mixed with the contrast solution can be estimated using the area under the time-enhancement curve in a two-step process: where Q is the volumetric flow rate in unit of millilitre per second (or litre per minute), Mi is the total mass of iodine injected determined from equation (12), the integral of C(t) is the area under the time-enhancement curve sampled at the control volume and has a unit of (HU x second), which can be converted to the unit of (mg / mL x second) with the conversion factors that were previously determined from our phantom experiments (So A et al, Medical Physics 2016;43(8):4821).

[0102] As the volumetric flow rate describes the volume of blood passes through per unit time, the volume of blood mixed with iodine, Vi, can be determined by the following equation: where Tenis the duration of time that the signal intensity in the control volume (image voxel) is higher than the baseline level, which can be determined graphically from the measured timeenhancement curve. The density of contrast agent in the control volume (image voxel) can be estimated using the following equation:

[0103] The time rate of change of the tracer mass in a control volume (image voxel) can be determined from the following equation: where dc / dt is the change in tracer concentration in the image voxel per unit time, Vf is the fractional volume of tracer solution that passes through the image voxel during this period of time. Equation (16) can be further expressed in this form: 02 OCTOBER 2025 (02.10.2025) where AHU is the difference in CT number in the image voxel between the two selected time points; At is the difference in time between the two selected time points; d is the factor for converting the Hounsfield Unit (CT number) to tracer (iodine) concentration (see So et al, Medical Physics 2016); Vt, as defined in Equation (12), is the total volume of tracer solution injected into the patient; Ten, as defined in Equation (14), is the duration of time that the signal intensity in the image voxel is higher than baseline. The terms within the first bracket on the right side of Equation (17) gives dc / dt. The terms within the second bracket on the right side of Equation (17) is equivalent to the fraction of the total volume of tracer solution involved during the selected time duration.

[0104] Figure 6B shows another example (alternative to the example shown in Figure 6A) of estimating blood flow 80a in a blood vessel of interest based on the time-enhancement curve. The estimation of blood flow 80a can be achieved by determining a fractional flow reserve (FFR) value based on the time-enhancement curve. The determination of an FFR value can include input 92 of the predicted area under the time-enhancement curve (AUC) from step 76. A flow rate 94 can be determined based on the predicted area under the time-enhancement curve, for example using an indicator-dilution principle as expressed in Equation 1 presented in co-owned PCT / CA2019 / 050668 filed 16 May 2019. A flow velocity can be determined 95 based on the flow rate and a calculated cross-section area of a lumen of the blood vessel at the sampling site, for example using Equation 2 presented in co-owned PCT / CA2019 / 050668 filed 16 May 2019. A flow pressure 96 can be determined from the flow velocity, for example using Bernoulli’s equation as expressed in Equations 3A or 3B presented in co-owned PCT / CA2019 / 050668 filed 16 May 2019. Based on flow pressure 96 determined from at least two sampling sites a pressure gradient can be calculated, and an FFR value can be determined based on the calculated pressure gradient and a systolic blood pressure value, for example using Equation 11 presented in co-owned PCT / CA2019 / 050668 filed 16 May 2019. The determined FFR value can be communicated or displayed to a technician / operator or other end-user through any conventional computer or display device.

[0105] The blood flow imaging system and method have been mathematically validated. Mathematical analysis described in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in coowned PCT / CA2021 / 051189 fded 26 August 2021 show numerous examples of deriving blood 02 OCTOBER 2025 (02.10.2025) flow characteristics from a dynamic contrast-enhanced dynamic CT imaging session that can all be incorporated within the current disclosure.

[0106] Fluid motion can typically be assessed by two approaches. The first approach is by monitoring the movement of individual particles in the fluid over time (Figure 7A). Navier-Stokes Equations (derived from Newton’s Second Law) can be used to describe the movement of any individual particle in the fluid in any direction. The second approach is by monitoring the passage of a small fraction of fluid within a fixed region (volume) over time (Figure 7B). The monitoring region or frame of reference (rectangle outlined with dark dashed line in Fig. 7B) does not move over time. The small fraction of fluid contains many individual particles that cannot be resolved, but this second approach is computationally less intensive compared to the first approach, and the results give a good approximation of the fluid movement with a reasonably good spatial resolution. Without wishing to be bound by theory, the second approach provides the basis of the analytic imaging method disclosed herein.

[0107] In CT, an image voxel, or a block of image voxels, is selected as the fixed region to monitor the movement of fluid (e.g. blood) over time. This frame of reference is called a control volume and the surface on each side of the control volume is called a control surface (as illustrated in Figure 8A). Fluid can move in and out of the control surface in any direction.

[0108] For illustration, an example shown in Figure 8B depicts movement of a small amount of fluid (grey region outlined with a dashed grey line) with respect to a control volume (CV). Here, the CV is shown in a two-dimensional view. At time 0, the fluid of interest (the grey region marked with mid-grey stripes) fills the CV completely. At a later time (time O+At), the same fluid of interest starts to move out of the CV in the flow direction shown in Figure 8B. The dark-grey stripes denote the portion (mass) of fluid that leaves the CV at time O+At and the vacant space within the CV is filled by the incoming fluid (marked with light-grey stripes). The mid-grey stripes represent the portion of fluid that remains in the CV at time O+At. If the mass of incoming fluid that fills the vacant space of CV equals the mass of fluid that leaves the CV, then the flow condition is considered as steady. If the two masses (light-grey and dark-grey portions) are not equal to each other, then the flow is considered as unsteady. It should be noted that unsteady implies the flow is either turbulent, or in the transition between the steady and turbulent states. At any time, the total mass of the fluid of interest (the fluid region delineated by the dashed grey line in Fig. 8B) is unchanged. That is, the mass in the area marked with the mid-grey stripes at time 0 should equal the total mass in the area marked with the mid-grey and dark-grey stripes at time O+At. The movement of fluid with respect to a CV can be described using the Reynolds Transport Theorem. 02 OCTOBER 2025 (02.10.2025)

[0109] Figures 9A and 9B shows schematic illustration for two different approaches of dynamic image acquisition and reconstruction that is a basis for absolute or relative flow velocity assessment. Figure 9A shows image acquisition and reconstruction at the same cardiac phase (e.g. 75% R-R interval or diastole) over multiple time points. Figure 9B shows image acquisition covering a full cardiac cycle (systole and diastole); multiple image sets corresponding to different cardiac phases are reconstructed.

[0110] The blood flow imaging technology described herein uses supervised machine learning (ML) or supervised deep learning (DL) algorithm to simplify the clinical workflow for functional assessment in vascular diseases with CT. Herein, we use the cardiac application for illustration but the blood low imaging technology can be applied to other vascular diseases as well.

[0111] In a routine cardiac CT test, coronary CT angiography (CCTA) is acquired first to determine if a patient has stenosis in a coronary artery. Prior to the CCTA scan, a bolus tracking (BT) scan is initiated to determine the optimal acquisition time for the CCTA scan. A BT scan is a real-time tracking technique in which the signal intensity or enhancement in a region of interest (ROI) is monitored at consecutive time points following an intravenous bolus injection of contrast solution into the patient. When the signal intensity in that monitoring ROI reaches a pre-defined threshold value, the CCTA scan is automatically executed at a delayed time (approximately 5 to 7 seconds) to ensure the CCTA scan captures the peak or near peak contrast enhancement in the coronary arteries. Alternatively, a non real-time tracking timing bolus (TB) scan can be used instead of a real-time tracking BT scan for estimating the peak enhancement time in the artery of interest. A TB scan is similar to a dynamic perfusion scan, with the exception that the volume and iodine concentration of contrast solution injected for the TB scan are usually less compared to those used in a CCTA scan or perfusion scan. Furthermore, a TB scan is set up as an independent series from a CCTA scan. While a BT scan is used for illustration of an ML / DL implementation, it will be understood that the same concept can be easily extended with a TB scan used instead of a BT scan.

[0112] In a typical clinical workflow, if there is absence of stenosis in the CCTA images, no further functional assessment is required. If the CCTA scan confirms the presence of stenosis, particularly if the degree of the stenosis is between 40 to 90% in lumen diameter, then a dynamic perfusion scan can be acquired during maximal vasodilation for functional evaluation of the stenosis (i.e. to determine if the stenosis is flow-limiting). In some situations, a dynamic perfusion 02 OCTOBER 2025 (02.10.2025) scan is also acquired at rest to determine the magnitude of blood flow increase from baseline for a more accurate functional assessment.

[0113] Thus, as summarized in Figure 10, in examples of clinical workflow progression of a cardiac CT test, the BT and CCTA scans are typically performed in a cardiac CT test while the perfusion scans are optional (depending on the CCTA findings). Both the BT and CCTA scans are performed in the resting condition.

[0114] Figure 11 shows a schematic of an example of blood flow imaging incorporating a machine learning model:

[0115] 1. combine the BT scan and CCTA scan as a pseudo truncated dynamic perfusion scan.

[0116] 2. track the temporal changes of enhancement in each coronary artery from the pseudo truncated dynamic perfusion scan.

[0117] 3. use ML or DL to predict the area under the time-enhancement curve in a coronary artery during the whole first-pass phase based on the findings in (2).

[0118] 4. use ML or DL to predict the area under the time-enhancement curve in a coronary artery during maximal vasodilation (hyperemia) from the AUC predicted in (3).

[0119] 5. use the AUC predicted in (4) coupled with the analytic algorithms disclosed in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021 to derive blood flow variables such as fractional flow reserve (FFR).

[0120] A similar approach can be applied in other scenarios. More specifically, one example is the application of ML or DL to predict the area under a full coronary time-enhancement curve from a genuine truncated dynamic rest perfusion scan covering 5 or more time points (Figure 12). Another example is the application of ML or DL to predict the area under a full coronary time-enhancement curve from a genuine truncated dynamic stress perfusion scan covering 5 or more time points (Figure 13). Moreover, ML or DL can also be applied to predict the area under a time-enhancement curve at one location in a coronary artery in a short-axis plane from the time-enhancement curve sampled at the same location in a coronary artery in the axial plane.

[0121] Both ML and DL are artificial intelligence techniques that teach computers to learn from training dataset in a way similar to how humans learn from past experiences. Furthermore, supervised ML and DL refer to the use of labeled training dataset for computer learning. A difference between ML and DL is the number of hidden layers and neurons used to achieve the learning goal. A DL architecture usually has more layers to handle more difficult tasks (such as 02 OCTOBER 2025 (02.10.2025) object segmentation). Contrasting ML and DL is for convenience of discussing the segmentation task as DL is advantageous for the segmentation task, while both ML and DL can perform the AUC prediction task.

[0122] DL is subset of ML, and therefore the term ML encompasses DL, and where ML and DL is contrasted it is intended to contrast non-deep learning to DL approaches, and not to suggest that ML and DL are mutually exclusive categories of computer learning models.

[0123] Machine Learning Approach. In an example of a ML approach, the computer is trained to optimize a linear regression model that predicts the area under a full coronary time-enhancement curve by learning from some labeled images provided to the computer. The training image set comes from either a pseudo or genuine truncated dynamic perfusion scan. The training images are assigned with labels corresponding to the AUC values of full coronary time-enhancement curves. The labels (ground truth) are derived from a genuine full dynamic perfusion scan acquired from the same patient. The features where the computer can learn from can be image based or non-image based or both (Figure 14).

[0124] Image-based features are features that can be observed or measured from image data and may include but are not limited to the following variables: enhancement (pixel signal intensity) in coronary arterial lumen, thickness of myocardial wall surrounding the left ventricle, size of left atrium, size of left ventricle, diameter of coronary artery, morphology of coronary artery, degree of stenosis in coronary artery, etc.

[0125] Non-image based features that are not measured from the image data and may include but are not limited to the following variables: patient’s age, sex, weight, heart rate, blood pressure, x- ray tube voltage, x-ray tube current, disease state, contrast-injection rate, iodine concentration in contrast solution, etc.

[0126] Figure 15 and Figure 16 provide schematic illustrations on how different image-based features can be used to train a ML algorithm to make predictions about the AUC values.

[0127] In the first illustration, the coronary enhancement at different time points can be combined into a single image-based feature with one of the following ways (Figure 15):

[0128] 1. The difference in Hounsfield Unit (HU, unit of CT image pixel value) between any two time points in the non-baseline (enhanced) phase.

[0129] 2. The difference in HU between any two time points in the non-baseline phase divided by the time difference between the two selected time points.

[0130] 3. The area under the truncated time-enhancement curve between any two time points in the non-baseline phase. 02 OCTOBER 2025 (02.10.2025)

[0131] 4. Any of the #1 to #3 with unit conversion from HU to iodine concentration (in milligram per mL).

[0132] In the second illustration, the myocardial wall thickness in different regions can be used individually or collectively as input feature(s) for training the ML algorithm (Figure 16). Furthermore, as illustrated in Figures 15 and 16, image and non-image features can be used conjunctionally as the input features for the training. The computer will determine which of the input features have the highest correlation with the labelled feature (AUC values) and keep them in the regression model for future prediction on the AUC values.

[0133] Deep Learning Approach. As illustrated in Figure 17, a basic artificial neural network consists of three layers: an input layer, a hidden layer and an output layer. A deep artificial neural network is a network with multiple hidden layers. The input layer is where the input values are accepted by the network, and the output layer is where the final prediction is made. In between, the hidden layers take the input values and perform some computations based on the input variables. The results generated by the hidden layers are propagated to the output layer where the final prediction is made (i.e. regression for the current example).

[0134] In one example, the hidden layers consist of: multiple convolution layers, multiple pooling layers, a flattened layer, multiple dense layers (or fully connected layers).

[0135] A CT image has 512x512 = 262144 pixels. For any greyscale image like a CT image, every pixel in the image has a value ranging from 0 to 255. Conventionally, a darker pixel has a lower value and a brighter pixel has a higher value. A CT image is not directly passed to the dense layers. Instead, a mathematical operation called convolution is used to extract only the desired information from the image as the inputs for the dense layers. During the convolution process, multiple filters are overlayed to a two-dimensional CT image and sliced through the image (from top left to bottom right). Each filter is a mathematical matrix with a small dimension such as 3x3 or 5x5 pixels. The extracted image information from the convolution with each filter is passed through an activation function and the results are stored in a new matrix with a reduced dimension (also called a feature map). Afterwards, a pooling filter (usual dimension is 2x2 pixels) is applied to the feature map for reducing the dimension of the feature map while preserving the most distinguished information. A pooled feature map is generated as a result. This image processing sequence (convolution -> activation -> pooling) can be repeated for several times and the dimension of the feature and pooled feature maps are reduced after each iteration. After the final iteration, the pooled maps are converted to a one-dimensional vector (flattened layer) as the inputs for the dense layers. 02 OCTOBER 2025 (02.10.2025)

[0136] Each convolution filter is designed for a specific pattern recognition, and as such, many filters are needed for more complicated image patterns. The number of feature map generated after a convolution is the same as the number of convolution filter applied (Figure 18).

[0137] Each dense layer has many interconnected nodes called neurons. At each neuron, each input variable is multiplied by a weight (a bias may also be added) before undergoing a non-linear transformation by passing through an activation function. The purpose of this computation is to determine whether a specific input variable should be passed to the next layer.

[0138] A dynamic image set can be handled by a DL algorithm in the following way: First, generate multiple inputs with a convolution stack for each image corresponding to a specific time point. Second, apply concatenation to the flattened layer corresponding to each time point to aggregate the temporal information from all the images in a dynamic series. Last, the concatenation flattened layer provides the inputs to the first dense layer (Figure 19).

