System and Methods of Prediction of Ischemic Brain Tissue Fate from Multi-Phase CT-Angiography in Patients with Acute Ischemic Stroke using Machine Learning

Abstract

The invention relates to systems and methods for predicting ischemic brain tissue fate from multi-phase CT-angiography. More specifically, systems and methods are provided that enable meaningful prediction of core, penumbra and perfusion from mCTA images using software that has been trained via machine learning to interpret mCTA images.

Claims

1.-41. (canceled)

42. A method of deriving and presenting information useful in diagnosing medium vessel occlusion (MeVO) in a current patient comprising the steps of: from a plurality of CT images showing hypoperfused regions of the current patient; i. quantifying a hypoperfused tissue volume in the current patient; ii. comparing the hypoperfused tissue volume from step i to threshold volume parameters defining a MeVO event and determining if the hypoperfused tissue matches volume parameters of a MeVO event; and, iii. if a MeVO event is determined, display a MeVO event determination.

43. The method as in claim 42 wherein steps i and ii includes quantifying a hypoperfused tissue shape in the current patient and comparing the hypoperfused tissue shape to threshold shape parameters defining a MeVO event and determining if the hypoperfused tissue shape matches shape parameters of a MeVO event.

44. The method as in claim 42 wherein steps i and ii includes quantifying a hypoperfused tissue location in the current patient and comparing the hypoperfused tissue location to threshold location parameters defining a MeVO event and determining if the hypoperfused tissue location matches location parameters of a MeVO event.

45. The method as in claim 42 wherein steps i and ii includes quantifying involved cortex.

46. The method as in claim 42 wherein steps i and ii includes quantifying hypoperfused tissue confluence in the current patient and comparing the hypoperfused tissue confluence to hypoperfused tissue confluence parameters defining a MeVO event and determining if the hypoperfused tissue confluence matches hypoperfused tissue confluence of a MeVO event.

47. The method as in claim 44 further comprising the steps of correlating the hypoperfused tissue location to corresponding hypoperfused locations from historical patient data wherein historical patient data includes data marking past MeVO events; determining a best fit of historical patient image data and marking current patient images with MeVO location data derived from the historical patient image data.

48. The method as in claim 47 wherein the historical patient data with past MeVO events includes data quantifying proximal voxel location relevant to a past MeVO event within a past patient record.

49. The method as in claim 47 wherein historical patient data records have been previously analyzed to derive 2D and/or 3D relationships between level 1-3 vessels.

50. The method as in claim 49 wherein the historical patient data records have been previously analyzed to define volumes of tissue as level 1, level 2, or level 3 tissue, and wherein each volume of level 1, level 2, or level 3 tissue has at least one, equal, distal or proximal relationship with an adjacent volume of tissue.

51. The method as in claim 50 further comprising the step of, after step iii, examining changes in contrast densities in adjacent proximal volumes across two or more phases of CTA images for the current patient and based on those changes marking changes in contrast density as normal flow or abnormal flow.

52. The method as in claim 51 further comprising the step of discarding volumes showing normal flow from further analysis.

53. The method as in claim 52 further comprising the step of utilizing volumes showing normal flow as a baseline for contrast density analysis.

54. The method as in claim 53 further comprising the step of marking volumes showing abnormal flow for further analysis.

55. The method as in claim 54 further comprising the step of analyzing zones where contrast abruptly transitions from no contrast to significant contrast between adjacent images or vice versa to identify vessels of interest.

56. The method as in claim 55 further comprising the step of marking and displaying zones where contrast abruptly transitions on CTA images of the current patient.

57. The method as in claim 42 further comprising the steps of providing at least one output selected from any one of or a combination of: a) presence or not of MeVO; b) zone of interest marking and c) vessel of interest.

58. A method of deriving and presenting information useful in diagnosing medium vessel occlusion (MeVO) in a current patient comprising the steps of: from a plurality of CTA images showing at least one hypoperfused region of the current patient; i. identifying the at least one hypoperfused region and correlating the at least one hypoperfused regions to one or more corresponding hypoperfused regions from within historical patient data; and, ii. deriving and identifying immediately proximal vessels/zones in the current patient based on best match(s) to the historical patient data and marking the proximal vessel/zones as predicted MeVO locations on current patient CT images.

59. The method as in claim 58 where the CT images are mCTA images.

60. The method as in claim 58 further comprising the steps of quantifying a hypoperfused tissue shape in the current patient and comparing the hypoperfused tissue shape to threshold shape parameters defining a MeVO event and determining if the hypoperfused tissue shape matches shape parameters of a MeVO event.

61. The method as in claim 58 further comprising the steps of quantifying a hypoperfused tissue location in the current patient and comparing the hypoperfused tissue location to threshold location parameters defining a MeVO event and determining if the hypoperfused tissue location matches location parameters of a MeVO event.

62. The method as in claim 58 further comprising the steps of quantifying involved cortex.

63. The method as in claim 58 further comprising the steps of quantifying hypoperfused tissue confluence in the current patient and comparing the hypoperfused tissue confluence to hypoperfused tissue confluence parameters defining a MeVO event and determining if the hypoperfused tissue confluence matches hypoperfused tissue confluence of a MeVO event.

64. The method as in claim 58 further comprising the steps of correlating the hypoperfused tissue location to corresponding hypoperfused locations from historical patient data wherein historical patient data includes data marking past MeVO events; determining a best fit of historical patient image data and marking current patient images with MeVO location data derived from the historical patient image data.

65. The method as in claim 64 wherein the historical patient data with past MeVO events includes data quantifying proximal voxel location relevant to a past MeVO event within a past patient record.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0123] Various objects, features and advantages of the invention will be apparent from the following description of particular embodiments of the invention, as illustrated in the accompanying drawings. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of various embodiments of the invention.

[0124] Similar reference numerals indicate similar components.

[0125] FIG. 1 is a graph showing typical time density curves and acquisition of mCTA images.

[0126] FIG. 1A is a graph showing 3 typical mCTA data points vs. time in accordance with one embodiment of the invention.

[0127] FIG. 2 is a representative graph showing a typical impulse residue function calculated from a deconvolution algorithm in CTP software in accordance with the prior art.

[0128] FIG. 3 is a flowchart showing an overall training and testing strategy of a machine learning model to predict ischemic infarct, penumbra and perfusion status in accordance with one embodiment of the invention.

[0129] FIG. 3A is a flowchart showing a training and testing strategy of machine learning models to predict (a) ischemic infarct, penumbra and (b) perfusion status in accordance with one embodiment of the invention.

[0130] FIG. 4 is an illustration showing d.sub.1, d.sub.2 and dd patterns in accordance with one embodiment the invention showing representative contrast delay for different patterns.

[0131] FIG. 4A is an illustration showing d.sub.1, d.sub.2 and dd patterns as in FIG. 4 overlaid with representatives zones for color coding in accordance with one embodiment of the invention.

[0132] FIG. 48 is a representative graph of data points from a set of mCTA images having p0-p3 data points and showing how tissue intensity curves (TICs) from past studies have been matched to the data points.

[0133] FIG. 4C is a figure illustrating a process of machine learning in accordance with one embodiment of the invention.

[0134] FIG. 5 is a flowchart showing details of training and testing each of three machine learning models.

[0135] FIG. 5A is a flowchart showing details of patient inclusion.

[0136] FIG. 6 is an overview of input images and input data including patient meta data and input artery/vein functions together with representative output data that may be displayed after modelling and prediction.

[0137] FIGS. 6A and 68 are representative mCTA and prediction maps from two case studies.

[0138] FIG. 7 is a representative image showing anterior, middle, and posterior arteries at levels 2 and 3 in accordance with the prior art.

