Multiphase computed tomographic angiography (mCTA) provides time variant images of pial vasculature supplying brain in patients with acute ischemic stroke (AIS). To develop a machine learning (ML) technique to predict tissue perfusion and infarction from mCTA source images.
284 patients with AIS were included from the Precise and Rapid assessment of collaterals using multi-phase CTA in the triage of patients with acute ischemic stroke for Intra-artery Therapy (Prove-IT) study. All patients had non-contrast computed tomography, mCTA, and computed tomographic perfusion (CTP) at baseline and follow-up magnetic resonance imaging/non-contrast-enhanced computed tomography. Of the 284 patient images, 140 patient images were randomly selected to train and validate three ML models to predict a pre-defined Tmax thresholded perfusion abnormality, core and penumbra on CTP. The remaining 144 patient images were used to test the ML models. The predicted perfusion, core and penumbra lesions from ML models were compared to CTP perfusion lesion and to follow-up infarct using Bland-Altman plots, concordance correlation coefficient (CCC), intra-class correlation coefficient (ICC), and Dice similarity coefficient.
Mean difference between the mCTA predicted perfusion volume and CTP perfusion volume was 4.6 mL (limit of agreement [LoA], –53 to 62.1 mL;
A ML based mCTA model is able to predict brain tissue perfusion abnormality and follow-up infarction, comparable to CTP.
Infarct predicted using computed tomographic perfusion (CTP) at admission is often used in treatment decision making in patients with acute ischemic stroke (AIS) [
We therefore aim to develop a machine learning (ML) based technique to estimate brain tissue perfusion abnormality and predict core and penumbra similar to what CTP does in patients with AIS.
Data were from the Precise and Rapid assessment of collaterals using multi-phase CTA in the triage of patients with acute ischemic stroke for Intra-artery Therapy (Prove-IT) study [
Subjects who had (1) baseline NCCT and mCTA; (2) baseline CTP imaging with ≥8 cm z axis coverage; (3) had reperfusion assessed on conventional angiography after thrombolysis treatment (intravenous tissue plasminogen activator, EVT, or both) with the modified thrombolysis in cerebral infarction [mTICI]); and (4) had 24/36-hour follow-up imaging on diffusion magnetic resonance imaging or NCCT were included in this analysis. Patient inclusion and exclusion are shown in
NCCT with 5 mm slice thickness was obtained, followed by mCTA with arch to vertex coverage in the first (arterial) and skull base to vertex coverage the second (peak venous) and third (late venous) phase. Detailed mCTA acquisition parameters have been published previously [
Iodinated contrast agent 45 mL were injected at a rate of 4.5 mL/sec followed by a 40 mL saline bolus injected at a rate of 6 mL/sec. Image acquisition started 5 seconds after contrast injection and 24 passes over 66 seconds were performed with 5 mm section thickness and a cranio-caudal coverage of 8 cm.
Each CTP study was processed using commercially available delay-insensitive deconvolution software (CT Perfusion 4D, GE Healthcare, Waukesha, WI, USA). Absolute maps of cerebral blood flow (mL/min/100 g), cerebral blood volume (mL/100 g), and Tmax (seconds) were generated. Average maps were created by averaging the dynamic CTP source images. Time-dependent Tmax thresholds confirmed previously (
NCCT and mCTA images were first skull stripped [
The perfusion map used in this analysis was a Tmax map thresholded using previously published time-dependent thresholds [
We developed three ML models: (1) core model; (2) penumbra model; (3) perfusion model. A 2-stage training mechanism was developed to train two ML models to predict core and penumbra respectively. The detailed training and testing strategy is shown in
Further, the 140 patient images used for training and internally validating the penumbra and core models were reused to train and internally validate the third random forest classifier (perfusion model). For deriving and testing this model, time-dependent Tmax thresholded maps from CTP were used as reference standard [
All three random forest models shared the same self-designed features as inputs. NCCT Hounsfield units (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 TIC for each voxel and used for deriving and testing the three random forest classifiers. These were: (1) average and standard deviation of 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. 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. 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 internal 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 (determined from the validation cohort), followed by isolated island removal and morphological operation, to generate the mCTA predicted volume.