[0139] In summary, both ML and DL methods can predict the AUC of a truncated timeenhancement curve from the image features presented in the dynamic perfusion images (non-image features may also be used), but can differ in input features. When a trained ML algorithm is used, the user will need to provide the input features with aids of image processing tools in the software. For instance, place a sampling ROI in the pseudo or genuine dynamic image set to obtain a truncated time-enhancement curve, or place a measuring tool across the myocardium to measure the wall thickness. The ML algorithm will then take the input features and make prediction on the AUC values, from which blood flow variables such as FFR can be computed with the analytic steps disclosed in our previous technologies. On the other hand, when a trained DL algorithm is used, the user only needs to load the dynamic image set and the algorithm will automatically identify the relevant input features and predict the AUC values and subsequently the blood flow variables.

[0140] The blood flow imaging system and method have been validated by experimental testing. Experimental testing results demonstrate the ability of the blood flow imaging system and method to determine one or more of several blood flow characteristics. The following experimental examples are for illustration purposes only and are not intended to be a limiting description.

[0141] Experimental Exemplification: Experimental Example 1 (Relationship between AUC and degree of lumen narrowing).

[0142] As explained in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021, the volumetric flow rate, F (shown as Q in Equation 1), in a blood vessel can be estimated with the indicator-dilution principle: 02 OCTOBER 2025 (02.10.2025) n =Mif C(t)dt(1)where m in Eq. (1) is the mass of tracer in the blood vessel, the integral in the denominator in the equation is the area under the time -enhancement curve (AUC). The integral in the denominator is the integral of tracer (interchangeably referred to as contrast agent) concentration as a function of time at a region of interest, and therefore AUC represents a total sum of tracer concentration time product at the region of interest, and a calculated AUC can readily be converted to the tracer concentration time product, optionally using a conversion factor correlated to the scanner energy for greater accuracy of unit conversion. Since F and AUC are inversely proportional to each other, a smaller AUC reflects a higher volumetric flow rate. Our data shows a clear inverse non-linear relationship between the degree of stenosis in a coronary artery and the AUC of a coronary time-enhancement curve (Figure 20).

[0143] In a stenosed coronary artery, the magnitude of increase in volumetric flow rate from baseline (rest condition) is attenuated in the post-stenotic coronary segment if the stenosis exceeds a certain degree (-40% luminal narrowing), and it is reflected by a larger ratio of the stress AUC to rest AUC in this segment compared to the pre-stenotic segment (Figure 21). The graphs shown in Figure 20 and Figure 21 collectively indicate that the AUC of a coronary time-enhancement curve is predictable if other information such as the degree of coronary stenosis is known, which supports the notion of using ML or DL algorithms to learn and predict the AUC values.

[0144] Experimental Exemplification: Experimental Example 2 (Relationship between timeenhancement curve and heart rate).

[0145] Although the symmetry of a coronary time-enhancement curve following a bolus intravenous injection of contrast media is relatively unaffected by the patient’s heart rate, the width of the curve may be partially dependent on the patient’s heart rate (Figure 22). This indicates that non-image feature such as patient’s heart rate may also be a useful input variable for the ML or DL algorithms to make prediction on AUC.

[0146] Experimental Exemplification: Experimental Example 3 (Shape of time-enhancement curve at different vascular sizes).

[0147] Figure 23 shows that the fitted time-enhancement curve obtained from the myocardium (after removing the contrast retention and recirculation effects) has a similar symmetry compared to the fitted time-enhancement curve obtained from the feeding epicardial coronary artery. This finding indicates that the proposed ML and DL methods can also be applied to estimate the first- 02 OCTOBER 2025 (02.10.2025) pass time-enhancement curve at the microvascular level with either a pseudo or genuine truncated dynamic perfusion image set.

[0148] Experimental Exemplification: Experimental Example 4 (Relationship between myocardial wall thickness and coronary diameter and AUG).

[0149] As discussed in Experimental Example 1, the AUC is different between the rest and stress (hyperemic) conditions. During maximal hyperemia arising from intravenous adenosine administration, the patient’s cardiac contractility and coronary artery diameter may increase substantially compared to the baseline levels. Therefore, image features such as myocardial wall thickness and coronary diameter (or radius or circumference) may be used as the input variables for the ML or DL algorithm to predict AUC (Figure 24).

[0150] Experimental Exemplification: Experimental Example 4 (Time-enhancement curves in different tomographic views).

[0151] As explained in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021, the pressure gradient between two sampling slices A and B in a blood vessel can be estimated with the Bernoulli’s equation: where PA and PB are the coronary flow pressure (in Pascal or Pa) in slice A and B respectively, p is the density of blood (g / cm3), g is the Earth’s gravity (980 cm / s2), hAand hBare the relative height (in cm) above or below a reference plane from the center point in slice A and B, V and VB are the flow velocities at slice A and B respectively, and PL is the pressure (energy) loss due to friction and / or turbulence. VA and VB should be calculated from the time-enhancement curves sampled at the tomographic slices perpendicular to the direction of blood flow. Due to the nature of arterial curvature, the imaging slices are not always perpendicular to the direction of flow. In this case, the tomographic images should be reformatted into the desired view before sampling the time-enhancement curves for blood flow calculation (Figure 25).

[0152] The axial view is the default tomographic plane for CT image reconstruction. As discussed above, for quantitative blood flow assessment, short-axis reformation is necessary to obtain the time-enhancement curves at the slices perpendicular to the direction of blood flow. However, reformatting a full dynamic (4D) image set can be quite time consuming. The proposed ML or DL approach can be used to predict a short-axis time-enhancement curve based on the timeenhancement curve sampled from the axial plane without the need of image reformation. Figure 26 02 OCTOBER 2025 (02.10.2025) compares a short-axis time-enhancement curve with an axial time-enhancement curve for one coronary artery. It can be seen that the short-axis curve is slightly more fluctuated than the axial curve. However, the details of the curve shape are irrelevant to the FFR calculation and only the area under the curve matters. This reduces the complexity of the ML and DL learning and makes the proposed methods more feasible for clinical applications.

[0153] Experimental Exemplification: Experimental Example 5 (Training a machine learning model to predict the AUC of a time-enhancement curve during maximal hyperemia (stress) from the AUC of a time-enhancement curve sampled at rest).

[0154] This example illustrates the machine learning approach to predict the AUC of a timeenhancement curve during maximal hyperemia from the AUC of a time-enhancement curve sampled during the resting condition. While the proposed approach works for both coronary and myocardial time-enhancement curves as well as other cardiovasculatures of interest, an example using coronary time-enhancement curves is provided here. In this approach, the cardiac CT studies with dynamic rest and stress perfusion scans performed on the same patient on the same day were used for training a linear regression model in supervised machine learning. This approach is suitable for predicting continuous numerical values, such as the area under curve of a timeenhancement curve. The implementation of machine learning was achieved using Python and Tensor flow. Figure 27 shows the coronary time-enhancement curves sampled across the RCA (right coronary artery) stenosis in a patient during rest and maximal hyperemia (stress). The sampled coronary time-enhancement curves were denoised with a modified gamma variate function, and the AUC of the denoised (fitted) curve corresponding to the rest condition was used as the input (independent) variable to the machine learning regression. Other input (independent) variables include non-image data such as the patient’s age, heart rate and blood pressure measured during the rest condition.

[0155] Figure 28 shows the scatter plot of the AUC of the measured stress time-enhancement curve against the AUC of the corresponding measured rest time-enhancement curve sampled in different right coronary arteries with different degrees of stenosis. It can be seen that the AUCs of the rest and stress time-enhancement curves tend to increase with the degree of stenosis in the artery of interest albeit some overlapping between different groups of AUC data.

[0156] The AUC data shown in Figure 28 was divided into the training and test (validation) data sets (80% of the data was used for training) for machine learning. The computer codes shown in Figure 29 illustrates three hidden layers were implemented in the neural network and linear regression was used for this task. 02 OCTOBER 2025 (02.10.2025)

[0157] The training and test (validation) results are shown in Figure 30. The crosses in the graph are the AUCs of the stress coronary time-enhancement curves predicted by the trained neural network against the true AUCs of the stress coronary time-enhancement curves. The training and test scores were 0.904 and 0.837, respectively (the highest score is 1). The results suggest that the trained machine learning model can predict the stress AUC from the measured rest AUC and other non-image data measured at rest (e.g. heart rate, blood pressure) with a high accuracy. The accuracy of the machine learning model can be improved further by training with more image and non-image data and optimizing some of the training settings.

[0158] The flow diagram in Figure 31 illustrates the machine learning scheme performed in this example. The flow diagram in Figure 32 illustrates how the trained machine learning model can be used for clinical hemodynamic assessment.

[0159] Experimental Exemplification: Experimental Example 6 (Training a machine learning model to predict the AUC of a full time-enhancement curve from the AUC of a truncated time- enhancement curve).

[0160] This example illustrates a machine learning approach to predict the AUC of a full coronary time-enhancement curve from a truncated coronary time-enhancement curve sampled from a simulated truncated rest perfusion scan. As mentioned before, a full curve is defined as the one covering the baseline and the entire upslope and downslope. Prior to the training, a pseudo truncated perfusion scan was first simulated by removing the images after the peak enhancement in the dynamic perfusion images (Figure 33a). The AUC of the truncated curve was then computed (Figure 33b). This AUC with that of the full (untruncated) curve were used as a pair of input variables to train a machine learning linear regression model to understand the relationship between the two AUC values and make correction prediction for unseen cases.

[0161] The curve shown in Figure 33 is a simulation of a pseudo truncated time-enhancement curve synthesized by combining the image data from a bolus tracking (BT) scan with the image data from the subsequent diagnostic CT angiography (CTA) scan. The simulation is based on a realistic clinical scenario where the diagnostic CTA scan is triggered approximately 4 to 7 seconds after the trigger threshold in the monitor region is reached (the trigger threshold in the selected monitor region is set to 50 HU in this simulation and the trigger time gap is specific to each scanner and is a function of time to reconfigure the scanner for BT scan settings to diagnostic CTA scan settings). Therefore, the truncated time-enhancement curve has a larger time gap between the 3rd and the 4th data points. Besides perfusion scan, the proposed machine learning approach will work for any type of scan, including test bolus (TB) scan, bolus tracking (BT) scan, and diagnostic 02 OCTOBER 2025 (02.10.2025)

[0162] CTA scan. Although the contrast volume and X-ray energy used for each scan may be different (Table 1), the computation of AUC should not be affected if the conversion from image pixel enhancement to tracer (iodine) concentration can be performed using a conversion factor specific to the X-ray energy used for the scan.

[0163] Table 1. Comparison in contrast volume and X-ray energy used between different CT scans (for cardiac).

[0164] Figure 34 shows the scatter plot of the AUC of measured full time-enhancement curve against the AUC of related measured truncated time-enhancement curve sampled from different coronary arteries with different degrees of stenosis. The AUC data was divided into the training and test datasets for machine learning (approximately 80% was used for training). Similar to Experimental Example 5, a linear regression analysis was used in machine learning for this task. The preliminary training results (Figure 35) showed a high train / test score and a low train / test score, suggesting the trained machine learning regression model can be used to predict the AUC of a full coronary time-enhancement curve even if the time-enhancement curve is not acquired completely to cover the entire upslope and downslope.

[0165] The flow diagram in Figure 36 illustrates the machine learning scheme discussed in this example. The flow diagram in Figure 37 illustrates how the trained machine learning can be used for clinical hemodynamic assessment.

[0166] Experimental Exemplification: Experimental Example 7 (Training a machine learning model to predict the AUC of a time-enhancement curve in the short-axis tomographic view from the AUC of a time-enhancement curve in the default axial tomographic view).

[0167] As illustrated in Figure 26, coronary time-enhancement curves can be sampled from dynamic cardiac perfusion images at different tomographic orientations. In CT, the default image view is axial, but a perfect cross-sectional view of the arterial lumen may not be seen in a given axial slice (ie., an axial slice may have a non-orthogonal orientation relative to a central longitudinal axis of a blood vessel or relative to direction of blood flow). Currently, reformatting the axial images into the short-axis orientation (ie., a slice having an orthogonal orientation relative to a central longitudinal axis of a blood vessel or direction of blood flow) requires a specialized 02 OCTOBER 2025 (02.10.2025) software to perform volumetric image reformation. Figure 38 shows the preliminary results of using a similar machine learning approach to those described in Experimental Examples 5 and 6 to predict the AUC of a coronary time-enhancement curve in a short-axis tomographic slice from the AUC of a coronary time-enhancement curve sampled in the corresponding axial slice.

[0168] Experimental Exemplification: Experimental Example 8 (Training a deep neural network for segmenting the coronary arteries and myocardium in contrast-enhanced cardiac CT images).

[0169] In contrast to the approaches discussed in Experimental Examples 5 and 6, this example illustrates a deep learning approach for segmenting the coronary arteries and myocardium in contrast-enhanced cardiac CT images. Such segmentation facilitates relevant image features to be extracted to generate input variables for the subsequent machine learning. In other words, this approach is a combination of deep learning and machine learning. Figure 39A and 39B show the labeling of different image features, including the coronary arteries, ascending aorta and left ventricular myocardium, with different colours using the MATLAB’s Medical Image Labeler. These image labels were used to train a deep neural network called U-NET for segmenting these features in unseen contrast-enhanced cardiac CT images. The computer codes shown in Figure 39C demonstrate the implementation of a three-dimensional U-NET using MATLAB. The U-NET is based on a convolution neural network (CNN) mentioned in previous sections but with some modifications that can yield a more precise segmentation than CNN. The main difference between the two neural networks is that in CNN, an image is downsampled and converted into a vector which is used for object classification. On the contrary, in U-NET, an image is downsampled to a pixel level followed by upsampling. The additional sampling step allows U-NET to achieve a more precise detection of the structure details in the image. In other words, the CNN approach can provide information about what an object in the image is, whereas the U-NET approach can also provide this information plus the detailed spatial information about the object (where exactly the object is located in the image).

[0170] The segmentation task requires an excellent spatial resolution to accomplish, because the cardiovasculature of interest (such as a coronary artery) may not stay in the exact same location during a time-series of scan images (such as obtained in a dynamic perfusion scan) due to the patient’s residual respiratory and cardiac motion during image acquisition. Therefore, a deep neural network that can better preserve the spatial information of the images (i.e. U-NET) is more suitable than a standard convolution neural network (CNN) that is less capable of preserving the spatial information of the images for our task. Examples of medical image segmentation with U-NET and CNN have been previously described (Ronneberger, O., Fischer, P., Brox, T. (2015). U-Net: 02 OCTOBER 2025 (02.10.2025)

[0171] Convolutional Networks for Biomedical Image Segmentation. In: Navab, N., Homegger, J., Wells, W., Frangi, A. (eds) Medical Image Computing and Computer- Assisted Intervention - MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, Cham) (Krizhevsky A, Sutskever I, Hinton GE (2012) ImageNet classification with deep convolutional neural networks. Adv Neural Inf Process Syst 25).

[0172] Due to the large number of CT images in a dynamic perfusion scan covering the whole heart (e.g. 14 cm axial scan coverage with 1 mm image slice thickness) and large number of pixels in each CT image (512x512), Image segmentation with U-NET requires a powerful computer to achieve. For example, the minimum requirement for RAM is typically 64 GB and the graphic card can be a NVIDIA RTX 3060 Ti (8 GB GDDR6). However, the preferred hardware requirement is usually higher to ensure an efficient deep learning. For example, a NVIDIA RTX A5000 graphic card (24 GB GDDR6) has been used by some researchers for deep learning. Therefore, a standard laptop or desktop computer built for non-heavy duties is likely not suitable for this task.