[0139] FIGS. 7A-7H are a sequence of images in a patient showing a progression of a diagnostic process using mCTA prediction maps and MeVO prediction tools together with conventional imaging. FIG. 7A shows an unremarkable non-contrast CT scan in a patient having an acute stroke with aphasia; FIG. 7B shows the area of the brain that is hypoperfused resulting in the patients symptoms; the map could be created using an mCTA prediction map; FIG. 7C shows a zone of interest predicted using the MeVO tool; FIG. 7D shows identification of a particular vessel of interest following further processing with the MeVO tool; FIG. 7E shows vessel occlusion from conventional angiography; FIG. 7F shows a follow-up image following successful reperfusion; FIG. 7G is a flowchart outlining a process of providing useful information to a physician diagnosing MeVO in accordance with one embodiment of the invention; and FIG. 7H shows a representative 2D and partial 3D map of the relationship of L1-L3 voxels.

[0140] FIG. 8 shows multiphase CTA predicted infarct probability maps compared to CTP (Tmax) map and follow up infarct. There are 3 examples, each labelled as per rows.

[0141] FIG. 8A show Bland-Altman plots of (a) mCTA penumbra volume predicted using the proposed model vs. follow up infarct volume for 44 patients who did not achieve acute reperfusion (mTICI 0/1); (b) mCTA infarct volume predicted using the proposed model vs. follow up infarct volume for 100 patients who had acute reperfusion (mTICI 2b/2c/3); and (c) mCTA perfusion volume predicted using the proposed model vs. time-dependent Tmax predicted infarct volume for all 144 patients in the test cohort.

DETAILED DESCRIPTION OF THE INVENTION

[0142] With reference to the figures, systems, and methods for predicting ischemic brain tissue fate from multi-phase CT-angiography (mCTA) are described. More specifically, systems and methods are described that enable meaningful prediction of core, penumbra and perfusion from mCTA images using software that has been trained via machine learning to interpret mCTA images.

[0143] Terms used herein have definitions that are reasonably inferable from the drawings and description.

Introduction

[0144] Various aspects of the invention will now be described with reference to the figures. For the purposes of illustration, components depicted in the figures are not necessarily drawn to scale. Instead, emphasis is placed on highlighting the various contributions of the components to the functionality of various aspects of the invention. A number of possible alternative features are introduced during the course of this description. It is to be understood that, according to the knowledge and judgment of persons skilled in the art, such alternative features may be substituted in various combinations to arrive at different embodiments of the present invention.

[0145] A primary objective is to obtain from a relatively small number of mCTA images (typically 3-phases of CTA images), a meaningful prediction of core, penumbra and perfusion using a methodology and software that has been trained to interpret mCTA images.

[0146] As noted above, mCTA does not have the granularity of data that a CTP study provides. Hence, without additional boundaries and/or knowledge to interpret the mCTA data, the mCTA images are, on their own, not effective in accurately quantifying core, penumbra and perfusion.

Data Used in Building a Prediction Model

[0147] The inventors have determined that by using image data from past CTP studies, and specifically from a first groups of patients that have undergone: [0148] a. CTP and no reperfusion treatment, [0149] b. CTP and reperfusion treatment, and [0150] c. follow-up Non-Contrast CT (NCCT)/Diffusion weighted (DW) MRI,
as well as image data from past mCTA studies from a second group of patients that have undergone: [0151] d. mCTA and no perfusion treatment, [0152] e. mCTA and reperfusion treatment; and, [0153] f. follow-up Non-Contrast CT (NCCT)/Diffusion weighted (DV) MRI,
a database of scenarios that shows the “start to finish” outcomes for these groups of patients can be utilized to build and test prediction models.

[0154] In accordance with the invention, models have been developed and trained with the objective of being able to interpret current mCTA images in a clinical setting to create clinically meaningful core, penumbra and perfusion maps at the time treatment decisions are being made.

[0155] For background, an ischemic stroke patient that has gone through a CTP and/or mCTA diagnosis and treatment protocol may have an outcome that is anywhere between a full recovery (no core) or poor recovery (significant core). This same patient may have been subjected to either a reperfusion treatment or no reperfusion treatment.

[0156] It has been determined that by studying the diagnostic and follow-up images of a number of these patients, patterns of effects can be observed across the population. For example, past data from a patient cohort (eg. 40 patients having an M1 occlusion) having undergone CTP and follow-up studies, the CTP studies will have determined a range of Tmax, CBV and CBF values that enabled CTP maps to be created that showed core, penumbra and perfusion predictions for these patients. These patients will have undergone (or not) treatments as well as follow-up imaging that verifies an outcome. Similarly, for a different patient cohort, mCTA studies have been used to make treatment decisions. Again, treatments will have been undertaken (or not) as well as follow-up imaging that verifies an outcome.

[0157] By using this data from past patients within models and training the models to interpret past mCTA images, it has been determined that at the time of diagnosis and the time that treatment decisions are being made with a current patient, these models can be utilized to fit mCTA data within the models to create predictive maps (like those obtained by CTP) that can be utilized by the physician to give an idea of the likelihood of success of a treatment. For example, a decision to treat or not to treat may be made given the relative likelihood of success based on a predicted core/penumbra and/or perfusion status.

Building and Testing the Prediction Models

[0158] In this invention, mCTA images were analyzed against the boundaries defined by the above databases using machine learning procedures. As noted above, mCTA images are effective in diagnosing and making treatment decisions; however, until now have been unable to be used as tools to quantitatively predict core, penumbra and perfusion status.

[0159] Thus, the models sought to determine if information from mCTA images can be correlated to data from CTP studies, be then used to create core/penumbra/perfusion maps (ostensibly at the time of diagnosis and treatment decision) and then based on follow-up images demonstrated that the prediction maps correlated well to the final outcome as determined by final outcome images.

[0160] The models were built based on knowledge of the flow of contrast dye through affected and unaffected tissues in the cerebral arteries.

[0161] FIG. 1 shows representative curves of the flow of dye at one location (i.e., a zone of interest) through unaffected (i.e., contralateral vessels; shown as the artery input function) and affected (i.e., ipsilateral vessels; shown as the tissue density curve). As can be seen, contrast dye will generally flow through affected and unaffected tissues differently showing variations as different times to peak dye as well as different slopes for the rise and fall of dye.

[0162] It is understood that for a CTP study, up to 50 sequences images would be taken such that the contrast dye curves as shown in FIG. 1 would have a typical resolution as shown by the curves whereas for an mCTA study, images would only be taken at time points t.sub.0,t.sub.1−t.sub.3 thus having substantially lower resolution of contrast dye flow as compared to the CTP curves.

[0163] For example, as shown in FIG. 1A, only three data points, S1, S2 and S3 representing contrast density at different times are obtained.

[0164] Returning to FIG. 1 and with reference to FIG. 2, the difference in flow is explained as follows. As the dye enters the cerebral vessels, the flow bifurcates to the ipsilateral and contralateral vessels. For both CTP and mCTA studies, a series of x-ray images are taken as the bolus of dye enters these vessels. As the contrast flows steadily through the cerebral arteries, the x-ray images will show the relative concentration of dye throughout the arteries at a given time and location on both sides. The relative flow rate of the dye will determine how long the dye will be seen. That is, fast flowing dye will rise in concentration rapidly but also decrease in concentration rapidly (as per the artery input function), whereas slower moving dye will take longer to accumulate and longer to clear (as per the tissue density curve).

[0165] Thus, images obtained at different times will show directly and indirectly, the flow of contrast through the brain arteries at the different times. In unaffected vessels, contrast will appear and will have substantially disappeared between to and t.sub.3. Further, contrast will peak around t.sub.1; be dropping away by t.sub.2 and be less than about 25% of the peak of t.sub.1 by t.sub.3.

[0166] For stroke affected tissues, shown as the tissue density curve, the flow of contrast will be time-delayed where for a given location, if contrast is being held up, the peak flow will be time-shifted to a later time, the peak contrast may be lower as compared to unaffected tissues and the time to clear and rate of clearance may be different.

[0167] As shown in FIG. 2, after the data has been obtained, for a CTP study the data is processed by deconvolution algorithms to create impulse residual functions and enable various parameters to be assessed.

[0168] FIG. 2 shows a number of the physiological parameters that can be derived from the contrast density curves and the impulse residual functions. As introduced above, these include: [0169] a. Mean Transit Time (MTT) which represents the length of time in seconds that it takes for blood to move from arteries to capillaries to veins. MTT=CBV/CBF. An increase in MTT indicates a vasodilatory response to reduced blood flow. [0170] b. Time to Maximum (Tmax) which represents the time at which the maximum value of the residual function occurred and represents a delayed arrival of contrast agent [0171] c. Cerebral Blood Volume (CBV) is the area under the FIG. 2 curve and is the volume flow rate through cerebral vasculature per unit time (ml/100 g of brain tissue) [0172] d. Cerebral Blood Flow (CBF) is the maximum value of the FIG. 2 curve and is the amount of blood flowing through capillaries per unit time per unit tissue (ml/min/100 g of brain tissue). It can be used to identify areas of hypoperfusion. Infarct core show decreased CBF by <30%.

[0173] Other parameters can be derived including: [0174] e. Time to Peak (TPP) which represents the time in seconds to reach peak voxel enhancement. TPP is an indicator of delayed flow in the setting of stenosis or occlusion and is increased when abnormal. [0175] f. Mismatch Volume is the difference in volume between total hypoperfused area and core infarct and equals penumbra. Mismatch ratio is the ratio of total hypoperfused area and core infarct.

[0176] Tmax and CBF are the main parameters used to determine core and penumbra.

[0177] From FIGS. 1 and 1A, for a mCTA study, contrast density values are obtained from images taken at t.sub.1, t.sub.2 and t.sub.3. Hence for a given location, as shown in FIG. 1, data points a, b and c are obtained for the ipsilateral side and data points d, e and f are obtained from the contralateral side. Importantly, as minimal data is obtained (as compared to CTP), the peak contrast density between t1 and t2 has not been measured and thus the peak level is not known from a direct measurement.

Machine Learning Models—Development and Testing Overview

[0178] Testing and evaluation protocols were developed using three machine learning models including, a core, penumbra and perfusion model, explained in detail below.

[0179] The core model seeks to predict the volume of core, the penumbra model seeks to predict the volume of penumbra and the perfusion model seeks to predict tissue perfusion status.

[0180] FIG. 3 is an overview of the process for model development, testing, use and refinement. Images from past patients having undergone CTP, reperfusion and follow-up studies 30a as well as images from patients having undergone CTP, no reperfusion and follow-up studies 30b are each subjected to feature extraction analysis 30c and density/acquisition time analysis 30d. Data from these analyses is used to build and train the models 30e. Data from a current patient having undergone an mCTA study 30f can then be introduced into the model 30e to output predictive penumbra/core/perfusion maps 30g for clinical use. Follow up studies on the current patient (eg. NCCT) 30h can be subsequently introduced to the model to enable model refinement.

[0181] FIG. 3A shows details/example of the process for more specific development/derivation of each of the (a) penumbra, core/infarction prediction maps and (b) perfusion prediction maps. As shown in FIG. 3A(a), images from 96 patients having undergone CTP, reperfusion, and follow-up as well as images from 44 patients having undergone CTP, no reperfusion and follow-up were analyzed for feature extraction. Test mCTA images (including follow up images) were subjected to the same process. After training with all images, using the penumbra model, the model was then tested on the mCTA images (without the follow up images) to predict penumbra for the mCTA images. A similar strategy was then used for infarction/core prediction.

[0182] Feature extraction involves analyzing density and acquisition time of areas of interest, namely those areas that may be showing abnormal flow of contrast (ipsilateral side) and the corresponding features on the contralateral side where flow is normal. That is, the steps of feature extraction will examine a baseline density level and look for changes in density across each image. Those areas where density is showing change above a threshold level is marked for further analysis whereas those areas where density does not change above the threshold level will not be marked and not subject to further analysis.

[0183] More specifically, zones of interest may be determined by evaluating the following: [0184] 1) average and standard deviation of Hounsfield units (HUs) across 3-phase CTA images; [0185] 2) coefficient of variance of HUs in 3-phase CTA images; [0186] 3) changing slopes of HUs between any two phases; [0187] 4) peak of HUs in 3-phase CTA images; [0188] 5) time of peak HU.

[0189] The size of the zone of interest may be variable and/or adjusted depending on the desired resolution. For example, features were calculated in zones centered at each voxel at three scales (3×3×3, 7×7×7, and 11×11×11 voxels) and then normalized using z-score method.

[0190] Analysis of changing slopes between images at different times provides useful information about how quickly contrast agent may be flowing into or out of affected tissues at the scale of individual or a defined number of voxels obtained from the mCTA imaging information. As shown in FIG. 1A, for three phases, three signal intensity (SI) values can be determined (eg. SI.sub.1-SI.sub.3) relative to a baseline HU (density) value. From these data points, the difference d in SI values between points, enables calculation of the slope of signal intensity change between phases. Slopes may be positive + or negative −. For example, d.sub.1, d.sub.2 and dd are defined as follows:

[00001]$\begin{array}{c}{d}_1={\mathrm{Sl}}_2-{\mathrm{Sl}}_1\\ d_2={\mathrm{Sl}}_3-{\mathrm{Sl}}_2\\ \mathrm{dd}=d_2-d_1\\ ={\mathrm{Sl}}_3-{\mathrm{Sl}}_2-\left({\mathrm{Sl}}_2-{\mathrm{Sl}}_1\right)\\ ={\mathrm{Sl}}_1-2{\mathrm{Sl}}_2+{\mathrm{Sl}}_3\end{array}$

[0191] dd is the 2.sup.nd derivative of the change of slope between the two lines. The slopes of each of d.sub.1, d.sub.2 and dd are calculated together with their sign (i.e., + or −) and used as a basis for understanding the contrast delay for a particular location which can then be used to assign a tissue health value to that location.

[0192] FIG. 4 is a representation of how the slopes may vary based on the relative delay of contrast arriving at a particular location. For example, if there is substantially no delay of contrast arriving at the ipsilateral location of interest, one would expect flow characteristics of normal tissue. Three-point line pattern 40, shows points X, Y and Z defining two lines having slopes d.sub.1, d.sub.2. As shown in FIGS. 1, X, Y and Z are similar to points d, e and f on the artery input function. Similarly, three-point line pattern 42, shows points A, B and C defining two lines having slopes d.sub.1′, d.sub.2′ which are similar to points a, b and c on the tissue density curve of FIG. 1 and shows the delay. As contrast is increasingly held up, the tissue density curve as shown in FIG. 1 will be time-shifted to the right; thus, producing three-point line patterns similar to those shown in FIG. 4.

[0193] Accordingly, depending on the time delay, each three-point line pattern will have a range of profiles as shown in FIG. 4 where the time delay increases from the right to left direction.

[0194] The ±patterns of the calculated d.sub.1, d.sub.2, dd values represent different scenarios of contrast flow as shown in Table 2 and FIG. 4A which shows the different patterns together with representative circulation value ranging from 1 (good flow) to 5 (poor flow) which can be indicative of relative tissue health. Information about how contrast is flowing to a location and out of a location can be determined from each three-point line function.

TABLE-US-00002 TABLE 2 d-Slope Interpretation and Representative Tissue Health Coding on Ipsilateral Side Relative Group Circulation Patterns d.sub.1 d.sub.2 dd Interpretation Value 44a − − + Voxel showing shortest delay 1 (if any) of contrast reaching the location. 44b − − − Voxel showing next shortest 2 delay in contrast reaching the location. Contrast has reached the voxel quickly and is showing a slight delay in contrast washout 44c, d + − − Voxel showing peak contrast 3 density at around the time the phase 2 image was taken. 44e + + − Voxel showing that contrast 4 density is starting to peak at around the time the phase 2 image was taken. 44f + + + Voxel showing greatest delay. 5 Across all phases, contrast density is continuing to rise. ** combinations of [− + −] and [+ − +] are not mathematically plausible.