[0173] Approximately 60 and 20 contrast-enhanced cardiac CT image sets were used for training and testing the U-NET, respectively. Each CT image sets had labels for coronary artery (i.e. RCA, LAD, LCX and LM) and the background (any feature unrelated to the coronary arteries). We first focused on training the U-NET for segmenting the left main (LM) and the left circumflex (LCX), and the test accuracy for segmenting the LM and LCx was 70.5%. A higher test accuracy can be expected by using more complex image sets for training and adjusting the training settings such as the learning rate. Figure 40 illustrates how the training results from a deep learning model can be used as the input labels to train a machine learning model described in Experimental Examples 5-7.

[0174] As demonstrated above, a use of machine learning model can significantly streamline the clinical workflow associated with blood flow imaging by employing a hybrid computer-leaming- analytic approach. As explained in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in coowned PCT / CA2021 / 051189 filed 26 August 2021, blood flow variables are dependent on many other physiological and morphological variables. If a computer needs to learn the relationship between each individual blood flow variable and all of its dependent variables, it will require a very large amount of training / validation data, image feature extractions and labeling, and computation time. The approach described herein restricts the computer learning to a relevant and central parameter, such as the area under a time-enhancement curve (AUC), which is needed for the analytic steps to derive all the blood flow characteristics described in co-owned 02 OCTOBER 2025 (02.10.2025)

[0175] PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021.

[0176] Advantageous features of the currently disclosed hybrid computer-leaming-analytic blood flow imaging approach are numerous, including for example predicting a hyperemic timeenhancement curve from a rest time-enhancement curve, predicting a hyperemic time-enhancement curve at a microvascular level from a macrovascular rest timeenhancement curve, or predicting a time-enhancement curve at one location in a blood vessel in one orientation from a timeenhancement curve sampled at the same location in a blood vessel in another orientation.

[0177] The currently disclosed hybrid computer-leaming-analytic blood flow imaging approach can offer reliable functional assessment of vascular diseases with CT while simplifying the clinical workflow through:

[0178] 1. deriving blood flow characteristics without performing a genuine dynamic perfusion scan or performing a short or truncated dynamic perfusion scan.

[0179] 2. deriving blood flow characteristics without administering a vasodilator.

[0180] 3. deriving blood flow characteristics without reformatting a dynamic (four -dimensional) image set from one tomographic view to another.

[0181] Experimental Exemplification: Experimental Example 9 (Training a machine learning model to predict functionally significant stenosis).

[0182] This experimental example describes a dynamic angiographic imaging (DAI) to evaluate whether a narrowed artery is functionally significant. Supervised machine learning is incorporated into DAI to predict the functional significance of arterial narrowing directly from the patient’s dynamic image data, and fractional flow reserve (FFR) calculation is not a requirement for this functional assessment.

[0183] Although coronary artery disease and dynamic contrast-enhanced computed tomography (CT) are used for illustration in this experimental example, this supervised machine learning functional assessment can also be used in other cardiovascular diseases and / or dynamic contrast- enhanced magnetic resonance imaging (MRI).

[0184] Figure 41 shows a flow chart of a supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and real rest / stress image data. More specifically, in this example, supervised machine learning is applied to determine the functional significance of coronary artery disease (CAD) using dynamic contrast-enhanced images acquired at rest or stress without computation of fractional flow reserve (FFR). FFR is defined as the coronary pressure distal to a stenosis to the coronary pressure proximal to a stenosis 02 OCTOBER 2025 (02.10.2025) during maximal hyperemia. A functionally significant CAD is defined as the physiological condition where the FFR value in a narrowed coronary artery is less than 0.80 or ischemia (reduced blood flow) is present in the downstream myocardium.

[0185] A supervised machine learning algorithm (ML algorithm 1) is developed and trained to determine whether a coronary artery stenosis is functionally significant based on at least one nonimage data (patient’s sex, age, height, weight, body-mass index, heart rate, blood pressure, symptoms, comorbidities; contrast concentration, contrast volume, contrast injection rate; X-ray energy, X-ray tube current, MR pulse sequence) and at least one image data acquired from dynamic angiographic imaging (Figure 41). In dynamic angiographic imaging, a cardiovascular region of interest is scanned over multiple times after contrast administration, capturing at least a portion of both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest.

[0186] Image data includes one or more options in (a) to (j): a. Area under curve (AUC), peak enhancement, upslope, downslope, full-width half maximum (FWHM: width of a spectrum curve measured between those points on the y-axis which are half the maximum amplitude), volumetric flow rate or flow velocity deriving from the time-enhancement curves sampled at the following locations (Figure 47): i. Prelesion slice at least 1 cm away from upper border of stenosis ii. Postlesion slice at least 1 cm away from lower border of stenosis iii. Immediate slice adjacent (proximal) to upper border of stenosis (entry slice) iv. Immediate slice adjacent (distal) to lower border of stenosis (exit slice) b. Pressure loss (PLs) due to an abrupt change in the arterial lumen size deriving from the time-enhancement curves sampled at the following locations: i. Immediate slice adjacent (proximal) to upper border of stenosis (entry slice) ii. Immediate slice adjacent (distal) to lower border of stenosis (exit slice) c. Image data of the prelesion and postlesion slices in (a) can be fed to the neural network of a supervised machine learning algorithm as independent predictors (inputs) or a single combined predictor as follow (Figure 41): i. Post - Pre (difference between Post and Pre) o For example, AUCpost- AUCpre; FWHMpost- FWHMpre; etc. ii . Post / Pre (rati o of Post to Pre) o For example, AUCpost / AUCpre; FWHMpost / FWHMpre; etc. 02 OCTOBER 2025 (02.10.2025) d. Image data of the entry and exit slices in (a) and (b) can be fed to the neural network of a supervised machine learning algorithm as independent predictors (inputs) or a single combined predictor as follow (Figure 41): i. Exit - Entry (difference between Exit and Entry) i. For example, AUCexit - AUCentry; PLSexit - PLSentry; etc. ii . Exit / Entry (ratio of Exit to Entry) i. For example, AUCexit / AUCentry; PLSexit / PLSentry; etc. e. At least two different contrast enhancement levels (in Hounsfield Unit) obtained from a time-enhancement curve sampled at the prelesion, postlesion, entry or exit slice. One of these enhancement levels must be the peak enhancement level. These enhancement levels are fed to the neural network of a supervised machine learning algorithm as individual input predictors NOT a single combined predictor (such as slope or AUC). f. Degree of stenosis. g. Length of stenosis. h. Number of stenosis in a coronary artery. i. Location of stenosis in a coronary artery (i.e. proximal / mid / distal segment or branch) j. Left or right artery involved (i.e. whether a stenosis is located in a right or left coronary artery).

[0187] For ease of understanding the image data in (a)-(e), Figure 47 schematically illustrates a location of a prelesion slice, a postlesion slice, an entry slice and an exit slice relative to a stenosis within an artery. The variable “d” denotes the minimum distance between the prelesion slice and the upper border of the stenosis, and between the postlesion slice and the lower border of the stenosis. It is equal to 1 cm.

[0188] Figure 42 shows a flow chart of a dual supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and virtual stress image data. In situations where dynamic contrast-enhanced images are acquired at rest only, the image data described in (a) and (b) can be replaced by the virtual stress image data predicted by a pre-trained machine learning model based on the rest time-enhancement curves sampled with dynamic angiographic imaging.

[0189] Figure 43 shows a flow chart of a dual supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and virtual image data corresponding to a tomographic orientation different from the tomographic orientation where real image data is acquired. The image data described in (a) and (b) can be derived from a virtual 02 OCTOBER 2025 (02.10.2025) time-enhancement curve corresponding to any tomographic view predicted by a pre-trained machine learning model based on a genuine time-enhancement curve sampled in a default tomographic view.

[0190] Figure 44 shows a flow chart of a dual supervised machine learning algorithm that predicts the functional significance of arterial lumen narrowing based on non-image data and image data corresponding to virtual full (untruncated) time-enhancement curves. The image data in (a) and (b) can be derived from a virtual full time-enhancement curve predicted by a pre-trained machine learning model based on a genuine truncated time-enhancement curved.

[0191] Figure 45 shows a flow chart of a supervised machine learning algorithm that predicts the presence or absence of downstream microvascular dysfunction in the presence of functionally significant lumen narrowing in the upstream artery. In situations where functionally significant coronary artery disease is confirmed, machine learning is used to determine whether such functionally significant disease is contributed by the epicardial coronary artery or the downstream microcirculation or both. In this approach, a supervised machine learning algorithm is developed and trained based on at least one non-image data and more than one image data. The image data includes one or more of the following options: at least one time-enhancement curve in an epicardial coronary artery, at least one time-enhancement curve in the myocardium supplied by the same epicardial coronary artery, coronary flow rate (in mL per unit time), coronary flow velocity (in cm per unit time), coronary flow reserve (CFR), fractional flow reserve (FFR), myocardial blood flow (in mL / min / g), iodine concentraion in myocardium (in mg / mL).

[0192] Figure 46 shows a flow chart of a dual supervised machine learning algorithm that predicts which image set of a dynamic contrast-enhanced imaging sequence corresponds to peak arterial enhancement. The prediction is used to select the correct images for subsequent anatomical and functional assessments of arterial lumen narrowing. Coronary CT angiography (CCTA) images can be derived from the rest or stress dynamic contrast-enhanced images acquired with dynamic angiographic imaging without acquiring additional test bolus / timing bolus and CCTA scans. In this approach, a supervised machine learning algorithm (ML algorithm 2) is developed and trained based on the dynamic contrast-enhanced images with proper labels denoting which volumetric image set corresponds to the peak contrast enhancement in the coronary arteries. The trained machine learning model (ML algorithm 2) is then used to determine which volumetric image set in an unseen dynamic contrast-enhanced image set corresponds to the highest contrast enhancement in the coronary arteries. The image set identified by the trained machine learning model (ML algorithm 2) is duplicated and transferred to a new location for anatomical analysis of the coronary 02 OCTOBER 2025 (02.10.2025) arteries in a similar fashion to CCTA images. Next, another supervised machine learning algorithm (ML algorithm 3) is used to segment the coronary arteries from the selected peak enhanced image set and output coordinates to locate the coronary arteries in each image of the unseen dynamic contrast-enhanced image set. This machine learning algorithm (ML algorithm 3) is developed and trained based on the dynamic contrast-enhanced images with proper labels denoting the coronary arteries and background. Next, an appropriately sized search region of interest (such as a 10x10 or 20x20 pixel block) is placed automatically based on the coordinates of the coronary arteries in each slice of the coronary arteries at the time point corresponding to the peak enhancement. Within each search RO I, the pixel or block of pixels (such as a 2x2 pixel block) exhibiting the highest signal (in Hounsfield Unit) is selected for each time point to generate a time-enhancement curve. Timeenhancement curve is automatically sampled at multiple locations in a coronary artery without user intervention.

[0193] Finally, the time-enhancement curves sampled along each coronary artery automatically are fed to the supervised machine learning model mentioned above (ML algorithm 1) to predict the functional significance of coronary artery disease. hi summary, this approach uses three different supervised machine learning models to achieve a one-stop comprehensive anatomical and functional assessment of coronary artery disease (Figure 46). The total effective dose (in milli-Sievert or mSv) associated with this approach is no more than 10 mSv.

[0194] For ML algorithms 2 and 3, a narrow-field-of-view training approach can be employed to facilitate the accuracy of neural network training. In this approach, the dynamic contrast-enhanced images are truncated first so that the neural network training can be focused solely on the region of interest. For example, in the cardiac applications, only the heart and the nearby aorta are kept in the field-of-view, while the lungs and other thoracic regions are removed. This approach can minimize the impact of contrast solution circulation in other parts of the thorax on the machine learning segmentation of the heart and coronary arteries at the peak enhancement.

[0195] CCTA and DCE scans are traditionally performed as two separate scans. The CCTA scan is always performed first and the DCE scan (also referred to as a perfusion scan) is an optional scan. The DCE scan is needed only if the CCTA scan identifies moderate degree of narrowing in the coronary arteries. The CCTA scan also requires a precedent test bolus or bolus tracking scan to ensure the CCTA scan is performed at the peak contrast enhancement in the coronary arteries. This conventional approach has several limitations, such as additional radiation dose and additional contrast dose (if a test bolus scan is used). Furthermore, if the patient’s heart rate varies between 02 OCTOBER 2025 (02.10.2025) the test bolus and CCTA scans, the CCTA scan may not capture the peak enhancement phase, leading to suboptimal image quality. Moreover, this approach requires the CCTA images to be reviewed immediately by a Radiologist / Cardiologist while the patient remains on the scanner table to determine if the DCE scan is needed. It is clearly not ideal for patient throughput and clinicians’ packed schedules. The approach of acquiring DCE images for functional assessment has several advantages compared to the existing approach. First, it does not need a test bolus or bolus tracking scan for CCTA. All the contrast phases are captured in the DCE scan. The trained machine learning model will select the optimal contrast phase for CCTA. Second, there is no need for the CCTA images to be interpreted immediately, because the DCE scan can be analyzed later to obtain FFR and myocardial perfusion if needed (if CCTA identifies moderate coronary stenosis). Third, only one contrast injection is needed which is favourable to patients at risk of renal problems. The DCE scan comprises multiple low-dose volumetric scans covering approximately 20 seconds. The X-ray tube current setting (which controls the number of X-ray photons used per scan) for a DCE scan is usually much lower than that for a CCTA scan, leading to a much lower tube currentexposure time product per scan (Figure 61). The tube current-exposure time product is directly proportional to the radiation dose received by the patient. Still, anatomical assessment of the coronary arteries can be performed with the DCE images corresponding to the peak enhancement phase without significant image degradation compared to the standard CCTA images (Figure 57). If the patient shown in Figure 61 had a DCE scan covering 9 passes (time points), the total radiation dose of the DCE scan would be identical to that of the CCTA scan.

[0196] Figure 49 shows a screenshot image capturing part of the codes in a program written in Python to execute a supervised machine learning for binary classification of functionally significant coronary artery disease based on the image and non-image variables input to the neural network. Selected codes are labeled in the left margin and explanations for the selected codes are as follows: 1. TensorFlow (library for machine learning) was imported to program. 2. The file “XY.csv” contained the dataset for neural network training. 3. In this example, the prelesion peak and slope, and the postlesion peak and slope, in the XY.csv file (Figure 50) were not used for the training. 4. This neural network had 3 layers with 32 neurons in each layer. The final layer employs a sigmoid function to predict a binary outcome (functionally significance or non functionally significance). 5. Optimization algorithm (Adam) and loss function (binary cross entropy) used in neural network training. 6. Dataset was shuffled and divided into training and validation sets. In this example, 90% of samples were assigned to training and 10% were assigned to validation. 7. Logic of a training step was customed. 02 OCTOBER 2025 (02.10.2025)

[0197] Figure 50 shows a screenshot image capturing a dataset used to train a neural network for binary classification of functionally significant coronary artery disease using the Python program illustrated in Figure 49. The dataset was saved in a file called “XY.csv” (the code reading this file was shown in Figure 49). In this example, 6 cases were included in the training with 3 of them had functionally significant coronary artery disease (binary classification = 1). Catheter-based fractional flow reserve measurement was performed in all 6 cases to determine the functional significance of the stenosis as the ground truth. Each row in Figure 50 corresponds to the image and non-image variables of each patient. Five cases were used for training and one case was used for validation. Tables 2 and 3 together reproduce the details of the image in Figure 50 for greater clarity.

[0198] Table 2.

[0199] Table 3. 02 OCTOBER 2025 (02.10.2025)

[0200] Figure 51 shows two cases from the training dataset shown in Figure 50. One case (a, c, e; Fig 10 case #5) had a functionally significant narrowing in the left anterior descending artery. The other case (b, d, f; Fig 10 case #2) did not have a functionally significant narrowing in the left anterior descending artery. The time-enhancement curves sampled from the prelesion, postlesion, entry and exit slices were very different between the two cases. The approximate sampling location of the entry, exit and postlesion slices in each case were illustrated by the dotted (entry / exit) and solid (postlesion) arrows in (a) and (d).

[0201] Figure 52 shows the training score and loss corresponding to the inputs “entry AUC” and “exit AUC” returned by the program illustrated in Figure 9. The lesion entry and exit slices refer to the slices adjacent to the top and bottom of the lesion - they are closer to the lesion compared to the prelesion and postlesion slices. For example, the postlesion slice is usually located at 1 cm distal to the bottom of the lesion, according to the recommendation from clinical experts. To highlight the usefulness of the entry AUC and exit AUC on predicting the functional significance of stenosis, other image-based variables input to the neural network (such as "prelesion AUC" and "postlesion AUC" shown in Figure 50) were not used for training in this example (see the code in Figure 49).

[0202] Figure 53 shows the prediction of functional significant coronary artery disease made by the trained neural network (Figure 53b) agreed perfectly with the ground truth (Figure 53a). There are six circles or crosses in each graph. Each circle and cross in the graphs represent one case shown in Figure 50. Dark color in Figure 53a represents true positive cases (i.e. the patients had functionally significant coronary artery disease) and light color represents true negative cases (i.e. the patients had no functionally significant coronary artery disease). The graph in Figure 53b shows the predictions made by the trained neural network were correct.

[0203] Figure 54 shows a screenshot image capturing part of the codes in a program written in MATLAB to execute a supervised machine learning for segmenting the coronary arteries in contrast-enhanced CT images. Selected codes are labeled in the left margin and explanations for the selected codes are as follows: 1. Define location of image data. 2. Define location of labelled data. 3. Define dataset for training. 4. Define image size (number of pixels per slice and number of slices) and number of encoders in UNET. 5. Define UNET training settings. 6. Execute UNET training.

[0204] Figure 55 shows example of contrast-enhanced CT images used for supervised machine learning for coronary artery segmentation. Each image file is a single slice in the training dataset. 02 OCTOBER 2025 (02.10.2025)

[0205] Figure 56 demonstrates detection of the coronary arteries in a contrast-enhanced CT image with supervised machine learning, (a) Original CT image, (b) Original CT image with labels of the coronary arteries and background, (c) Prediction of the coronary arteries made by the trained neural network. The image scale in (b) and (c) was adjusted to facilitate the visualization of the labeled and predicted coronary arteries shown in these images.

[0206] Figure 57 compares visualization and lumen diameter measurement of the right coronary artery (RCA) with the standard CCTA image set (a, b, c) and the dynamic angiographic image set at peak contrast enhancement (d, e, f). The differences in visualization and lumen diameter are minimal between the two approaches.

[0207] Figure 58 compares (a) a patient with a functionally significant proximal-to-mid stenosis in an epicardial coronary artery without downstream microvascular dysfunction versus (b) a patient with a functionally non-significant proximal-to-mid stenosis in an epicardial coronary artery with downstream microvascular dysfunction.

[0208] Figure 59. (a to d) Prelesion and postlesion coronary time-enhancement curves sampled in the patient in Figure 58a. There was a substantial difference in the peak and AUC between the pre- and post- curves in this patient, (e to f) Prelesion and postlesion coronary time-enhancement curves sampled in the patient in Figure 58b. There were minimal differences between the pre- and postcurves in this patient.

[0209] Figure 60. (a and c) Myocardial time-enhancement curve sampled in the myocardium supplied by the artery shown in Figure 58a. (b and d) Myocardial time-enhancement curve sampled in the myocardium supplied by the artery shown in Figure 58b. (e) The peak and upslope of the two myocardial curves were noticeably different. Figures 58 to 60 collectively illustrate the differences of the coronary and myocardial time -enhancement curves can be used to differentiate CAD patients with and without microvascular dysfunction.

[0210] Figure 61. Comparison of the X-ray tube current (mA) and exposure time (s) settings between a clinical CCTA scan (a) and a DCE scan (b) on the same patient. The tube current exposure time product (mAs) of the CCTA and DCE scans were 247.5 and 27.5 mAs respectively.

[0211] Experimental Exemplification: Experimental Example 10 (Calculating distribution of tracer mass in a plurality of cardiovasculatures of interest, and calculate stenosis-induced pressure loss).

[0212] This example discloses new methods to assign tracer mass in each coronary artery for the calculation of volumetric flow rate, flow velocity pressure loss due to lumen constriction and pressure loss due to lumen expansion mentioned in Section 1(a) and 1(b). Equations (1), (2) and (8) appeared in in co-owned PCT / CA2019 / 050668 filed 16 May 2019. This approach for assigning 02 OCTOBER 2025 (02.10.2025) tracer mass may be generalized to other cardiovasculature, may be adapted to MR imaging, and may be useful in absence or presence of machine learning models.

[0213] Tracer mass assignment

[0214] Volumetric flow rate in a coronary artery can be calculated using the following equation: m

[0215] F = ~h - faCa(t)dt

[0216] (1) where F is volumetric flow rate, m is the mass of tracer, Ca(t) is the tracer concentration in the artery at time t. The lower and upper limits a and b in the integral denote the start and end time of the first-pass circulation phase of tracer in the artery (omitting the recirculation phase). Volumetric flow rate can be further converted to flow velocity as follow:

[0217] F

[0218] V = — nrA

[0219] (2) where r is the radius of the artery. Herein discloses a new method to estimate the tracer mass entering into each coronary artery. The distribution of tracer mass in each coronary artery is calculated as follow:

[0220] _FTRCA mRCA ~mtotal ' J ' yj

[0221] (3)

[0222] _FFLAD mLAD —mtotal ' J ' yj

[0223] (4) mLM —mLAD +mLCX

[0224] (6) where mRcA , WILAD, THLCX and mLMis the tracer mass distributed in the artery RCA, LAD, LCX and left main (LM) artery respectively; mMaiis the total tracer mass injected into the patient, / is the fraction of the total tracer mass; TRCA , TLAD and TLCx is the number of myocardial segments supplied by the artery RCA, LAD and LCX respectively. The variable / can consume any value 02 OCTOBER 2025 (02.10.2025) between 0.03 and 0.06 (or 3% to 6%) for the rest condition, and any value between 1 (0.03 to 0.06) to 3 (0.03 to 0.06) for the stress condition (or 100% to 300% of the rest condition value). By default, TRCA, TRAD and TRCX are set to 5, 7 and 5 respectively. However, these numbers can be adjusted according to the following conditions: (i) dominance of the coronary system, (ii) presence of collateral circulation, (iii) presence of the ramus intermedins branch. Conditions (i), (ii) and (iii) can be assessed with the individual’s CCTA images.

[0225] For instance, the posterior descending artery (PDA) supplies the inferior portion of the interventricular septum. Most patient’s coronary system is “right dominant” where the PDA is supplied by the RCA. In a “left dominant” coronary system, the PDA is supplied by the LCX. In a “co-dominant” coronary system, the PDA can be supplied by both the RCA and LCX. Additionally, the ramus intermedins is a variant coronary artery originating from trifurcation of the left main artery . It can supply the myocardium in a similar fashion to the obtuse marginal branches of the left circumflex artery (LCX) or the diagonal branches of the left anterior descending artery (LAD).

[0226] Stenosis-induced pressure loss

[0227] PLS is the pressure loss due luminal narrowing (stenosis). PLs can be further approximated using the following equation:

[0228] PLS=[(T’LC + PLE) ' P ' £l]

[0229] (7) where PLC and PLE are the head loss due to abrupt reduction and expansion of the cross- sectional area of a blood vessel, respectively, p is the density of blood and g is the acceleration due to Earth’s gravity.

[0230] The head loss due to an abrupt expansion of the cross-sectional area of the lumen (PLE) is estimated using the following equation: where Vin is the in-stenosis flow velocity and Vpost is the post-stenosis flow velocity, g is acceleration due to Earth’s gravity. The head loss due to an abrupt constriction of the cross- sectional area of the lumen (PLC) is estimated with the following equation: 02 OCTOBER 2025 (02.10.2025)

[0231] The factor "0.5" in the right-hand side of equation (9) is an approximation, and can fluctuate between 0.1 to 1.0. The value of this factor changes proportional to the severity of the degree of stenosis in the artery of interest. The pair of PLE and PLC values computed with equations (8) and (9) should be used to derive PLSusing equation (7) for all the downstream slices. In the situation where there is a second stenosis in the downstream slices, the pair of PLE and PLC values corresponding to the second stenosis should be calculated in a similar fashion as for the first stenosis. Then, the two pairs of PLE and PLC values (corresponding to 1st and 2nd stenosis) should be summed and used to update the PLS value for all the slices downstream of the 2nd stenosis.

[0232] To estimate PLE and PLC, the in-stenosis flow velocity (Vin) is needed. According to equation (1) and (2), Vin can be estimated based on the time-enhancement curve sampled within the stenosis. However, due to partial volume averaging and artifacts arising from calcification, measurement of in-stenosis time-enhancement curve may not be accurate. Alternatively, Vin can be estimated without in-stenosis time-enhancement curve based on the continuity principle: where r^ and rpreare the in-stenosis and pre-stenosis lumen radius, respectively. For example, if the stenosis is 50%, then rin= 0.5 rpre, and the ratio of rpreand rinbecomes 2. As another example, if the stenosis is 70%, then rin= 0.3rpre, and the ratio of rpreand rinbecomes 3.33. Both Vpre and rpreshould be measured at the same normal segment of a coronary artery.

[0233] When tracers travel along the artery, a portion of the tracers is lost to side branches. The tracer mass remains in the main artery after passing through a side branch can be estimated as follow (Figure 48): 02 OCTOBER 2025 (02.10.2025) where mpreb is the tracer mass in the main artery before a side branch, mb is the tracer mass in a side branch, mpostb is the tracer mass in the main artery after a side branch; rpreis the radius of the main artery before a side branch, rb is the radius of the side branch, rpostb is the radius of the main artery after a side branch.

[0234] Validation results for Experimental Example 10 are presented in Figure 62. (A)-(B) Assessment of FFR across the prelesion and postlesion locations in the RCA artery (see the cross in each image). (C) Calculation of the residual tracer percentage at different locations along the RCA based on Equations 3, 4, 5, 6 and 11. The residual tracer percentages are shown in column J in the Excel screenshot. There were six branches between the two locations shown in (A) and (B). (D) Plot of residual tracer percentage in the RCA after passing through each branch. (E)-(J) The timeenhancement curves sampled from the prelesion and postlesion locations were fed to an in-house software to compute the prelesion and postlesion flow velocities using Equation (1) and (2). Equations (7) to (9) were also used to compute the pressure loss arising from the stenosis. Finally, the fractional flow reserve (FFR) was computed using the Bernoulli’s Equation. Our algorithm showed the FFR was 0.83, which agreed well with the invasive (catheter-based) FFR measurement (0.83) at the same postlesion site.

[0235] Experimental Exemplification: Experimental Example 11 (Selecting a peak contrast enhanced image from a plurality of corresponding images).

[0236] The code in Figure 63A (Figures 63B and 63C are magnified views of Figure 63A) illustrates the convolution neural network employed in a machine learning model for predicting the time point corresponding to peak arterial enhancement in dynamic CT myocardial perfusion images. A graphical representation of the same neural network is provided in Figure 64. This neural network has three hidden layers, including two convolution layers (code lines 4 and 8) followed by one fully connected layer (code line 13). Each convolution layer has 16 nodes (neurons) and the fully connected layer has 64 nodes. As shown in code line 2, the in-plane resolution of the input images were 128x128 pixels, which is lower than the default in-plane resolution of CT images (512x512 pixels). Although the example shown here is for the coronary 02 OCTOBER 2025 (02.10.2025) application, which aims to predict the peak coronary enhancement in dynamic myocardial perfusion images, our training targets on predicting the peak aortic enhancement. This approach is justified by the fact that the peak signal enhancement in the coronary arteries always occur at the same time as that in the ascending aorta (because blood flow is fast and the coronary arteries are connected to the ascending aorta). Since the aorta is much larger than the coronary arteries (the normal aortic diameter is about 40 mm), and the training goal is to predict peak aortic enhancement, high-resolution images are not necessary for this task. Therefore, the input images were degraded to 128x128 by combining adjacent 4x4 pixels in the images to reduce the training time. Image resolution reduction for the training images is further demonstrated in the code shown in Figure 65 (code line 23 to 41). The same section of code also revealed every 8 consecutive slices were averaged to reduce the total number of slices in each training image set. The final section of the code (line 54 to 58) shows the dataset was rescaled from 0 to 1. This normalization allows a faster convergence to reach the optimal solution, which is beneficial if the input image features have widely different ranges (in our case this feature is the arterial enhancement in Hounsfield Unit).

[0237] Figure 66 shows a partial list of the studies used for training the convolution neural network. There were 186 cases used for the training with the MATLAB deep learning toolbox. This toolbox randomly shuffled these cases and used 70% and 30% of the cases for learning and validation. Furthermore, the optimizer was set to the Adam (Adaptive Moment Estimation) algorithm. Most study subjects (go by the case number in column A) shown in this figure had both rest and stress dynamic perfusion image sets. Each perfusion image set has approximately 10 to 14 time points (column B) and the time point corresponding to the peak aortic / coronary enhancement is visually identified by an experience reader as ground truth.

[0238] Figure 67 shows the code (lines 3 to 6) for loading the matrix of weights learned during training to detect specific features in the dataset (such as the peak enhancement in the ascending aorta) and the dataset (perfusion images). The peak enhancement prediction is executed using the trained network in line 9. The result would be plotted by executing the code in line 10.

[0239] The plot in Figure 68A (Figure 68B is a magnification of the plot shown in Figure 68A) shows the training and validation results. The Y-axis represents the ground truth. The number label in each row of the Y-axis represents the time point corresponding to the true peak aortic / coronary enhancement in a dynamic perfusion image set. For instance, the “4” in the second row (from top) of the Y-axis represents time point 4. The “5” in the fourth row of the Y-axis represents time point “5”. The double-digit numbers in the Y-axis represent cases with two enhancement peaks. For 02 OCTOBER 2025 (02.10.2025) instance, the “34” in the first row represents time points “3 and 4”. The “45” in the third row represents time points “4 and 5”, and so forth. Theoretically, the peak enhancement should appear solely at one time point after a bolus injection of contrast. However, due to the limited frequency of image acquisition, the true peak of a time-enhancement curve may be missed and only the nearby data point on each side of the peak is sampled, leading to two similar peaks. The X-axis of the plot represents the time points predicted by the trained model. The number labels of the X-axis can be interpreted in a similar manner.

[0240] The number in each box along the diagonal in the graph represents the number of cases correctly predicted by the trained model. For instance, the box at the top left comer of the graph along the diagonal has a number “20”. It means that 20 cases with the peak aortic / coronary enhancement at time points “3 and 4” in the corresponding dynamic perfusion image sets were correctly predicted by the trained model. Another example is the box at the low right comer of the graph along the diagonal has a number “18”. It means that 18 cases with the peak enhancement at time points “9 and 10” in the corresponding dynamic perfusion image sets were correctly predicted by the trained model. All the numbers that are not along the diagonal of the graph represent the cases where the trained model failed to make the correct prediction. For example, the box located in the 3rd row (from top) and the 5th column (from left) has a number “1”. It means that the trained model predicted the peak enhancement at time points “5 and 6” (see the X-axis label), which is wrong because the ground truth is time points “4 and 5” instead (see the Y -axis label). As shown in the plot, the trained model made incorrect prediction only in 4 out of 186 cases. The result suggests the proposed approach for predicting peak arterial enhancement is feasible.