[0195] Each group pattern can be provided with threshold values to determine which group a particular line pattern may be categorized within. Different groups may be defined with individual, or a range of colors used for subsequent color mapping of a particular voxel. The above analysis is performed for each voxel of an image volume of interest.

[0196] To obtain meaningful prediction data, group patterns are matched to past data showing similar patterns.

[0197] As shown in FIG. 4B, the group pattern determined above can be matched to past patient images showing similar group patterns. For example, FIG. 4B shows 3 tissue input curves, TIC, TIC1 and TIC2 which represent the data from 3 CTP studies and with “matching” phase data points p1-p3 and baseline p0. As each of the CTP studies have determined penumbra, core and perfusion maps, data from past studies provides a range of values for penumbra, core and perfusion which can thus be used to provide a basis for predicting these values for the current patient and used to generate 2- and/or 3-dimensional color-coded image/maps. That is, data from the past studies can be processed by appropriate deconvolution algorithms to generate CT perfusion maps such as Tmax, CBV and CBF for each TIC from the past studies.

Model Development-Machine Learning

[0198] Referring back to FIGS. 3 and 3A, as noted above, mCTA images as well as NCCT at baseline are introduced into the model and matched to the best fit within the CTP datasets to provide both an mCTA penumbra prediction and mCTA core prediction.

[0199] Hence, a prediction of Tmax, CBV and CBF for each can be estimated for the 3-phase mCTA study based on the CT perfusion maps from the past studies.

[0200] In one embodiment, as shown in FIG. 4C, machine learning techniques such as deep neural network can be used to interpolate a TIC having p0-pm points at discrete time points to create a predicted TIC represented by continuous data points CO-Cn, which can then be utilized to build predicted CT perfusion maps.

[0201] In further detail, as shown in FIG. 5, in one embodiment, the following steps were undertaken in processing the mCTA images to build the models.

[0202] Step 1-Analysis of data from mCTA images [0203] a. Pre-processing mCTA training images. This includes spatially aligning 3-phase mCTA and NCCT at baseline in order to correct patient movement during acquisition; and standardizing density values of mCTA by subtracting density value of NCCT from each phase CTA at each voxel location. [0204] b. Analyzing the mCTA images and identifying features of interest for each voxel location wherein the features of interest are derived from density values across three phases and acquisition time of each phase of the mCTA. These features of interest are extracted and includes data from infarcted and normal tissue on the ipsilateral side as defined by the follow-up imaging. This information was then normalized by comparing it with the ipsilateral hemisphere via a z-scoring method. The z-scoring method was performed by subtracting the mean value of an image and divided by the standard deviation of the image.

[0205] Step 2-Training [0206] a. Training a random forest classifier using the features of interest from step 1b while using follow up infarct segmentation as an indicator. For example, a “1” is assigned to represent an infarct voxel whereas a “0” is assigned to represent normal tissue. This information is then used to generate a penumbra probability map, indicating how likely a voxel in mCTA will be infarcted if no reperfusion is achieved. Random forest classifier is an ensemble learning method for classification that operates by constructing a multitude of decision trees at training time and outputting the class that is the mode of the classes (classification) of the individual trees.

[0207] Step 3-Applying a Threshold [0208] a. Applying a threshold on the probability map from step 2 to generate a binary mask of penumbra tissue. For example, a “1 in this binary mask represents an infarct voxel whereas a “0” represents normal tissue.

Other Modelling Techniques

[0209] The machine learning model may be constructed using other modelling techniques including support vector machine, neural network and/or k nearest neighbor techniques.

Model Validation

[0210] After predicted core, infarction and perfusion maps are created, correlation to the actual” outcome can be made to determine the accuracy of the models. As discussed below in the validation study, the models were statistically validated.

Additional System and Model Functionality

[0211] The models are effective in assisting a physician in making treatment decisions during diagnosis. As discussed above, for patients with acute ischemic stroke, time to treatment is well correlated to patient outcome; hence obtaining effective information to enable a physician to make a treatment decision as soon as possible is desired. As such, the steps to determine and present core, penumbra and perfusion status from the time current mCTA images are introduced into the system are ideally completed in 10 minutes or less.

[0212] Additional data can also be introduced into the past patient database and presented as additional information to the physician. In one embodiment, the past patient database includes information about patient outcome following treatment or not as may be input after an NCCT follow-up study has been completed. Thus, upon creating a prediction map as described above for the current patient, data from one or more of the closest match past patient studies describing patient outcome may be presented to the physician. Such outcome data may be a quantified and standard assessment score as known. For example, a past patient study may include treatment information that a successful thrombectomy at a particular region of interest was completed within 40 minutes of images being obtained and the patient made a good recovery. A second past patient study showing similar map information may include information that treatment occurred in 90 minutes and that patient recovery was poor. As such, the physician can use this information as additional information to evaluate if they should initiate a specific treatment.

[0213] In other embodiments, likelihood of success of a treatment may be presented and be correlated to any one of or a combination of predicted core, penumbra and/or perfusion status.

[0214] In other embodiments, features of interest and patterns relating to occlusion location, core, penumbra and perfusion, include any one of or a combination of the first-order statistics, such as mean and histogram of HU values, and texture features, such as gray-level co-occurrence matrix and gray level run length matrix.

[0215] The features of interest relating to occlusion location, core, penumbra and perfusion are calculated at different scales; for a given voxel corresponding to the axial imaging, the features are calculated at low, median, and high-resolution scales.

[0216] In various embodiments, the features of interest mostly contributed to occlusion location, core, penumbra and perfusion are automatically or manually selected using feature selection technique in order to improve prediction accuracy, reduce overfitting, and reduce training time. The feature selection technique includes univariate selection, feature importance, and correlation matrix with heatmap.

[0217] In various embodiments, each probability map is thresholded to generate infarct core and/or penumbra and/or perfusion volume for the axial imaging slice.

[0218] In other embodiments, morphological operations including image dilation and/or erosion and component analysis are applied after thresholding to remove isolated islands.

[0219] In another embodiment, the machine learning model enables prediction of a combination of core and penumbra from a multiple label machine learning model, that is, label 1 denotes core, label 2 denotes penumbra, and label 3 denotes normal tissue. The single model can predict core and penumbra at the same time.

[0220] In other embodiments, as shown in FIG. 6 features at a patient level are selectively used as inputs, such as artery and vein input functions manually or automatically selected, together with meta data including age, sex, NIHSS, ASPECTS, occlusion site.

Case Examples

[0221] FIG. 6 shows a hypothetical example, in which, artery and vein input functions and meta data are used as additional features to refine the developed machine learning model and better predict tissue status, especially for the patients with small or no occlusions.

[0222] FIGS. 6A and 6B are examples of application of the system to create effective diagnostic maps. FIG. 6A shows representative mCTA images for an 83 year old female showing a left ICA occlusion with a stroke onset to CT time of 95 mins, NIHSS 23, and ASPECTS 10.

[0223] FIG. 6B shows representative mCTA images for an 87 years old male showing left distal M1 occlusion with a stroke onset to CT time of 208 mins, NIHSS 28, and ASPECTS 5.

Medium Vessel Occlusion (MeVO) Tool and Use

[0224] In various embodiments as illustrated in FIGS. 7-7H, the system may additionally be used to assist in diagnosing MeVO and provide different outputs to a diagnosing physician with varying levels of detail that may be relevant to treatment decisions. As described below, levels of output may include: [0225] a) a simple indication that MeVO (as opposed to LVO or SVO) is likely present, [0226] b) identification of zones of interest where MeVO may be present; and/or, [0227] c) specific identification of locations/vessels where MeVO may be present.

[0228] Improved MeVO detection is achieved through utilization of CTA images and/or with prediction maps (e.g., core/penumbra/perfusion) together with additional functionality within the system including anatomical maps built from a plurality of patients and/or knowledge obtained by prediction models/learning algorithms as described above.

[0229] For example, in one embodiment, as with the general prediction map system described above, the MeVO system/tool is trained with past images and used to create effective prediction maps for a current patient that can be used to locate and quantify hypoperfused tissue and subsequently evaluate if the parameters of the hypoperfused tissue are indicative of MeVO.

[0230] In accordance with one embodiment, the steps of identifying MeVO may be achieved automatically or semi-automatically by the following general process: [0231] a. Assess location of affected tissue from various combinations of CTA images, CTP studies, mCTA studies as may be available. [0232] b. From the hypoperfused area, define a general zone of interest where MeVO may be present. [0233] c. Determine additional parameters of the affected tissue including the shape, size/volume, confluence, involvement of the cortex and sub-cortical white matter and knowledge of the known supply by vessels in that region. [0234] d. Determine the 2D/3D position, size, and shape of the hypoperfused area. The position, size and shape is determined by assembling voxels showing affected tissue characteristics that share a boundary with adjacent voxels showing affected tissue in both proximal and distal positions. The assembled voxels define a quantifiable volume (i.e., based on a calculated number of linked voxels) and shape characteristics (e.g., a characteristic “cone”, “frusto-conical” or “wedge” shape in the distal direction). The location of the volume is compared to known regions of the brain based on general knowledge of brain anatomy or specific knowledge of brain anatomy from a match/correlation analysis (as described below). Table 3 shows various characteristics that can be assessed. [0235] e. Determine if the hypoperfused volume is LVO, MeVO or SVO based on the analysis in step d.

TABLE-US-00003 TABLE 3 Characteristics of LVO, MeVO and SVO Characteristic LVO MeVO SVO Volume (ml) Range: 80-400 Range: 25-80 <25 Common: 150-180 Common: 30-70 Shape Cone/Wedge Cone/Wedge Often indeterminate Confluence Yes Yes No Cortex Involved Mostly Maybe Sub-cortical Involved Involved Maybe white matter Possible sites 1-2 6-20 >20 (often beyond of Occlusion resolution of CTA) per hemisphere [0236] f. If it is determined that MeVO is likely, additional analysis can be conducted depending on the desired outputs as described above. If greater precision is desired a zone of interest analysis can be conducted to highlight a zone(s) on the images where further investigation could be conducted. A vessel of interest analysis may also be conducted to identify a vessel(s) of interest. Generally, to conduct these analyses: [0237] i. Contrast densities from adjacent proximal voxels from one or more phases of CTA images are searched for variations in contrast density that may signal that contrast is flowing normally or abnormally within one or more nearby voxels. [0238] ii. Voxels that indicate normal contrast flow may be discarded from further analysis and/or be utilized as a baseline for determining if contrast is abnormal or normal in nearby voxels. [0239] iii. Voxels indicating abnormal contrast flow may be flagged for further analysis. [0240] iv. For those voxels showing abnormal flow, additional analysis is conducted to highlight zones/vessels where contrast abruptly transitions from no contrast to significant contrast between adjacent images or vice versa. [0241] v. Highlighted zones and/or vessels may be automatically marked on CTA images as a suggestion to the physician to focus attention in a particular area.

[0242] The foregoing is illustrated by the following illustrative example. As noted above, FIG. 7 is a representative figure showing the structure and distribution of level 1-3 vessels. From known anatomy, a general pattern of interconnected volumes can be built through successive levels from proximal to distal vessels as shown schematically in FIG. 7H. That is, L1 vessels as represented by individual voxels may be positioned in 2D/3D space. L2 vessels are represented by voxels surrounding or branching off the L1 vessels to a defined distance and L3 vessels are represented by voxels surrounding the L2 vessels to a defined distance.

[0243] Thus, from a prediction map, a hypoperfused zone may be identified and correlated to a 3D location (for example, a particular M2/M3 zone) and thus to a general location in the brain. With knowledge that vessels proximal to that location are generally perfused by adjacent areas in a known direction, corresponding proximal voxels on the images may be flagged for additional investigation.

[0244] Importantly, voxels that may be distal and/or beyond a particular threshold distance from the hypoperfused area may be discarded from further processing. Similarly, proximal voxels beyond a threshold distance may also be discarded.

[0245] Further processing can look for a variety of changes within those flagged voxels, including normal and abnormal contrast flow and/or collateral filling from a later CTA image.

[0246] In one embodiment, different phases of voxels (e.g., from mCTA) are overlaid with respect to one another to help identify a “missing vessel” i.e., one where no contrast is directly observed but contrast behaviour nearby suggests its presence.

[0247] FIG. 7G describes a process of providing additional information to a physician in accordance with one embodiment to assist in diagnosing MeVO. In this example, output to the physician is derived from a combination of current patient images 90a and historical images 90b which are utilized by a model 90c as generally described above. From the model, prediction maps (eg. core/penumbra/perfusion) 90d are analysed to identify hypoperfused regions 90e in a current patient typically marked by color coding. Analysis to determine if the hypoperfused volume is likely LVO, MeVO or SVO is conducted. If MeVO is suggested 90e, these maps are correlated to specific 2D areas/volumes in the brain of the current patient and can then be compared and correlated to corresponding hypoperfused areas from historical data images 90f. The best fit historical data images, having associated L1-L3 mapping data, is/are used to identify corresponding proximal zones in the current patient images 90g, which are marked as zones of interest for the physician to examine for further MeVO diagnosis.

[0248] In various embodiments, historical images may be filtered to limit the dataset to MeVO images only 90i.

[0249] In addition, and prior to comparison with current patient images, 2D/3D relationships between level 1-3 vessels can be derived 90j and as shown in FIG. 7H, wherein numerous volumes of tissue are assembled into a successive branch structure in 2D/3D space. FIG. 7H shows a graphical representation of voxels labelled according to their level and that based on their level may touch voxels of the same label or one level higher/lower. That is, a level 1 voxel can only directly touch/communicate with a level 2 voxel etc. Shown as a representative 2D and partial 3D map, FIG. 7H shows two zones where a number of L3 voxels are grey representing the color code for hypoperfused tissue as shown at one level (shown as smaller dotted ellipse). From this, the system can initially search for voxels distal to most proximal hypoperfused level as a starting point to calculate the location, volume and shape of the hyperperfused volume. In FIG. 7H, this is shown as the volume defined between the two dotted ellipses (ie. a general frustoconcial shape). The detection of the location of MeVO is determined by analysis of proximal voxels that are contiguous with those L3 voxels as a likely location of an occlusion.

[0250] For illustrative purposes only, non-confluent voxels are shown which are unlikely to be present in a typical MeVO case.

MeVO Application

[0251] In various clinical settings, the MeVO tool can be used to assist in treatment/triaging decisions. As shown in FIG. 7G, a first level output 901 may simply provide output that MeVO likely exists. If this information is provided based on imaging conducted at a smaller center some distance from a main center, the diagnosing physician may be able to make an effective triaging decision to transfer the patient to the larger center, where the expertise in providing reperfusion treatment can be provided in a reasonable time frame. In this embodiment, the determination of MeVO may be based on various factors including the size, shape, location, confluence of the hypoperfused area and/or involvement of cortical grey matter and may include quantitative outputs regarding the likelihood of the diagnosis (e.g., 90% probability the occlusion is MeVO).

[0252] At other centers, particularly where treatment options may be available, additional outputs may be provided. These may include marking zones of interest as per analysis conducted at step 90h and/or conducting further analysis 90k that allows more specific identification of vessels of interest 90n.

Case Example

[0253] An 88-year-old female, arriving from home presented with expressive aphasia and mild right sided weakness since 2h; NIHSS on presentation: 10.

[0254] A perfusion map from CTP or a predictive perfusion map from mCTA as described above was obtained indicating an area of brain was hypoperfused.

[0255] The size and location characteristics of the hypoperfused area indicated a likely occluded vessel in an adjacent and proximal vessel. Based on the volume of tissue that is hypoperfused, an estimate of the size of vessel is made. The MeVO tool predicts and marks one or more areas where the occlusion is likely to be allowing the physician to quickly focus attention on those areas.

[0256] In various embodiments, past patient images are subjected to machine learning analysis to refine the precision of locating potential occlusion sites based on evaluations and variations across multiple past images.