[0241] Several illustrative variants of a method or system for blood flow imaging have been described above. Further variants and modifications are described below. Moreover, guiding relationships for configuring variants and modifications are also described below. Still further variants and modifications are contemplated and will be recognized by the person of skill in the art. It is to be understood that guiding relationships and illustrative variants or modifications are provided for the purpose of enhancing the understanding of the person of skill in the art and are not intended as limiting statements.

[0242] For example, several machine learning methods have been described for assisting various aspects of image processing / analysis in blood flow assessment using DCE CT or MRI including, for example: (1) extracting the DCE images corresponding to the optimal arterial enhancement for anatomical assessment of the arteries of interest; (2) delineating the arteries of interest across 02 OCTOBER 2025 (02.10.2025) multiple slices to facilitate automatic sampling of the time-enhancement curves; (3) estimating the area under a virtual hyperemic time-enhancement curve from a sampled resting time-enhancement curve to spare the need of pharmacologic vasodilation; (4) determining the functional significance of a vascular stenosis based on non-image and image features such as the area under curve to guide patient treatment. Multiple forms of neural networks are used for the above tasks. Tasks 3 and 4 use the standard neural network; Task 1 relies on a convolution neural network (CNN), which is a subtype of neural network; Task 2 relies on a U-shaped convolution neural network (a subtype of CNN). One difference between Task 3 (prediction of area under the curve) and Task 4 (prediction of functionally significant stenosis) lies in the final layer of the neural network. In Task 3, a linear regression model is used to predict a continuous numerical value (i.e., the area under the curve). In Task 4, a sigmoid function is used to predict a binary outcome (i.e. presence or absence of functionally significant stenosis). Neither Tasks 3 nor Task 4 use a CNN or U-NET because the images are not directly fed into the machine learning model. Instead, image features such as the area under the curve are extracted as numerical values first before machine learning training. One advantage of this approach is that it improves training accuracy. For instance, in cardiac applications, coronary time-enhancement curves are used to predict the flow and pressure across a stenosis, thereby determining its functional significance. However, the coronary arteries are adjacent to the heart, which has multiple heart chambers. The temporal changes in enhancement in each heart chamber resulting from the fast circulation of contrast are unrelated to the stenosis within the coronary arteries. Yet, these irrelevant enhancements in the chambers will contaminate the coronary enhancement during the convolution process. The extracted image features will less likely be correlated with the desired prediction.

[0243] As another example, the blood flow imaging method 20 as shown Figure 2 is merely illustrative, and should not be considered as limiting to the blood flow imaging method as one or more steps shown in Figure 2 can be substituted or removed as desired for a specific implementation. For example, in a specific implementation CT scanning of a subject may be geographically or temporally displaced from image reconstruction. An example, of a contemplated variant blood flow imaging method includes both projection data from CT scanning and image reconstruction occurring at a prior stage and reconstructed images are stored for analysis at either a later date or for analysis by a third party. The variant blood flow imaging method can initiate by obtaining the stored image data. Contrast agent signal data can then be extracted from the stored image data, optionally without explicitly identifying a target blood vessel in the image data. A time-enhancement curve is generated based on the contrast agent signal data, the time- 02 OCTOBER 2025 (02.10.2025) enhancement curve having an upslope plotted from data points obtained during an increase phase of the contrast agent signal data, and a downslope plotted from data points obtained during a decline phase of the contrast agent signal data. A flow velocity value is then determined according to the same method steps shown in Figure 6 A or Figure 6B.

[0244] As another example, the blood flow imaging method and system are not limited to computed tomography (CT) scanning, and can readily be adapted to other imaging modalities that have sufficient spatial resolution to image blood vessels and exhibit proportional increase in signal intensity in a ROI as a function of the mass of contrast agent present in the RO1 (more contrast agent or tracers results in a higher signal in the ROI), including MRI and other X-ray imaging techniques (ie., X-ray imaging techniques other than CT imaging), including for example fluoroscopy. X-ray based scans are a form of medical imaging comprising transmission of a high frequency electromagnetic signal that becomes attenuated as it passes through the body of a subject with the remaining signal captured by a detector for subsequent analysis. Data for the Experimental Examples was acquired with a single-energy CT (SECT) scanner. Most clinical CT scanners use single-energy acquisition. However, dual-energy CT (DECT) scanners are also available. Dualenergy CT refers to two X-ray energy sources used for scanning an object instead of a single X-ray energy source. Existing literature shows that dual-energy CT can perform dynamic CT acquisition just like single-energy CT. From the image processing aspect, nothing changes and methods described herein or in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021, such as methods based on Reynolds Transport Theorem or Indicator-Dilution Principle or Bernoulli’s equation, can be applied in both SECT and DECT.

[0245] An alternative to X-ray based scans is Magnetic Resonance Imaging (MRI), which has well-recognized medical imaging applications including for example, imaging to diagnose disease in soft tissues such as the brain, lungs, liver, muscles, and heart. MRI scans involve the application of a magnetic field to a patient and the transmission of radio frequency pulses. Resonance energy is emitted by the patient and picked up by a receiver / detector that captures scan data for subsequent analysis. To improve image clarity, both X-ray scans and MRI scans involve the oral or intravenous administration of a contrast agent to a patient. Contrast agents for X-ray imaging techniques include for example iodine-based contrast agents. Contrast agents for MRI imaging techniques include for example gadolinium -based contrast agents. Scan data acquired from X-ray based scanner de vic es / sy stems are often referenced as scan data or projection data interchangeably, 02 OCTOBER 2025 (02.10.2025) while scan data acquired from MRI scanner devices / systems are typically referenced as scan data. Thus, the term scan data is understood to encompass the term projection data.

[0246] The methods described herein and as adapted from co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021, demonstrated with dynamic contrast-enhanced CT imaging data obtained after an intravenous bolus injection of iodine-based contrast agent, are also applicable for dynamic MRI imaging data obtained after intravenous bolus injection of Gadolinium-based (Gd) contrast agent. We have demonstrated with a preclinical study in co-owned PCT application no. PCT / CA2019 / 050668 (filed 16 May 2019) that a time-enhancement curve in a region of interest can be obtained from dynamic contrast- enhanced MRI imaging in a similar manner to dynamic contrast-enhanced CT imaging. Furthermore, the temporal change in signal intensity (e.g. Ti relaxation time) over time is induced by the movement of Gd contrast agent in the region of interest, and the magnitude of signal alteration is closely related to the concentration of Gd-based molecules (tracers). When a low concentration of Gd contrast agent is used, the change in MRI signal intensity and contrast concentration in a region of interest exhibits a relatively linear relationship. Moreover, with respect to co-owned PCT / CA2021 / 051189 filed 26 August 2021, this linear relationship facilitates the estimation of the time rate of change of mass of tracer (dm / dt in Equation 11 in co-owned PCT / CA2021 / 051189 filed 26 August 2021) with the RTT method to derive flow velocity.

[0247] Contrast agents (also referred to as tracers) for various imaging modalities are established in the current literature and continue to be an active area of development for new alternatives. The blood flow imaging method and system may accommodate any suitable combination of contrast agent and imaging modality provided that the imaging modality affords sufficient temporal and spatial resolution to image a cardiovasculature of interest, for example a blood vessel of interest or a portion of a blood vessel of interest or a heart chamber of interest or a portion of a lumen of a heart chamber of interest.

[0248] The blood flow imaging method and system can include a selection of a target voxel or pixel in the acquired image data (ie., acquired pixel data) or image data outputted by a computer learning model and analysis of the pixel data in the selected voxel or pixel. While voxels provide precision to volumetric imaging, voxel based assessment can also be disadvantaged by large data sets that are unwieldy to manage given the bandwidth of common computers. The systems and methods described herein provide an efficient manipulation of large data files that permits interactive visualization and fine temporal resolution with near real-time assessment using commonly available computers. 02 OCTOBER 2025 (02.10.2025)

[0249] A voxel is the smallest 3D element of volume and is typically represented as a cube or a box, with height, width and depth dimensions (or 3D Cartesian coordinate x, y and z dimensions). Just as 2D images are made of several pixels (represented as squares, with height and width, or x and y dimensions) and the smaller the pixel the better the quality of the picture, the same concept applies to a 3D data volume. In data acquisition, each three-dimensional voxel represents a specific x-ray absorption. A voxel stated as isotropic means that all dimensions of the isotropic voxel are the same and typically the isotropic voxel is a perfect cube, with uniform resolution in all directions. In contrast, a voxel stated as anisotropic or non-isotropic means that the anisotropic voxel is not a perfect cube, such that all dimensions of the voxel are not the same (ie., at least one dimension of the anisotropic voxel is different than other dimensions) or that the anisotropic voxel includes partial voxel units (typically more than one voxel unit). The systems and methods described herein provide an efficient manipulation of image data that permits operability with selection of one or both of isotropic or non-isotropic target voxels.

[0250] The terms ROI and target voxel are related, as an ROI in reconstructed 3D image data will encompass either a target voxel or a block of neighboring target voxels. In reconstructed 2D image data, an ROI will encompass either a target pixel or a block of neighboring target pixels, and therefore the terms ROI and target pixel are also related. The terms voxel and pixel are related as voxel is a 3D analog of a pixel. Voxel size is related to both the pixel size and slice thickness. Pixel size is dependent on both the field of view and the image matrix.

[0251] A selected isotropic target voxel may be a single isotropic voxel or a continuous block of neighboring or adjacent voxels where the block is isotropic. A selected non-isotropic target voxel will typically encompass more than one voxel unit, but may approach a volume of a single voxel unit or may be a continuous block of neighboring or adjacent voxels where the block is non- isotropic. A non-isotropic block of voxels can include parts of voxels at its boundary as would be expected if the target voxel is a non-square shape such as a circle or triangle. Thus, blocks of target voxels or target pixels need not be limited to full voxel or pixel units as an ROI of various shapes (including circles, triangles or even irregular shapes) may be accommodated, and an ROI may defining a block of neighboring voxels or pixels with partial voxels at the boundary of the ROI.

[0252] The elapsed time of an imaging scan procedure, equivalent to the time duration of scan data acquisition, can be varied as desired provided that the imaging scan captures at least a portion of an increase phase or a decline phase of contrast agent at the sampling site so as to obtain sufficient data to estimate or predict area under the time-enhancement curve. Generally, to capture a portion of both increase and decline phases an imaging scan of greater than 5 seconds is needed. In certain 02 OCTOBER 2025 (02.10.2025) examples, imaging scans can be configured to capture scan data for greater than 6 seconds, greater than 7 seconds, greater than 8 seconds, greater than 9 seconds or greater than 10 seconds. Although not constrained by an upper time limit and not constrained by the transit time of contrast agent, most often imaging scans will not extend significantly beyond the expected transit time of contrast agent at a sampling site.

[0253] The number of images (also referred to as frames or individual scans) analyzed to predict the area under the time-enhancement curve can be varied as desired provided that the number of images cumulatively captures at least a portion of an increase phase or a decline phase of contrast agent at the sampling site so as to obtain sufficient data to estimate shape of the time-enhancement curve. Generally, to capture both increase and decline phases an imaging scans of greater than 5 images is needed. In certain examples, imaging scans can be configured to capture scan data for greater than 6 images, greater than 8 images, greater than 10 images, greater than 12 images, greater than 14 images, greater than 16 images, greater than 18 images, or greater than 20 images. Additionally, imaging scans configured to capture at least 10 images are observed to benefit consistency of peak value determinations and curve shape; signal intensity values need not be extracted from all of the at least 10 images, but the at least 10 images often provides a large enough set of images to select a subset of appropriate time-distributed images (typically 5 or more images) that leads to consistency of estimating curve shape.

[0254] The blood flow imaging method and system is considered dynamic due to analysis of a plurality of images as distinguished from static techniques that evaluate a single image. Most commercially available CT angiography techniques are static. Furthermore, commercially available CT angiography techniques that are minimally dynamic (evaluating 2 to 3 images) do not recognize or consider benefits of acquiring scan data from both the increase phase and decline phase of contrast agent transit or generating a time-enhancement curve or predicting an area under a time-enhancement curve using a computer learning model. Furthermore, CT angiography studies that obtain 2 or 3 images at slightly different time frames, for motion correction, or for the doctor to select the best image that is least affected by motion, may also be considered a static technique.

[0255] A plurality of images, for example at least 5 images, for predicting an area under a timeenhancement curve are considered to be a plurality of corresponding images with the correspondence of images referring to a time-ordered sequence of multiple images located in the same sampling site or slice or in a group of adjacent sampling sites or slices. Thus, correspondence of images is spatially limited to a single sampling site or slice or a group of adjacent sampling sites or slices (or to a single ROI or a group of adjacent ROIs), and correspondence of images does not 02 OCTOBER 2025 (02.10.2025) include sampling sites or slices spatially separated to be upstream versus downstream of a source of blood flow aberration. For example, when determining a blood flow characteristic comprises a comparison of corresponding values calculated from first and second time-enhancement curves, the first time-enhancement curve may be from a first plurality (or set) of corresponding images from a first sampling site or slice located upstream of a suspected source of a blood flow aberration and the second time-enhancement curve may be from a second plurality (or set) of corresponding images from a second sampling site or slice located downstream of the suspected source of the blood flow aberration. In this example, the first set of corresponding images will not be intermingled with the second set of corresponding images as the first and second sampling sites are spatially separated by an intervening suspected source of blood flow aberration.

[0256] Each set or plurality of corresponding images is optionally time-ordered or time-resolved to benefit input to a computer learning model to predict an area under a time-enhancement curve. The time-enhancement curve can have an upslope, a peak and a downslope. Time-ordering benefits computer learning and computer prediction of an area under the time-enhancement curve. Timeordering provides an upslope of the time enhancement curve interpolated from time-specific contrast agent signal data points acquired during an increase phase of contrast agent transit, and the downslope of the time enhancement curve interpolated from time-specific contrast agent signal data points acquired during a decline phase of contrast agent transit. Accordingly, acquisition of scan data and reconstruction of image data occurs with reference to a time-ordering scheme such that each set of corresponding images obtained from the image data can be arranged in a time- ordered sequence. A time-ordering scheme can be any convenient scheme including a time stamp with a real-time identifier, a relative-time identifier such as elapsed time from bolus injection, or any customized time identifier that can be used for identifying absolute or relative time of each image and time-resolved sequencing of the set of corresponding images. Established protocols for time intervals between contrast agent administration and image acquisition may be adopted in devising a time ordering scheme. Furthermore, established timing techniques, for example bolus tracking, may be adopted to optimize timing of scan acquisition and time-ordering of image data.

[0257] The time-enhancement curve is a plot of contrast agent signal intensity versus time derived from scan data of a contrast agent transit at a single sampling site or a group of adjacent sampling sites. The time-enhancement curve may also be referred to as a time-density curve, signal intensity time curve, time-dependent signal intensity, time-intensity curve among other variations. The machine learning model does not need to plot a time-enhancement curve (plots shown in the Figures are for illustration for easier comparison to technology described in co-owned 02 OCTOBER 2025 (02.10.2025)

[0258] PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021) and can predict an area under a time-enhancement curve directly from inputted images by directly assessing enhancement with the inputted images or selected portions therein. The term enhancement within the term time-enhancement curve refers to an increase in measured contrast signal intensity relative to a baseline or reference value such as signal intensity measured at a minimal level of contrast agent or measured at a residual level of contrast agent or measured in absence of contrast agent. Qualitative terms describing a contrast agent transit, such as prior to entry, entry, wash-in, increase phase, decline phase, wash-out, clearance and subsequent to clearance, are referenced to a bolus injection event or more generally a contrast agent administration event, such that each of these terms, except prior to entry, describing a portion of a contrast agent transit that occurs subsequent to an associated injection or administration event. The term prior to entry may correspond to a time range that may begin earlier than the injection or administration event.