[0257] As above, when the current patient images are introduced into the model, they are analyzed to find past patient images most correlated to the current images. As a result, the accuracy of predicting the location of the MeVO may be improved.

[0258] As shown in FIG. 7A for the above patient, an initial non-contrast CT scan was unremarkable. Following mCTA, a prediction map as shown in FIG. 7B was produced showing a specific hypoperfused region. Based on the relative position of the hypoperfused region, the MeVO tool determines and identifies a particular zone of interest in specific scan images as shown by hashed circles in FIG. 7C.

[0259] Additional analysis (manual or automatic) is conducted within that zone to identify an occluded vessel marked by the arrow in FIG. 7D. Depending on the protocols followed and for the purposes of illustration, occlusion was confirmed by conventional angiogram shown in FIG. 7E. Following reperfusion, follow-up imaging shows the opened vessel as shown in FIG. 7F.

Validation Study

[0260] As described above, Multiphase CT-Angiography (mCTA) provides time variant images of the pial vasculature supplying brain in patients with acute ischemic stroke (AIS). Described below is a machine learning (ML) technique that predicts infarct, penumbra and tissue perfusion from mCTA source images.

Study Methodology

[0261] 284 patients with acute ischemic stroke (AIS) were included. All patients had non-contrast CT, mCTA and CTP imaging at baseline and follow up MRI/NCCT imaging. Of the 284 patient images, 140 patient images were randomly selected to train and validate three ML models to predict infarct, penumbra, and perfusion parameter on CTP, respectively. The remaining unseen 144 patient images independent of the derivation cohort were used to test the derived ML models. The predicted infarct, penumbra, and perfusion volume from ML models was spatially and volumetrically compared to manually contoured follow up infarct and time-dependent Tmax thresholded volume (CTP volume), using Bland-Altman plots, concordance correlation coefficient (CCC), intra-class correlation coefficient (ICC), and Dice similarity coefficient (DSC).

Study Results

[0262] Within the test cohort. Bland-Altman plots showed that the mean difference between the mCTA predicted infarct and follow up infarct was 21.7 mL (limit of agreement (LoA): −41.0 to 84.3 mL) in the 100 patients who had acute reperfusion (mTICI 2b/2c/3), and 3.4 mL (LoA: −66 to 72.9 mL) in the 44 patients who did not achieve reperfusion (mTICI 0/1). Amongst reperfused subjects. CCC was 0.4 [95% CI: 0.15-0.55, P<0.01] and ICC 0.42% Cl: 0.18-0.50, P<0.01]; in non-reperfused subjects CCC was 0.52 [95% CI: 0.2-0.6, P<0.001] and ICC 0.6 [95% Cl: 0.37-0.76, P<0.001]. No difference was observed between the mCTA and CTP predicted infarct volume for the overall test cohort (P=0.67).

[0263] Multiphase CT Angiography is able to predict infarct, penumbra and tissue perfusion, comparable to CT perfusion imaging.

Study Background

[0264] Ischemic infarct core estimated using CT perfusion (CTP) at admission may be used in treatment decision making for patients with acute ischemic stroke (AIS). .sup.1-4 Classification of infarct core and penumbra is achieved using tissue perfusion estimates derived using a deconvolution algorithm from repeated serial imaging. The mismatch ratio between salvageable tissue (penumbra) volume and infarct core volume can be used for selecting patients presenting beyond 6 hours and up to 24 hours from last known well. .sup.3 CTP is limited by varying standardization of CTP parameter thresholds across different vendors, longer acquisition times and consequent susceptibility to patient motion, increased radiation dose, limited coverage (with some scanners) and the need for additional technical expertise to acquire the images. .sup.5-7

[0265] Multiphase computed tomographic angiography (mCTA) has been similarly used to select patients with AIS for endovascular therapy (EVT) in recent clinical trials. Advantages of this technique compared to CTP are simpler image acquisition, lower radiation exposure, no additional contrast compared to single-phase CTA, and whole-brain time-resolved images of pial arteries and veins beyond an occlusion while also determining thrombus location, size, vessel patency and tortuosity. .sup.10, 11 Multiphase CTA imaging has not been as commonly used to predict ischemic tissue fate on a voxel by voxel basis, in the same way as CTP imaging. However, recent studies have demonstrated that mCTA can be used to predict tissue fate regionally, similar to CTP. .sup.12-14 An ability to harness the advantages of mCTA while producing brain maps that estimate tissue perfusion and predict tissue fate is likely to be of significant clinical utility.

[0266] The study aimed to develop a machine learning based technique to estimate infarct core, penumbra and tissue perfusion in patients with acute ischemic stroke.

Study Materials and Methods

[0267] Data from the Prove-IT study (Precise and Rapid assessment of collaterals using multi-phase CTA in the triage of patients with acute ischemic stroke for IA Therapy), a multicenter study that acquired acute multimodal CT imaging including NCCT, multiphase CTA imaging (three phases), and CTP at baseline among ischemic stroke patients. .sup.10, 12 This study was approved by the local institutional review board.

Study Participants

[0268] Subjects who had (1) baseline non-contrast-enhanced CT (NCCT) and mCTA; (2) baseline CTP imaging with >=8 cm z-axis coverage; (3) had reperfusion assessed on conventional angiography after thrombolysis treatment (intravenous tPA, endovascular therapy, or both) with the modified thrombolysis in cerebral infarction [mTICI]); and (4) had 24/36-hour follow-up imaging on diffusion MRI or NCCT were included in this analysis. Patient inclusion and exclusion are shown in FIG. 5A. 284 patients, of whom, 196 patients had acute reperfusion (mTICI 2b/2c/3) and 88 patients did not (mTICI 0/1) were included.

Image Preprocessing

[0269] Each CTP study was processed using commercially available delay-insensitive deconvolution software (CT Perfusion 4D, GE Healthcare, Waukesha, WI). Absolute maps of cerebral blood flow (CBF, mL.Math.min.sup.−1.Math.(100 g).sup.−1], cerebral blood volume (CBV, mL.Math.(100 g).sup.−1], and Tmax (seconds) were generated. Average maps were created by averaging the dynamic CTP source images. Time-dependent Tmax thresholds confirmed previously, were used to generate baseline CTP thresholded maps (perfusion volume). .sup.6, 7

[0270] NCCT and mCTA images were first skull stripped. .sup.15 Three-phase CTA images were then aligned using rigid-body registration to account for patient movement. The aligned 3-phase CTA images were registered onto NCCT images using affine registration. Two radiologists (>5 years' experience) used ITK-SNAP and consensus to manually delineate the infarct region on follow-up DWI/NCCT imaging. .sup.16 The follow-up images along with manual infarct segmentations and CTP average maps were registered onto NCCT images, thus bringing all images into the same image space. When registration was sub-optimal, manual refinement of the registered infarct segmentations was attempted. The NiftyReg tool was used for all image registration tasks. .sup.17

Machine Learning Model

[0271] For the analysis, infarct core was defined as tissue that is infarcted on follow-up imaging even with reperfusion. Penumbra was defined as ischemic tissue that was not infarct core but infarcts on follow-up imaging when reperfusion is not achieved. These definitions of infarct core and penumbra are operational in context and not biological. The perfusion map used was a Tmax map thresholded using previously published time dependent thresholds. .sup.6-7

[0272] Three machine learning models were developed: (1) Infarct model; (2) Penumbra model; and, (3) Perfusion model.

[0273] A 2-stage training mechanism was developed to train two machine learning models to predict infarctcore and penumbra respectively. The detailed training and testing strategy is shown in FIG. 3A.