[0259] The blood flow imaging method and system described herein allows for determination of a blood flow characteristic. A blood flow characteristic may be any metric that assesses blood flow at a region of interest in a subject. A blood flow characteristic includes those described in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021, for example, flow rate, flow velocity, flow acceleration, flow pressure and reconstruction of heart-induced pulsation. Heart-induced pulsation refers to temporal variation of flow rate / flow velocity arising from the heart contraction and relaxation (which lead to forward ejection and backward suction of blood respectively). Rate, velocity, and acceleration are metrics of blood flow. The blood flow imaging technique can include other blood flow assessment techniques as desired, for example blood flow assessment or blood pressure assessment (using Bernoulli’s equation) as described in co-owned International PCT Application No. PCT / CA2019 / 050668 filed 16 May 2019 which also describes fractional flow reserve (FFR), blood flow reserve (BFR), and shear stress as blood flow characteristics that may be quantified; and also describes area under the curve, rate of change of area under the curve, peak (maximum value) of the curve, and blood volume as further examples of a blood flow characteristic.

[0260] A blood flow characteristic can be determined from raw signal intensity measurement or enhancement measurements. In CT, measured signal intensity can be stated as CT number, while enhancement infers a normalization against a reference value or a subtraction of signal intensities.

[0261] The determination of a blood flow characteristic can minimally require computer learning model prediction of an area under a time-enhancement curve and may optionally include computer 02 OCTOBER 2025 (02.10.2025) prediction of other parameters such as a time rate of change of a parameter including for example, time rate of change of signal intensity, time rate of change of enhancement, time rate of change tracer mass, time rate of change of flow velocity, time rate of change of flow pressure, and the like. The various time rate of change parameters are related as described in mathematical derivations provided in co-owned PCT / CA2019 / 050668 filed 16 May 2019 or in co-owned PCT / CA2021 / 051189 filed 26 August 2021.

[0262] Assessment of blood flow and determination of a blood flow characteristic can provide a diagnostic result. For example, predicting an area under time-enhancement curves at first and second sampling sites (or first and second RO Is in the same sampling slice) yields a first area under a time-enhancement curve and a second area under a time-enhancement curve; and estimating of the blood flow characteristic comprises a determination including corresponding values calculated from the first and second areas under time-enhancement curves. As another example, the first timeenhancement curve is generated from image data acquired with the blood vessel of interest in a hyperemic state and the second time-enhancement curve is generated from image data with the blood vessel of interest in a resting state, and the blood flow characteristic is a blood flow reserve determined based on a ratio of corresponding blood flow values calculated from the first and second time-enhancement curves. As another example, a relative flow velocity or absolute flow velocity may be determined at one or more ROIs. The blood flow characteristic value may in itself provide a diagnostic result. In further examples, corresponding values calculated from the first and second time-enhancement curves or first and second flow velocities are compared and a difference in the corresponding values beyond a predetermined threshold is indicative of a diagnostic result. Thresholds and corresponding diagnostic results can be adopted from relevant literature and medical guidelines. Furthermore, with repeated use of the blood flow imaging method and system, various correlations of metrics, thresholds and diagnostic results may be developed.

[0263] A region of interest (ROI) is an area on a digital image that circumscribes or encompasses a desired anatomical location, for example a blood vessel of interest or a portion of a lumen of a blood vessel of interest or heart chamber of interest or any other cardiovasculature of interest. The terms ROI and target voxel or target pixels are related as the ROI defines an area that encompasses one or more voxels (in 3D imaging) or one or more pixels (in 2D imaging). The terms voxel and pixel are related in that both rely on pixel data, but voxel is a 3D-analog of pixel and is an accumulation of pixel data from multiple slices in a 3D image.

[0264] Image processing systems permit extraction of pixel data from ROI on images, including for example an average parametric value computed for all pixels within the ROI. A sampling site is 02 OCTOBER 2025 (02.10.2025) the location of one or more imaging slices selected to assess a desired anatomical location, such as a blood vessel of interest or heart chamber of interest or any other cardiovasculature of interest. In some examples, analysis of a time-enhancement curve from a single ROI may be sufficient to determine a blood flow characteristic or metric. In other examples, a plurality of ROIs in a single sampling site or a plurality of ROIs in a plurality of sampling sites, or a plurality imaging slices may be analyzed to obtain a plurality of corresponding image sets and to generate a plurality of corresponding time-enhancement curves, and any number of the plurality of corresponding timeenhancement curves may be compared to determine a blood flow characteristic or blood flow metric. Conventional scanners can capture 3D image data for all or part of a blood vessel of interest or other cardiovasculature of interest, and possibly even all or parts of a plurality of vascular structures such as a plurality of blood vessels of interest. Furthermore, a scan can be subdivided into a plurality of slices as desired, and therefore interrogation of multiple sites or slices at an ROI, near an ROI, upstream of an ROI, downstream of any ROI, or any combination thereof, is feasible and convenient. In multi-slice or multi -site imaging modalities, simultaneous tomographic slices or sampling sites may be extracted per scan. Thus, the blood flow imaging method need not be limited to analysis of one or two time-enhancement curves for a scan of a contrast agent transit (entry to clearance) at blood vessel interest and a single scanning procedure with a single bolus injection of contrast agent can support a plurality of slices or sampling sites divided from the scan data as desired.

[0265] Motion correction or motion compensation processing of reconstructed image data may be used if ROIs benefit from adjustment to accommodate the movement of the vessel wall during the cardiac cycle. Rules-based or machine learned motion correction or compensation models are available, and may be used as desired for specific implementations.

[0266] A cardiovasculature of interest (also referred to as vascular structure of interest) may be any blood flow passage or lumen of the cardiovascular system (also referred to as the circulatory system), and may include any blood vessel of interest (including for example systemic arteries, peripheral arteries, coronary arteries, pulmonary arteries, carotid arteries, systemic veins, peripheral veins, coronary veins, pulmonary veins) or any heart chamber of interest or any heart aperture of interest that can be imaged by a contrast-enhanced imaging technique. The cardiovasculature of interest will typically have a diameter of at least about 0.1 mm, for example a diameter greater than 0.2 mm or a diameter greater than 0.3 mm. The cardiovasculature of interest, such as a blood vessel of interest or a designated portion of the blood vessel of interest, may be identified and targeted for contrast enhanced blood flow imaging to determine a diagnosis of a 02 OCTOBER 2025 (02.10.2025) cardiovascular disorder or a blood vessel disorder or to determine a predisposition to such disorder. A blood vessel of interest can be within any anatomical area or any organ (for example, brain, lung, heart, liver, kidney and the like) in an animal body (for example, a human body).

[0267] The blood flow imaging method is not limited to scan data acquired while a subject is in a hyperemic state (also referred to as hyperemic stress or vasodilatory stress) and time -enhancement curves generated from scan data acquired while a subject is in a non-hyperemic state (also referred to as a resting state) can produce a useful result. Inducing a hyperemic state is a well-known medical protocol in blood flow assessment and often includes administration of a vasodilator such as adenosine, sodium nitroprusside, dipyridamole, regadenoson, or nitroglycerin. Mode of administration of the vasodilator may vary depending on an imaging protocol and can include intravenous or intracoronary injection.

[0268] To determine a presence of a cardiovascular disorder at a cardiovasculature of interest, such as a blood vessel disorder at a blood vessel of interest, a blood flow characteristic will be analyzed based on at least one area under a time-enhancement curve, including for example a single area under a time-enhancement curve generated from pixel data of an ROI in a scan of a single sampling site, or as another example a plurality of areas under time-enhancement curves respectively generated from a corresponding plurality of sampling sites. In a case of stenosis a comparison of two sampling sites is beneficial to compare a blood flow characteristic determined at a sampling site upstream of the stenosis with a blood flow characteristic determined at a sampling site downstream of the stenosis. More generally, when a blood vessel of interest is identified, a plurality of sampling sites may be designated at or near the blood vessel of interest; a time- enhancement curve generated for each of the plurality of sampling sites; a desired blood flow characteristic based on a respective time-enhancement curve determined for each of the plurality of sampling sites; and comparing the determined blood flow characteristic of each of the plurality of sampling sites to determine a blood vessel disorder. Depending on a specific implementation determining of a blood flow characteristic at one or more sampling sites or determining presence of absence of a blood vessel disorder based on a comparison of blood flow characteristic at a plurality of sampling sites can provide a diagnostic result.

[0269] A cardiovascular disorder or a blood vessel disorder (may also be referred to as a vascular disorder) assessed by the method or system described herein can be any unhealthy blood flow aberration such as a functionally significant blood flow restriction or blood flow obstruction in a cardiac or non-cardiac blood vessel or any aberrant blood flow in a heart chamber or heart aperture that can compromise health of a subject including for example, unhealthy blood flow aberrations 02 OCTOBER 2025 (02.10.2025) symptomatic of Heart Chamber Abnormalities, Heart Valve Abnormalities (eg., Aortic Valve Disease), Heart Failure, Atherosclerosis (for example, plaque formation), Coronary Artery Disease, Carotid Artery Disease, Peripheral Artery Disease including Renal Artery Disease, Aneurysm, Raynaud's Phenomenon (Raynaud's Disease or Raynaud's Syndrome), Buerger's Disease, Peripheral Venous Disease and Varicose Veins, Thrombosis and Embolism (for example, blood clots in veins), Blood Clotting Disorders, Ischemia, Angina, Heat Attack, Stroke and Lymphedema.

[0270] The blood flow imaging method and system can be used to assess a suspected cardiovascular disorder or blood flow disorder, for example by providing a determination of a blood flow characteristic at a blood vessel of interest identified in a previous medical examination as possible source of an unhealthy blood flow aberration. Additionally, due in part to scan data capturing multiple blood vessels and the reduced time to process scan data, the blood flow imaging method and system may be used in a first instance to proactively assess blood flow in a specific blood vessel or specific group of blood vessels (for example, a pulmonary artery blood flow assessment) and may be implemented as a screening tool to be an initial indicator to identify a source of unhealthy blood flow aberration such as a functionally significant stenosis.

[0271] The blood flow imaging method does not require the scanned subject or patient to hold breath during a scan procedure. Breath -hold is an option in some examples. In other examples, motion correction or motion compensation processing of image data may be used for scan data acquired without breath-hold of the subject or patient. If desired, motion correction or motion compensation processing of image data may be used for scan data acquired with breath-hold, if RO Is benefit from adjustment to accommodate the movement of the vessel wall during the cardiac cycle. Rules-based or machine learned motion correction or compensation models may be used as desired for specific implementations.

[0272] Embodiments disclosed herein, or portions thereof, can be implemented by programming one or more computer systems or devices with computer-executable instructions embodied in a non-transitory computer-readable medium. When executed by a processor, these instructions operate to cause these computer systems and devices to perform one or more functions particular to embodiments disclosed herein. Programming techniques, computer languages, devices, and computer-readable media necessary to accomplish this are known in the art.

[0273] In an example, a non-transitory computer readable medium embodying a computer program for blood flow imaging may comprise: computer program code for obtaining image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; computer 02 OCTOBER 2025 (02.10.2025) program code for providing the image data or image features to a machine learning model to predict an area under a time-enhancement curve of the contrast agent within the cardiovasculature of interest; computer program code for selecting a region of interest within the cardiovasculature of interest in the image data; and computer program code for determining a blood flow characteristic through the region of interest based on the area under the time-enhancement curve. In another related example, the image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent. In still another related example, the computer readable medium further comprises computer program code for acquiring scan data of the cardiovasculature of interest from a X-ray based scan or a MRI scan, and reconstructing image data based on the scan data.

[0274] The computer readable medium is a data storage device that can store data, which can thereafter, be read by a computer system. Examples of a computer readable medium include readonly memory, random-access memory, CD-ROMs, magnetic tape, optical data storage devices and the like. The computer readable medium may be geographically localized or may be distributed over a network coupled computer system so that the computer readable code is stored and executed in a distributed fashion.

[0275] Computer-implementation of the system or method typically comprises a memory, an interface and a processor. The types and arrangements of memory, interface and processor may be varied according to implementations. For example, the interface may include a software interface that communicates with an end-user computing device through an Internet connection. The interface may also include a physical electronic device configured to receive requests or queries from a device sending digital and / or analog information. In other examples, the interface can include a physical electronic device configured to receive signals and / or data relating to the blood flow imaging method and system, for example from an imaging scanner or image processing device.

[0276] The technology area of medical imaging, and particularly CT or MRI imaging described herein came into existence as a result of commercially available computers, and moreover required significant advances in semiconductor technology to achieve an implementation in a medical clinic. For CT, implementation in a medical clinic was initiated in the 1970s, while for MRI medical imaging implementation was initiated in the 1980s.

[0277] As such, CT or MRI imaging is intimately tied to computers and computer function, and CT or MRI imags are communicated and processed using specialized medical imaging software to process and extract data points to generate time-enhancement curves. For example, CT or MRI 02 OCTOBER 2025 (02.10.2025) image data can have a range greater than 1000 pixel / voxel values (continuous, not discretized) that make it impossible for human experts to accurately quantify enhancement. In addition, CT or MRI image data are stored and processed in DICOM format requiring specialized software to assess enhancement and time information to accurately generate a time-enhancement curve. Neither the time component nor the enhancement component of the time-enhancement curve is assessed and quantified without specialized software.

[0278] Any suitable processor type may be used depending on a specific implementation, including for example, a microprocessor, a programmable logic controller or a field programmable logic array. Moreover, any conventional computer architecture may be used for computerimplementation of the system or method including for example a memory, a mass storage device, a processor (CPU), a graphical processing unit (GPU), a Read-Only Memory (ROM), and a Random-Access Memory (RAM) generally connected to a system bus of data-processing apparatus. Memory can be implemented as a ROM, RAM, a combination thereof, or simply a general memory unit. Software modules in the form of routines and / or subroutines for carrying out features of the system or method can be stored within memory and then retrieved and processed via processor to perform a particular task or function. Similarly, one or more method steps may be encoded as a program component, stored as executable instructions within memory and then retrieved and processed via a processor. A user input device, such as a keyboard, mouse, or another pointing device, can be connected to PCI (Peripheral Component Interconnect) bus. If desired, the software may provide an environment that represents programs, files, options, and so forth by means of graphically displayed icons, menus, and dialog boxes on a computer monitor screen. For example, any number of blood flow images and blood flow characteristics may be displayed, including for example a time-enhancement curve.

[0279] Computer-implementation of the system or method may accommodate any type of end-user computing device including computing devices communicating over a networked connection. The computing device may display graphical interface elements for performing the various functions of the system or method, including for example display of a blood flow characteristic determined for a cardiovasculature of interest. For example, the computing device may be a server, desktop, laptop, notebook, tablet, personal digital assistant (PDA), PDA phone or smartphone, and the like. The computing device may be implemented using any appropriate combination of hardware and / or software configured for wired and / or wireless communication. Communication can occur over a network, for example, where remote control of the system is desired. 02 OCTOBER 2025 (02.10.2025)

[0280] If a networked connection is desired the system or method may accommodate any type of network. The network may be a single network or a combination of multiple networks. For example, the network may include the internet and / or one or more intranets, landline networks, wireless networks, and / or other appropriate types of communication networks. In another example, the network may comprise a wireless telecommunications network (e.g., cellular phone network) adapted to communicate with other communication networks, such as the Internet. For example, the network may comprise a computer network that makes use of a TCP / IP protocol (including protocols based on TCP / IP protocol, such as HTTP, HTTPS or FTP).