[0274] Of 88 patients without acute reperfusion (mTICI 0/1), 44 patients (35 for training and 9 for validation) were randomly selected to derive a random forest classifier at the first stage for prediction of follow-up infarction in the non-reperfused patients (Penumbra model), while the remaining 44 patients with mTICI 0/1 independent of the derivation cohort were used to test this derived Penumbra Model. Of those 196 patients with mTICI 2b/2c/3, 96 patient images (70 for training and 26 for validation) randomly selected were first processed by the 1.sup.st stage Penumbra model, generating penumbra probability maps. These probability maps along with mCTA images were then used as inputs to derive the second random forest classifier at the second stage for infarct prediction (Infarct model) using follow up infarct manually segmented as a reference standard, while the remaining 100 patients with mTICI 2b/2c/3 reperfusion independent of the derivation cohort were used to test the derived Infarct Model. The final predictions are shown as infarct core and penumbra where penumbra is defined as affected tissue from the penumbra model minus affected tissue from the infarct core model (FIG. 5).

[0275] In order to show the ability of mCTA to estimate tissue perfusion at baseline compared to CTP imaging, the 140 patient images used for training and validating the Penumbra and Infarct models were reused to train and validate the third random forest classifier (Perfusion model). For deriving and testing this model, time dependent Tmax thresholded maps were used as reference standard. .sup.6-7 The 144 images used for testing Penumbra and Infarct models independent on the derivation cohort were used to test the Perfusion model.

[0276] All three random forest models shared the same self-designed features as inputs. NCCT HU values were first subtracted from 3-phase CTA images, leading to a 3-point time intensity curve (TIC) for each voxel. Several features were extracted from the time intensity curve (TIC) for each voxel and used for deriving and testing the three random forest classifiers.

[0277] These include: 1) average and standard deviation of Hounsfield units (HUs) across 3-phase CTA images; 2) coefficient of variance of HUs in 3-phase CTA images; 3) changing slopes of HUs between any two phases; 4) peak of HUs in 3-phase CTA images; 5) time of peak HU.

[0278] All these features were calculated in the neighborhood centered at each voxel at three scales (3×3×3, 7×7×7, and 11×11×11 voxels) and then normalized using z-score method. The hyper-parameters for each random forest model, such as the number of trees in the forest and the maximum depth of trees, etc., were optimized using 5-fold cross validation using the respective validation cohort. Specifically, in 5-fold cross-validation, all the original samples are randomly partitioned into 5 equal sized subgroups. Of the 5 subgroups, a single subgroup is retained as the validation data for testing the model, and the remaining t subgroups are used as training data. The cross-validation process is then repeated 5 times, with each of the 5 subgroups used exactly once as the validation data. The 5 results can then be averaged to produce a single estimation. Class weight was set to account for the imbalanced sample distribution based on the ratio of positive and negative samples. The random forest classifiers derived from the training and validation dataset was then applied to the test cohort to generate a probability map for each patient. The probability map was then thresholded by a fixed value of 0.35, followed by image post-processing, such as isolated island removal and morphological operation, to generate the mCTA predicted volume. The thresholding value was optimized and determined from the validation cohort.

[0279] The fixed thresholding value of 0.35 was achieved by maximizing the Dice coefficients between the thresholded binary mask and reference standard of follow up infarct segmentation while varying different discrete thresholding values using the validation cohort. Isolated island removal was used to discard small clustered random noise in the thresholded binary mask. Morphological operation includes image erosion and dilation followed by hole-filling in the binary mask.

Statistical Methods

[0280] Expert contoured follow up lesion volume (Follow up infarct volume) were used as standard reference to evaluate mCTA predicted infarct core and penumbra volume for the test cohort. Time-dependent Tmax thresholded volumes (CTP volume) were used as standard reference to evaluate the mCTA perfusion volume for the test cohort. Bland-Altman plots were used to illustrate mean differences and limit of agreement (LoA) between mCTA predicted and follow up infarct volume, and CTP volume. Literal and relative volume agreement between mCTA predicted and follow up infarct volume, and CTP volume were also assessed using concordance correlation coefficient (CCC) and intra-class correlation coefficient (ICC), respectively. Spatial agreement between mCTA predicted volume and follow up infarct volume, and CTP volume was assessed using Dice similarity coefficient (DSC). Rank sum test was used to assess the difference between any non-normally distributed data. All statistical analyses were performed using MedCalc 17.8 (MedCalc Software, Mariakerke, Belgium) and Matiab (The MathWorks, Inc., United States). A two-sided alpha <0.05 was considered as statistically significant.

Study Results

Study Participants

[0281] Patient characteristics are summarized in Table 4. No differences were observed between the derivation and test cohorts (all P>0.05).

TABLE-US-00004 TABLE 4 Patient characteristics. Derivation Test cohort cohort P Characteristics (N = 140) (N = 144) value Median age, year (IQR) 73 (62-79) 72 (62-80) .73 Sex, n(%) male 80 (57) 77 (53) .56 Median baseline 17 (7-23) 14 (6-18) .12 NIHSS (IQR) Median baseline 9 (8-10) 9 (8-10) .15 ASPECTS (IQR) Median onset-to-imaging 131 (94-226) 139 (88-294) .35 time, min (IQR) Median imaging-to-reperfusion 90 (68-115) 87 (64-125) .97 time, min (IQR) Median onset-to-reperfusion 245 (172-330) 240 (181-377) .71 time, min (IQR) Median follow-up infarct 22.2 (10.3-59.4) 25.9 (10.1-60.6) .60 volume, mL (IQR) Site of occlusion, n(%) ICA 22 (16) 26 (18) .76 MCA: M1 73 (52) 70 (48) .64 other 45 (32) 48 (33) .63 IQR, interquartile range; NIHSS, National Institutes of Health Stroke Scale. * p < 0.05.
Accuracy of mCTA In Predicting Follow up Infarct

[0282] FIG. 8 shows three examples of the mCTA prediction maps compared to CTP Tmax maps and follow up infarct. In particular, FIG. 8 shows multiphase CTA predicted infarct probability map compared to CTP (Tmax) map and follow-up infact. There are 3 examples, each labeled as per rows. Row I: patient with mTICI 2b, Row II: patient with mTICI 1, and Row III: patient with mTICI 3. Columns: mCTA phase 1 to 3, mCTA prediction perfusion maps, mCTA predicted infarct (red in column 5) and penumbra (blue in column 5) overlaid on the mCTA predicted perfusion map, CTP Tmax maps, time-dependent Tmax thresholds predicted infarct, infarct contoured in follow-up imaging. respectively. The penumbra is shown as affected tissue from the Penumbra model minus affected tissue from the Infarct model. Contour Color Code: inner or single contour “Red” R (core); outer contour—“Blue” B (penumbra). All areas within a red contour are red. All areas between a red contour and blue contour are blue.

[0283] FIG. 8A illustrates Bland-Altman plots (a) mCTA penumbra volume predicted using Penumbra model vs. follow-up infarct volume for 44 patients who did not achieve acute reperfusion (mTICI 0/1); (b) mCTA infarct volume predicted using Infarction model vs. follow-up infarct volume for 100 patients who had acute reperfusion (mTICI 2b/2c/3); and (c) mCTA perfusion volume predicted using Perfusion model vs. time-dependent Tmax predicted infarct volume for all 144 patients in the test cohort. FIG. 8A(a) illustrates a Bland-Altman agreement plot between mCTA predicted infarct core+ penumbra volume and follow up infarct volume for the 44 patients who did not receive acute reperfusion (mTICI 0/1) in the test cohort. The mean difference between this mCTA predicted infarct core+penumbra volume (median, 33.2; IQR, 20.6-53.2 mL) and follow up infarct volume (median, 26.8; IQR, 12.3-54.8 mL) was 3.4 mL (LoA: −66-72.9 mL, P=0.69). The CCC between the two volumes was 0.52 [95% CI: 0.2-0.6. P<0.001] while the ICC was 0.6 [95% Cl: 0.37-0.76, P<0.001]. The median DSC between the mCTA predicted lesion and follow up infarct was 26.5% (IQR, 12.9-39.3%).

[0284] FIG. 8A(b) illustrates a Bland-Altman agreement between the mCTA predicted infarct volume and follow up volume for 100 patients who achieved acute reperfusion (eTICI 2b/2c/3) in the test cohort. The mean difference between the mCTA infarct volume (median, 37; IQR, 23-58 mL) and follow up volume (median, 26; IQR, 13-54 mL) was 21.7 mL (LoA: −41.0-84.3 mL, P=0.48), CCC was 0.4 [95% CI: 0.15-0.55, P<0.01] and ICC 0.42 [95% Cl: 0.18-0.50, P<0.01]. The median DSC between the mCTA infarct and follow up infarct was 24.7% (IQR, 13.8-30.4%).

[0285] The association between infarct volume predicted by the mCTA Infarct and Penumbra models and follow up infarct volume in the whole test cohort is shown in Table 5.

Accuracy of mCTA Predicting Perfusion Status

[0286] FIG. 8A(c) illustrates a Bland-Altman agreement between the mCTA predicted perfusion volume and CTP volume for 144 patients in the entire test cohort. The mean difference between the mCTA perfusion (median, 40.5; IQR, 22.9-59 mL) and CTP volume (median, 26.9; IQR, 6.7-56.7 mL) was 4.6 mL (LoA: −53-62.1 mL, P=0.56), CCC was 0.63 [95% CI: 0.53-0.71, P<0.01] and ICC was 0.68 [95% Cl: 0.58-0.78, P<0.001]. The median DSC between mCTA predicted perfusion and CTP volume was 40.5% (IQR, 25.7-52.7%).

[0287] The association between the volume predicted by mCTA perfusion model and follow up infarct volume, and between the time dependent Tmax thresholded predicted infarct volume and follow up infarct volume in the whole test cohort is shown in Table 5.

TABLE-US-00005 TABLE 5 Comparisons between infarct volumes predicted by the derived mCTA models and CTP vs. follow up infarct volume (median, 24.8; IQR, 10.5-58.8 mL) in the test cohort (n = 144). Time dependent mCTA Tmax Infarct and mCTA thresholded Penumbra Perfusion model P model model (CTP) value Predicted volume 37.3 40.5 38.3 .67 (median [IQR], [21.3, 57.8] [22.9, 63] [15.0, 65.5] mL) Volume 21.7 20.4 22.3 .45 difference# [−44, 86.3] [−51.3, 92.1] [−42.6, 87.2] (mean [LoA], mL) DSC (median 22.5 21.7 23.2 .55 [IQR], %) [13.8, 30.4] [10.9, 31.2] [13.9, 33] CCC [95% CI] 0.43 0.41 0.45 N/A [0.18-0.58] [0.16-0.62] [0.32-0.54] ICC [CI] 0.5 0.47 0.54 N/A [0.29-0.64] [0.3-0.56] [0.3-0.64] IQR, interquartile range; LoA, limit of agreement; CI, confidence interval; DSC: Dice similarity coefficient between the predicted volume and follow up infarct volume; CCC: concordance correlation coefficient; ICC: intra-class correlation coefficient. N/A: Not applicable. #Volume difference is defined as follow up infarct volume minus model prediction, generated from Bland-Altman analysis

Study Discussion

[0288] Multiphase CT angiography (mCTA) is a quick and easy-to-use imaging tool in selecting patients with acute ischemic stroke (AIS) for revascularization therapy. .sup.10 The developed machine learning technique described in this study shows that tissue status can be automatically predicted from the mCTA just as it is currently done using CT perfusion imaging. These results demonstrate that mCTA using the methods proposed here has similar ability to CTP imaging in predicting tissue fate.

[0289] As such, the methodologies described herein can help physicians make clinical decisions regarding acute stroke treatment, especially in hospitals without CTP capabilities.

[0290] When comparing mCTA predicted infarct volume with follow up infarct volume in patients who achieved acute reperfusion (mTICI 2b/2c/3), the mean volume difference of 21.7 mL, CCC of 0.4, and ICC of 0.42 are fair. The mCTA predicted infarct volume agrees better with follow up infarct volume in patients who did not achieve acute reperfusion (mTICI 0/1/2) with a mean volume difference of 3.4 mL, CCC of 0.52, and ICC of 0.6. DSCs between mCTA predicted infarct and penumbra and follow up volume are relatively low (less than 30%). However, accurate spatial quantification of infarction in patients with AIS is complicated and likely influenced by many pathophysiological factors, such as collateral status, tissue tolerance to ischemia/hypoxia, cerebral autoregulation, leukoaraiosis, fluctuation in blood pressure, hyperglycemia and time to reperfusion. .sup.6, 7, 8 All these factors likely lead to discrepancy between infarct volume predicted at baseline and follow up imaging.

[0291] Of note, a recent paper from the HERMES group that used validated CTP software (i.e. RAPID, iSchemaView, Menlo Park, CA) showed similar DSC (median, 0.24; IQR, 0.15-0.37) between CTP predicted infarct volume and follow up infarct volume. .sup.19 When comparing mCTA predicted perfusion maps with CTP time dependent Tmax thresholded maps, the results show stronger agreement between the two measurements with a mean volume difference of 4.6 mL, CCC of 0.63, and ICC of 0.68. The median DSC of 40.5% between the mCTA predicted perfusion and CTP volume was also reasonable, suggesting good spatial overlap.

[0292] Imaging paradigms currently used in selecting patients with AIS for treatment include non-contrast CT, single-phase CTA, or CTP. CTP, however, requires additional radiation and contrast and specific acquisition protocols that are different from NCCT and CTA. CTP is sensitive to patient motion, a feature that invalidates that tool in almost 10 to 25% of patients. .sup.20 Eleven patients were excluded from this study as CTP maps generated by the software were corrupted due to the excessive patient motion during acquisition (FIG. 5A). Multiphase CTA is potentially less prone to patient motion while being capable of predicting perfusion maps appropriately (Supplements: FIG. 5A). CTP also adds additional costs to the health system that include cost of the scan and resources allocated to training personnel to acquire the scan. .sup.21, 22 Multiphase CTA has minimal additional radiation, no additional contrast material, whole-brain coverage, relative insensitivity to patient motion and minimal additional cost to acquire over a single phase CTA. .sup.10, 12, 13 The technique described here automates mCTA interpretation and provides physicians with tissue prediction as well as perfusion maps, just as CTP imaging does, thus potentially increasing their confidence in decision making (Table 5). FIG. 8 shows an example of penumbra and infarct core (and mismatch) predicted using mCTA. Of note, physiological definitions of infarct and penumbra are different from the operational definitions used in this study. Unlike conventional mCTA but similar to CTP, the technique described in the study is capable of detecting smaller perfusion lesions in the entire brain including the posterior circulation (Supplements: FIG. 8).

[0293] A strength of the developed machine learning technique is that it does not rely on deconvolution algorithms, which plays an essential role in current CTP processing. Although deconvolution methods can appropriately model perfusion status, the introduction of physiological variations in arterial delivery of contrast, the effects of collateral flow, and venous outflow components of cerebral perfusion, greatly increase the computational complexity. .sup.23, 24 The number of variables and the algorithms used to calculate these variables results in variability in generating CTP threshold values for estimating infarct and penumbra across different vendor software. Additionally, numerical solutions to deconvolution greatly relies on accurate selection of artery input function (AIF), a parameter that is case dependent and sensitive to noise, especially given the low signal to noise ratio of perfusion images, even when preprocessing, such as motion correction, temporal and spatial smoothing, and deconvolution regularization are applied. .sup.25, 26. The deconvolution free approach developed in this study can be easily integrated into any imaging paradigm using NCCT and mCTA as a post-processing step, potentially obviating the need for CTP.

[0294] In conclusion, infarct core, penumbra, and perfusion status can be automatically predicted from multiphase CTA imaging using machine learning. This technique shows comparable accuracy to CT perfusion imaging in measuring tissue status in patients with acute ischemic stroke. This work has the potential of assisting physicians in making treatment decisions in clinical settings.

[0295] Although the present invention has been described and illustrated with respect to preferred embodiments and preferred uses thereof, it is not to be so limited since modifications and changes can be made therein which are within the full, intended scope of the invention as understood by those skilled in the art.