[0281] To further demonstrate variation of disclosed subject matter, several illustrative aspects are presented, in numerical itemized order for clarity:

[0282] 1. in illustrative aspect 1, a computer implemented method for blood flow imaging comprising: obtaining CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; extracting a first image feature of a measured time-enhancement curve from the CT or MRI image data; providing the first image feature and at least one non-image feature to a machine learning model to generate a predicted value of an area under a simulated time-enhancement curve of the contrast agent within the cardiovasculature of interest, the predicted value simulating a second set of image acquisition parameters that are different than a first set of image acquisition parameters used to acquire the CT or MRI image data; converting the predicted value of area under the simulated time-enhancement curve to a total sum of contrast agent concentration time product in the cardiovasculature of interest.; determining a blood flow characteristic in the cardiovasculature of interest based on a ratio of mass of the contrast agent in the cardiovasculature of interest to the a total sum of contrast agent concentration time product in the cardiovasculature of interest.

[0283] 2. The method of aspect 1, further comprising including a baseline data point extracted from the CT or MRI image data in the measured time-enhancement curve and denoising the measured timeenhancement curve, the CT or MRI image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent to provide the baseline data point.

[0284] 3. The method of aspect 1, wherein the machine learning model is trained with training inputs of the image feature of the measured time-enhancement curve and at least one non-image feature, and 02 OCTOBER 2025 (02.10.2025) associated with a ground truth value of an expected area under the simulated time-enhancement curve.

[0285] 4. The method of aspect 1, wherein the image feature is an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope.

[0286] 5. The method of aspect 1, wherein the image feature is a measured value of an area under an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope; and calculating the area under the at least partial time-enhancement curve.

[0287] 6. The method of aspect 1, further comprising providing a second image feature based on enhancement of signal intensity, thickness of a wall of a cardiovasculature of interest, size of a cardiovasculature of interest, diameter of a cardiovasculature of interest, morphology of a cardiovasculature of interest, location of sampling site in a cardiovascular of interest, or degree of stenosis in a cardiovasculature of interest.

[0288] 7. The method of aspect 1, wherein the non-image feature is based on age, sex, weight, heart rate, blood pressure, x-ray tube voltage, x-ray tube current, gradient pulse sequence, disease state, contrast-injection rate, contrast agent volume, or contrast agent concentration.

[0289] 8. The method of aspect 1, wherein the first set of image acquisition parameters is different than the simulated second set of image acquisition parameters according to at least one parameter selected from the group consisting of: scan axis orientation relative to a longitudinal axis of the cardiovasculature of interest, anatomical location of scan, hyperemic or rest condition of a subject, time duration of scan, x-ray tube voltage, x-ray tube current, gradient pulse sequence, contrastinjection rate, contrast agent volume, or contrast agent concentration.

[0290] 9. The method of aspect 1, wherein determining the blood flow characteristic comprises determining an absolute flow velocity using Reynolds transport theorem by determining a time rate of change of contrast agent mass in the cardiovasculature of interest based on a time rate of change of signal enhancement measured from an at least partial timeenhancement curve, a factor for converting signal enhancement to contrast agent concentration, a predetermined volume value of injected contrast agent, and a time value of a duration of time that signal intensity is higher than a predetermined baseline threshold level; 02 OCTOBER 2025 (02.10.2025) determining density of the contrast agent in the cardiovasculature of interest based on the ratio defined in aspect 1, a time value of a duration of time that signal intensity is higher than a predetermined baseline threshold level, and the mass of contrast agent in the cardiovasculature of interest; and determining an area measured from the image data in the cardiovasculature of interest.

[0291] 10. The method of any one of aspects 1 to 9, wherein determining the blood flow characteristic comprises determining a flow pressure by applying Bernoulli’s equation expressed as

[0292] 11 . in illustrative aspect 1 1 , a system for blood flow imaging comprising: a memory for storing CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; a CT or MRI image processing component to extract a first image feature of a measured time-enhancement curve from the CT or MRI image data; a machine learning model to generate a predicted value of an area under a simulated timeenhancement curve of the contrast agent within the cardiovasculature of interest by inputting the first image feature and at least one non-image feature to the machine learning model, the predicted value simulating a second set of image acquisition parameters that are different than a first set of image acquisition parameters used to acquire the CT or MRI image data; and a processor executing instructions to communicate with the CT or MRI image processing component and the machine learning model, and convert the predicted value of area under the simulated time-enhancement curve to a total sum of contrast agent concentration time product in the cardiovasculature of interest, and determine a blood flow characteristic in the cardiovasculature of interest based on a ratio of mass of the contrast agent in the cardiovasculature of interest to the a total sum of contrast agent concentration time product in the cardiovasculature of interest.

[0293] 12. The system of aspect 11, further comprising including a baseline data point extracted from the CT or MRI image data in the measured time-enhancement curve and the processor executing instruction to denoise the measured time-enhancement curve, the CT or MRI image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent to provide the baseline data point. 02 OCTOBER 2025 (02.10.2025)

[0294] 13. The system of aspect 11, wherein the machine learning model is trained with training inputs of the image feature of the measured time-enhancement curve and at least one non-image feature, and associated with a ground truth value of an expected area under the simulated time-enhancement curve.

[0295] 14. The system of aspect 11, wherein the image feature is an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope.

[0296] 15. The system of aspect 11, wherein the image feature is a measured value of an area under an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope; and calculating the area under the at least partial time-enhancement curve.

[0297] 16. The system of aspect 11, further comprising providing a second image feature based on enhancement of signal intensity, thickness of a wall of a cardiovasculature of interest, size of a cardiovasculature of interest, diameter of a cardiovasculature of interest, morphology of a cardiovasculature of interest, location of sampling site in a cardiovascular of interest, or degree of stenosis in a cardiovasculature of interest.

[0298] 17. The system of aspect 11, wherein the non-image feature is based on age, sex, weight, heart rate, blood pressure, x-ray tube voltage, x-ray tube current, gradient pulse sequence, disease state, contrast-injection rate, contrast agent volume, or contrast agent concentration.

[0299] 18. The system of aspect 11, wherein the first set of image acquisition parameters is different than the simulated second set of image acquisition parameters according to at least one parameter selected from the group consisting of: scan axis orientation relative to a longitudinal axis of the cardiovasculature of interest, anatomical location of scan, hyperemic or rest condition of a subject, time duration of scan, x-ray tube voltage, x-ray tube current, gradient pulse sequence, contrastinjection rate, contrast agent volume, or contrast agent concentration.

[0300] 19. The system of aspect 11, wherein the processor executes instructions to determine the blood flow characteristic comprises determining an absolute flow velocity using Reynolds transport theorem by determining a time rate of change of contrast agent mass in the cardiovasculature of interest based on a time rate of change of signal enhancement measured from an at least partial timeenhancement curve, a factor for converting signal enhancement to contrast agent concentration, a 02 OCTOBER 2025 (02.10.2025) predetermined volume value of injected contrast agent, and a time value of a duration of time that signal intensity is higher than a predetermined baseline threshold level; determining density of the contrast agent in the cardiovasculature of interest based on the ratio defined in aspect 11, a time value of a duration of time that signal intensity is higher than a predetermined baseline threshold level, and the mass of contrast agent in the cardiovasculature of interest; and determining an area measured from the image data in the cardiovasculature of interest.

[0301] 20. The system of aspect 11, wherein the processor executes instructions to determine the blood flow characteristic comprises determining a flow pressure by applying Bernoulli’s equation expressed as

[0302] 21. in illustrative aspect 21, a computer implemented method for blood flow imaging comprising: obtaining CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; extracting at least one image feature from the CT or MRI image data; providing the at least one image feature and an associated at least one non-image feature to a machine learning model to generate a predicted value of area under a time-enhancement curve of the contrast agent within the cardiovasculature of interest, the machine learning model trained with training inputs of the at least one image feature with the at least one non-image feature, and associated with an area under a time-enhancement curve value as ground truth; converting the predicted value of area under the time-enhancement curve to a time rate of change of contrast agent concentration in the cardiovasculature of interest; determining a blood flow characteristic in the cardiovasculature of interest based on a ratio of mass of the contrast agent in the cardiovasculature of interest to the time rate of change of contrast agent concentration in the cardiovasculature of interest.

[0303] 22. The method of aspect 21, further comprising extracting a baseline data point from the CT or MRI image data prior to extracting the at least one image feature, the CT or MRI image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent to provide the baseline data point. 02 OCTOBER 2025 (02.10.2025)

[0304] 23. The method of aspect 21, wherein the machine learning model is trained with training inputs of the at least one image feature extracted from a time-enhancement curve generated from the image data and the at least one non-image feature including a least a heart rate or blood pressure, and associated with a ground truth value of an expected area under a simulated time-enhancement curve.

[0305] 24. The method of aspect 21, wherein the image feature is an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope.

[0306] 25. The method of aspect 21, wherein the image feature is a measured value of an area under an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope; and calculating the area under the at least partial time-enhancement curve.

[0307] 26. The method of aspect 21, further comprising providing a second image feature based on enhancement of signal intensity, thickness of a wall of a cardiovasculature of interest, size of a cardiovasculature of interest, diameter of a cardiovasculature of interest, morphology of a cardiovasculature of interest, location of sampling site in a cardiovascular of interest, or degree of stenosis in a cardiovasculature of interest.

[0308] 27. The method of aspect 21, wherein the at least one non-image feature comprises a first nonimage feature that is a heart rate or blood pressure, and a second non-image feature based on age, sex, weight, heart rate, blood pressure, x-ray tube voltage, x-ray tube current, gradient pulse sequence, disease state, contrast-injection rate, contrast agent volume, or contrast agent concentration.

[0309] 28. The method of aspect 21, wherein the predicted value simulates a second set of image acquisition parameters that are different than a first set of image acquisition parameters used to acquire the CT or MRI image data according to at least one parameter selected from the group consisting of: scan axis orientation relative to a longitudinal axis of the cardiovasculature of interest, anatomical location of scan, hyperemic or rest condition of a subject, time duration of scan, x-ray tube voltage, x-ray tube current, gradient pulse sequence, contrast-injection rate, contrast agent volume, or contrast agent concentration. 02 OCTOBER 2025 (02.10.2025)

[0310] 29. The method of aspect 21, wherein the scan capturing the image data is a perfusion scan, a bolus tracking (BT) scan, a test bolus (TB) scan, a diagnostic angiography scan, or a combination thereof.

[0311] 30. The method of aspect 21, wherein determining the blood flow characteristic comprises: determining an absolute flow velocity using Reynolds Transport Theorem; determining a flow pressure by applying Bernoulli’s equation; or determining a flow rate by applying Indicator- Dilution Principle.

[0312] 31. in illustrative aspect 31 , a computer implemented method for blood flow imaging comprising: obtaining image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; providing the image data to a machine learning model to predict an area under a timeenhancement curve of the contrast agent within the cardiovasculature of interest; selecting a region of interest within the cardiovasculature of interest in the image data; determining a blood flow characteristic through the region of interest based on the area under the time-enhancement curve.

[0313] 32. The method of aspect 31, wherein the image data comprises an increase phase or a decline phase.

[0314] 33. The method of aspect 32, wherein the image data comprise both an increase phase and a decline phase.

[0315] 34. The method of aspect 33, wherein the image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent.

[0316] 35. The method of any one of aspects 31 to 34, wherein the image data is a pseudo truncated dynamic perfusion scan.

[0317] 36. The method of aspect 35, wherein the pseudo truncated dynamic perfusion scan is a compilation of a bolus tracking (BT) scan capturing images at a plurality of time points and a diagnostic angiographic scan capturing an image at a single time point.

[0318] 37. The method of any one of aspects 31 to 34, wherein the image data is a genuine dynamic perfusion scan.

[0319] 38. The method of aspect 37, wherein the genuine dynamic perfusion scan is a genuine truncated dynamic perfusion scan, such that scan duration covers less than a complete increase phase and complete decline phase of a first-pass circulation of the contrast agent within the cardiovasculature of interest, and the scan is a test bolus (TB) scan or a diagnostic angiographic scan. 02 OCTOBER 2025 (02.10.2025)

[0320] 39. The method of aspect 37, wherein the genuine dynamic perfusion scan is a genuine full dynamic perfusion scan, such that scan duration covers a complete increase phase and complete decline phase of a first-pass circulation of the contrast agent within the cardiovasculature of interest, and the scan is a test bolus (TB) scan or a diagnostic angiographic scan.

[0321] 40. The method of any one of aspects 31 to 39, wherein the machine learning model is trained by extraction of an image feature from the image data.

[0322] 41. The method of aspect 40, wherein the image feature is based on enhancement of signal intensity, thickness of a wall of a cardiovasculature of interest, size of a cardiovasculature of interest, diameter of a cardiovasculature of interest, morphology of a cardiovasculature of interest, location of sampling site in a cardiovascular of interest, or degree of stenosis in a cardiovasculature of interest.

[0323] 42. The method of any one of aspects 31 to 39, wherein the machine learning model is trained by accompanying the image data with a non-image feature.

[0324] 43. The method of aspect 42, wherein the non-image feature is based on age, sex, weight, heart rate, blood pressure, x-ray tube voltage, x-ray tube current, gradient pulse sequence, disease state, contrast-injection rate, contrast agent volume, or contrast agent concentration.

[0325] 44. The method of any one of aspects 31 to 43, wherein determining the blood flow characteristic comprises determining an absolute flow velocity using Reynolds Transport Theorem.

[0326] 45. The method of any one of aspects 31 to 43, wherein determining the blood flow characteristic comprises determining a flow pressure by applying Bernoulli’s equation.

[0327] 46. The method of any one of aspects 31 to 43, wherein determining the blood flow characteristic comprises determining a flow rate by applying Indicator-Dilution Principle.

[0328] 47. The method of any one of aspects 31 to 46, further comprising acquiring scan data of the cardiovasculature of interest from a X-ray based scan or a MRI scan, and reconstructing image data based on the scan data.

[0329] 48. The method of aspect 47, wherein the scan data is acquired from a CT scan.

[0330] 49. The method of aspect 47, wherein the scan data is acquired from a MRI scan.

[0331] 50. The method of aspect 47, wherein the scan data is acquired from a scan having an elapsed time of greater than 5 seconds.

[0332] 51. The method of any one of aspects 47 to 50, further comprising: administering the contrast agent to a subject; and scanning the subject to obtain the scan data, the scan data capturing at least a portion of an increase phase or a decline phase of a contrast agent in a cardiovasculature of interest. 02 OCTOBER 2025 (02.10.2025)

[0333] 52. in illustrative aspect 52, a computer implemented method for blood flow imaging based on predicting an area under a time-enhancement curve, the method comprising: obtaining image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; providing the image data to a machine learning model to predict an area under a timeenhancement curve of the contrast agent within the cardiovasculature of interest, the predicted area under the time-enhancement curve representing the total sum of contrast agent concentration time product within the cardiovasculature of interest.

[0334] 53. The method of aspect 52, wherein the image data comprises an increase phase or a decline phase.

[0335] 54. The method of aspect 53, wherein the image data comprise both an increase phase and a decline phase.

[0336] 55. The method of aspect 54, wherein the image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent.

[0337] 56. The method of any one of aspects 52 to 55, wherein the image data is a pseudo truncated dynamic perfusion scan.

[0338] 57. The method of aspect 56, wherein the pseudo truncated dynamic perfusion scan is a compilation of a bolus tracking (BT) scan capturing images at a plurality of time points and a diagnostic angiographic scan capturing an image at a single time point.

[0339] 58. The method of any one of aspects 52 to 55, wherein the image data is a genuine dynamic perfusion scan.

[0340] 59. The method of aspect 58, wherein the genuine dynamic perfusion scan is a genuine truncated dynamic perfusion scan, such that scan duration covers less than a complete increase phase and complete decline phase of a first-pass circulation of the contrast agent within the cardiovasculature of interest.

[0341] 60. The method of aspect 58, wherein the genuine dynamic perfusion scan is a genuine full dynamic perfusion scan, such that scan duration covers a complete increase phase and complete decline phase of a first-pass circulation of the contrast agent within the cardiovasculature of interest.

[0342] 61. The method of any one of aspects 52 to 60, wherein the machine learning model is trained by extraction of an image feature from the image data. 02 OCTOBER 2025 (02.10.2025)

[0343] 62. The method of aspect 61, wherein the image feature is based on enhancement of signal intensity, thickness of a wall of a cardiovasculature of interest, size of a cardiovasculature of interest, diameter of a cardiovasculature of interest, morphology of a cardiovasculature of interest, or degree of stenosis in a cardiovasculature of interest.

[0344] 63. The method of any one of aspects 52 to 60, wherein the machine learning model is trained by accompanying the image data with a non-image feature.

[0345] 64. The method of aspect 63, wherein the non-image feature is based on age, sex, weight, heart rate, blood pressure, x-ray tube voltage, x-ray tube current, gradient pulse sequence, disease state, contrast-injection rate, or contrast agent concentration.

[0346] 65. The method of any one of aspects 52 to 64, further comprising determining a blood flow characteristic by determining an absolute flow velocity based on the predicted area under the timeenhancement curve and Reynolds transport theorem.

[0347] 66. The method of any one of aspects 52 to 64, further comprising determining a blood flow characteristic by determining a flow pressure based on the predicted area under the timeenhancement curve and Bernoulli’s equation.

[0348] 67. The method of any one of aspects 52 to 66, wherein the plurality of corresponding images is greater than 5 images.

[0349] 68. The method of any one of aspects 52 to 67, further comprising acquiring scan data of the cardiovasculature of interest from a X-ray based scan or a MRI scan, and reconstructing image data based on the scan data.

[0350] 69. The method of aspect 68, wherein the scan data is acquired from a CT scan.

[0351] 70. The method of aspect 68, wherein the scan data is acquired from a MRI scan.

[0352] 71. The method, of aspect 68, wherein the scan data is acquired from a scan having an elapsed time of greater than 5 seconds.

[0353] 72. The method of any one of aspects 68 to 71, further comprising: administering the contrast agent to a subject; and scanning the subject to obtain the scan data, the scan data capturing at least a portion of an increase phase or a decline phase of a contrast agent in a cardiovasculature of interest.

[0354] 73. The method of any one of aspects 1 to 10 and 21 to 72, further comprising determining a clinical condition of a cardiovasculature of interest based on an area under a time -enhancement curve or blood flow characteristic.

[0355] 74. in illustrative aspect 74, a system for blood flow imaging comprising: 02 OCTOBER 2025 (02.10.2025) a memory for storing image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; a processor configured to execute the method steps of any one of aspects 1 to 10 and 21 to

[0356] 73.

[0357] 75. in illustrative aspect 76, a computer readable medium embodying computer readable code for executing the method of any one of aspects 1 to 10 and 21 to 73.

[0358] 76. in illustrative aspect 76, a computer implemented method for assessing functionally significant stenosis comprising: obtaining CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; extracting at least one image feature from the CT or MRI image data, the at least one image feature extracted from an at least partial time-enhancement curve generated from the CT or MRI image data; providing the at least one image feature and an associated at least one non-image feature to a machine learning model to generate a prediction of a functionally significant stenosis within the cardiovasculature of interest, the machine learning model trained with training inputs of the at least one image feature with the at least one non-image feature, and associated with a binary (YES / NO) assessment of functionally significant stenosis as ground truth.

[0359] 77. The method of aspect 76, further comprising extracting a baseline data point from the CT or MRI image data prior to extracting the at least one image feature, the CT or MRI image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent to provide the baseline data point.

[0360] 78. The method of aspect 76, further comprising providing a value characterizing severity of stenosis to the machine learning model.

[0361] 79. The method of aspect 78, wherein the value characterizing severity of stenosis is a lumen diameter or a lumen cross-sectional area caused by the stenosis in a stenosed segment of the cardiovasculature of interest.

[0362] 80. The method of aspect 78, wherein the value characterizing severity of stenosis is determined as a ratio of reduction in lumen diameter caused by the stenosis in a stenosed segment of the cardiovasculature of interest relative to a lumen diameter in an adjacent segment that does not exhibit stenosis. 02 OCTOBER 2025 (02.10.2025)

[0363] 81. The method of aspect 78, wherein the value characterizing severity of stenosis is determined as a ratio of reduction in lumen cross-sectional area caused by the stenosis in a stenosed segment of the cardiovasculature of interest relative to a lumen cross-sectional area in an adjacent segment that does not exhibit stenosis.

[0364] 82. The method of aspect 78, wherein the value characterizing severity of stenosis is a length of stenosis measured from entry to exit of a stenosed segment of the cardiovasculature of interest.

[0365] 83. The method of aspect 78, wherein the value characterizing severity of stenosis is based on identifying a location of stenosis within the cardiovasculature of interest.

[0366] 84. The method of aspect 76, wherein the machine learning model is trained with training inputs of the at least one image feature extracted from a time -enhancement curve generated from the image data and the at least one non-image feature including a least one of patient’s sex, age, height, weight, body-mass index, heart rate, blood pressure, symptoms, comorbidities; contrast concentration, contrast volume, contrast injection rate; X-ray energy, X-ray tube current, MR pulse sequence, and associated with a ground truth of a binary (YES / NO) value of functionally significant stenosis as ground truth.

[0367] 85. The method of aspect 76, wherein the image feature is an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope, and providing at least one value based on the at least partial time-enhancement curve to a second machine learning model.

[0368] 86. The method of aspect 76, wherein the image feature is a measured value of an area under an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope; and calculating the area under the at least partial time-enhancement curve.

[0369] 87. The method of aspect 76, wherein the image feature includes one or more of: at least one timeenhancement curve in the cardiovasculature of interest, at least one time-enhancement curve in the downstream cardiovasculature supplied by the cardiovasculature of interest, flow rate of the cardiovasculature of interest or the downstream cardiovasculature, flow velocity of the cardiovasculature of interest or the downstream cardiovasculature, blood flow reserve (BFR) of the cardiovasculature of interest or the downstream cardiovasculature, fractional flow reserve (FFR) of the cardiovasculature of interest or the downstream cardiovasculature, blood flow of the 02 OCTOBER 2025 (02.10.2025) cardiovasculature of interest or the downstream cardiovasculature, iodine concentration in the cardiovasculature of interest or the downstream cardiovasculature.

[0370] 88. in illustrative aspect 88, a computer implemented method for assessing functionally significant stenosis comprising: obtaining CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; extracting at least one image feature from the CT or MRI image data; determining a blood flow characteristic from the at least one image feature according to the method of any one of aspects 1 to 10 and 21 to 73; providing the blood flow characteristic and an associated at least one non-image feature to a machine learning model to generate a prediction of a functionally significant stenosis within the cardiovasculature of interest, the machine learning model trained with training inputs of the at least one image feature with the at least one non-image feature, and associated with a binary (YES / NO) assessment of functionally significant stenosis as ground truth.

[0371] 89. The method of any one of aspects 76 to 88, further comprising determining with a machine learning model is whether a functionally significant stenosis is contributed by the cardiovasculature of interest or its downstream cardiovasculature microcirculation or both.

[0372] 90. The method of aspect 89, wherein the machine learning model is trained based on at least one non-image data and more than one image feature.

[0373] 91. The method of aspect 90, wherein the image feature includes one or more of: at least one timeenhancement curve in the cardiovasculature of interest, at least one time-enhancement curve in the downstream cardiovasculature supplied by the cardiovasculature of interest, flow rate of the cardiovasculature of interest or the downstream cardiovasculature, flow velocity of the cardiovasculature of interest or the downstream cardiovasculature, blood flow reserve (BFR) of the cardiovasculature of interest or the downstream cardiovasculature, fractional flow reserve (FFR) of the cardiovasculature of interest or the downstream cardiovasculature, blood flow of the cardiovasculature of interest or the downstream cardiovasculature, iodine concentration in the cardiovasculature of interest or the downstream cardiovasculature.

[0374] 92. in illustrative aspect 92, a system for assessing functionally significant stenosis comprising: a memory for storing image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; 02 OCTOBER 2025 (02.10.2025) a processor configured to execute the method steps of any one of aspects 76 to 91.

[0375] 93. in illustrative aspect 93, a computer readable medium embodying computer readable code for executing the method of any one of aspects 76 to 91.

[0376] 94. in illustrative aspect 94, a computer implemented method for selecting a peak contrast enhanced image from a plurality of corresponding images, comprising: obtaining a dynamic contrast enhanced CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest, the plurality of corresponding images including a plurality of image slices, and each plurality of image slices including a plurality of time points; providing the plurality of corresponding images to a machine learning model, the machine learning model selecting at least one image exhibiting peak contrast enhancement within the plurality of corresponding imagea, and the machine learning model outputting selecting at least one image exhibiting peak contrast enhancement; the machine learning model trained with training inputs of a plurality of corresponding dynamic contrast enhanced training images, and associated with a labelled image selected from the plurality of corresponding dynamic contrast enhanced training images, and the labelled image presented as ground truth.

[0377] 95. in illustrative aspect 95, a computer implemented method for calculating a distribution of contrast agent mass in a plurality of a coronary cardiovasculature of interest, comprising: obtaining CT or MRI contrast enhanced image data comprising a plurality of corresponding images, the plurality of corresponding images including a plurality of image slices at the same time point; determining lumen radius of the cardiovasculature of interest for each of the plurality of image slices; determining number of arteries branching from the cardiovasculature of interest as displayed within the plurality of image slices; determining a distribution of contrast agent mass in at least a first target slice upstream of a region of interest in the cardiovasculature of interest and a second target slice downstream of the region of interest based on a total mass of injected contrast agent, the determined lumen radius, the determined number of branch arteries, the percentage of the injected contrast agent entering the whole coronary system in either a rest or hyperemic condition. 02 OCTOBER 2025 (02.10.2025)

[0378] Embodiments described herein are intended for illustrative purposes without any intended loss of generality. Still further variants, modifications and combinations thereof are contemplated and will be recognized by the person of skill in the art. Accordingly, the foregoing detailed description is not intended to limit scope, applicability, or configuration of claimed subject matter.

Claims

02 OCTOBER 2025 (02.10.2025)WHAT IS CLAIMED IS:

1. A computer implemented method for assessing functionally significant stenosis comprising: obtaining CT or MRI image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; extracting at least one image feature from the CT or MRI image data, the at least one image feature extracted from an at least partial time-enhancement curve generated from the CT or MRI image data; providing the at least one image feature and an associated at least one non-image feature to a machine learning model to generate a prediction of a functionally significant stenosis within the cardiovasculature of interest, the machine learning model trained with training inputs of the at least one image feature with the at least one non-image feature, and associated with a binary (YES / NO) assessment of functionally significant stenosis as ground truth.

2. The method of claim 1, further comprising extracting a baseline data point from the CT or MRI image data prior to extracting the at least one image feature, the CT or MRI image data comprises at least one image capturing the cardiovasculature of interest prior to entry of the contrast agent to provide the baseline data point.

3. The method of claim 1 or 2, further comprising providing a value characterizing severity of stenosis to the machine learning model.

4. The method of claim 3, wherein the value characterizing severity of stenosis is a lumen diameter or a lumen cross-sectional area caused by the stenosis in a stenosed segment of the cardiovasculature of interest.

5. The method of claim 3, wherein the value characterizing severity of stenosis is determined as a ratio of reduction in lumen diameter caused by the stenosis in a stenosed segment of the cardiovasculature of interest relative to a lumen diameter in an adjacent segment that does not exhibit stenosis.

6. The method of claim 3, wherein the value characterizing severity of stenosis is determined as a ratio of reduction in lumen cross-sectional area caused by the stenosis in a stenosed segment of the cardiovasculature of interest relative to a lumen cross-sectional area in an adjacent segment that does not exhibit stenosis.

7. The method of claim 3, wherein the value characterizing severity of stenosis is a length of stenosis measured from entry to exit of a stenosed segment of the cardiovasculature of interest.02 OCTOBER 2025 (02.10.2025)8. The method of claim 3, wherein the value characterizing severity of stenosis is based on identifying a location of stenosis within the cardiovasculature of interest.

9. The method of any one of claims 1 to 8, wherein the machine learning model is trained with training inputs of the at least one image feature extracted from a time-enhancement curve generated from the image data and the at least one non-image feature including a least one of patient’s sex, age, height, weight, body-mass index, heart rate, blood pressure, symptoms, comorbidities; contrast concentration, contrast volume, contrast injection rate; X-ray energy, X-ray tube current, MR pulse sequence, and associated with a ground truth of a binary (YES / NO) value of functionally significant stenosis as ground truth.

10. The method of any one of claims 1 to 8, wherein the image feature is an at least partial timeenhancement curve and extracting the image feature comprises: generating the at least partial timeenhancement curve of the contrast agent based on the image data, the at least partial timeenhancement curve having at least an upslope or a downslope, and providing at least one value based on the at least partial time-enhancement curve to a second machine learning model.

11. The method of any one of claims 1 to 8, wherein the image feature is a measured value of an area under an at least partial time-enhancement curve and extracting the image feature comprises: generating the at least partial time-enhancement curve of the contrast agent based on the image data, the at least partial time-enhancement curve having at least an upslope or a downslope; and calculating the area under the at least partial time-enhancement curve.

12. The method of any one of claims 1 to 8, wherein the image feature includes one or more of: at least one time-enhancement curve in the cardiovasculature of interest, at least one timeenhancement curve in the downstream cardiovasculature supplied by the cardiovasculature of interest, flow rate of the cardiovasculature of interest or the downstream cardiovasculature, flow velocity of the cardiovasculature of interest or the downstream cardiovasculature, blood flow reserve (BFR) of the cardiovasculature of interest or the downstream cardiovasculature, fractional flow reserve (FFR) of the cardiovasculature of interest or the downstream cardiovasculature, blood flow of the cardiovasculature of interest or the downstream cardiovasculature, iodine concentration in the cardiovasculature of interest or the downstream cardiovasculature.

13. A system for assessing functionally significant stenosis comprising: a memory for storing image data comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest; a processor configured to execute the method steps of any one of claims 1 to 12.02 OCTOBER 2025 (02.10.2025)14. A computer readable medium embodying computer readable code for executing the method of any one of claims 1 to 12.

15. A computer implemented method for selecting a peak contrast enhanced image from a plurality of corresponding images, comprising: obtaining a dynamic contrast enhanced CT or MRI image data set comprising a plurality of corresponding images capturing at least a portion of one or both an increase phase and a decline phase of a contrast agent in a cardiovasculature of interest, the plurality of corresponding images including a plurality of image slices, and each plurality of image slices including a plurality of time points; providing the plurality of corresponding images to a machine learning model, the machine learning model selecting at least one image exhibiting peak contrast enhancement within the plurality of corresponding imagea, and the machine learning model outputting the selected at least one image exhibiting peak contrast enhancement; the machine learning model trained with training inputs of a plurality of corresponding dynamic contrast enhanced training images, and associated with a labelled image selected from the plurality of corresponding dynamic contrast enhanced training images, and the labelled image presented as ground truth.

16. A computer implemented method for calculating a distribution of contrast agent mass in a plurality of a coronary cardiovasculature of interest, comprising: obtaining CT or MRI contrast enhanced image data comprising a plurality of corresponding images, the plurality of corresponding images including a plurality of image slices at the same time point; determining lumen radius of the cardiovasculature of interest for each of the plurality of image slices; determining number of arteries branching from the cardiovasculature of interest as displayed within the plurality of image slices; determining a distribution of contrast agent mass in at least a first target slice upstream of a region of interest in the cardiovasculature of interest and a second target slice downstream of the region of interest based on a total mass of injected contrast agent, the determined lumen radius, the determined number of branch arteries, the percentage of the injected contrast agent entering the whole coronary system in either a rest or hyperemic condition.

Images