T test for normally distributed data, Fisher’s exact test for categorical data and the Rank sum test for non-normally distributed data was used to analyze any differences between groups. Time-dependent Tmax thresholded volumes (CTP volume) were used as reference standard to evaluate the mCTA predicted perfusion volume in the test cohort [
Patient characteristics are summarized in
Infarct volumes predicted by the mCTA models and those by CTP versus the reference standard (follow-up infarct volume) in the test cohort (n=144) are shown in
mCTA with its simple three phase image acquisition protocol is a quick and easy-to-implement imaging tool used in patients with AIS [
Imaging paradigms currently used for selecting patients with AIS for treatment include non-contrast CT, single-phase CTA, or CTP. Although CTP is widely used in large comprehensive stroke centers, its adoption in smaller, less academic hospitals and in primary stroke centers continues to be limited. Concerns about technical expertise needed for its implementation and, the additional radiation and contrast needed when compared to a stroke protocol based on NCCT and single-phase CTA limit its acceptability in smaller hospitals. CTP is also sensitive to patient motion, a feature that invalidates that tool in almost 10 to 25% of patients [
A strength of the developed ML 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 [
The correlation between the mCTA predicted core and penumbra volume and follow-up infarct volume are moderate with CCC and ICC ranging 0.4 to 0.5. The correlation between the mCTA perfusion volume and CTP perfusion volume is stronger with CCC and ICC of >0.6. The spatial overlap between the ML predicted volume or the CTP predicted volume and follow- up infarct volume appears weak with DSC of <30%. The moderate volume correlation and weak DSC can be partially attributed to infarct growth, which can occur despite endovascular reperfusion because of delay between imaging and reperfusion or incomplete reperfusion. Moreover, accurate quantification of ischemic infarct and penumbra in patients with AIS is complex and likely influenced by many pathophysiological factors, such as cerebral autoregulation, collateral responsiveness, tissue tolerance to ischemia and hypoxia, leukoaraiosis, etc. Weak DSC can also be explained by the limitations of co-registering different imaging modalities [
This study has several limitations. First, accurate quantification of ischemic infarct and penumbra in patients with AIS is complex and likely influenced by infarct location, time factors including stroke symptom onset-to-imaging and imaging- to-reperfusion time, and many pathophysiological factors, such as cerebral autoregulation, collateral responsiveness, tissue tolerance to ischemia and hypoxia, leukoaraiosis, etc [
In conclusion, as with CTP, core, penumbra, and perfusion status can be automatically predicted from mCTA imaging using ML. This work, therefore, has future potential for assisting physicians in making treatment decisions in clinical settings where CTP is not available.
Supplementary materials related to this article can be found online at
Optimal Tmax thresholds for infarction when reperfused <90, 90 to 180 minutes, and not reperfused
Patient characteristics in patients with mTICI 2b/3 and 0/1
Predicted volumes of different models compared to follow-up infarct volume between the patients with anterior circulation (ICA, MCA, ACA) and with posterior circulation (vertebral and basilar) occlusions
The authors have no financial conflicts of interest.
This study is funded through an operating grant from the Canadian Institute of Health Research (CIHR) and Alberta Innovate: Health Solution (AIHS).
Patient inclusion chart. CTP, computed tomographic perfusion.
Training and testing strategy of machine learning models to predict core, penumbra and perfusion status. (A) Derivation and testing of penumbra model and infarction model using follow-up infarct as reference standard. (B) Derivation and testing of the perfusion model using time-dependent Tmax thresholded map as reference standard. mCTA, multiphase computed tomographic angiography.
Multiphase computed tomographic angiography (mCTA) predicted infarct map compared to computed tomographic perfusion (CTP) time-dependent Tmax thresholded map when compared to follow-up infarct. (A) Patient who achieved reperfusion (modified thrombolysis in cerebral infarction [mTICI] 2b), (B) patient who did not achieve reperfusion, and (C) patient who achieved reperfusion (mTICI 3). Columns: mCTA phase 1 to 3, mCTA predicted perfusion maps, mCTA predicted core (red in column 5) and penumbra (blue in column 5) overlaid on the mCTA predicted perfusion map, CTP Tmax maps, CTP time-dependent Tmax threshold 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 core model.
Bland-Altman plots of (A) multiphase computed tomographic angiography (mCTA) infarct volume predicted using the penumbra model versus follow-up infarct volume for the 44 patients who did not achieve acute reperfusion; (B) mCTA infarct volume predicted using core model versus follow-up infarct volume for the 100 patients who achieved reperfusion; and (C) mCTA perfusion volume predicted using perfusion model versus time-dependent Tmax predicted infarct volume for all 144 patients in the test cohort. CTP, computed tomographic perfusion; SD, standard deviation.
An example shows the computed tomographic perfusion (CTP) maps (column 1–3) due to the excessive movement of the patient during CTP acquisition, versus multiphase computed tomographic angiography (mCTA) prediction (column 4) that correlates well with follow-up imaging (column 5). CBF, cerebral blood flow; CBV, cerebral blood volume.
An example shows the multiphase computed tomographic angiography (mCTA) prediction, computed tomographic perfusion map, and follow-up imaging of a patient with posterior circulation occlusion.
Failure cases from multiphase computed tomographic angiography (mCTA) prediction. (A) Row shows images from a patient who presented ultraearly with an onset-to computed tomography time of 21 minutes. The mCTA model significantly over-predicts follow-up infarct. (B) Row shows images from a patient without obvious occlusion; the mCTA model shows a false positive perfusion abnormality in the left posterior occipital region. (C) Row shows images of a patient with an internal carotid artery occlusion; the mCTA model under-estimates the perfusion abnormality. Column 1–3: mCTA predicted follow-up infarct, Tmax, and follow-up infarct imaging. NCCT, non-contrast-enhanced computed tomography; DWI, diffusion-weighted imaging.
Patient characteristics in the derivation and test cohorts in the study
Characteristic | Derivation cohort (n=140) | Test cohort (n=144) | |
---|---|---|---|
Age (yr) | 73 (62–79) | 72 (62–80) | 0.73 |
Male sex | 80 (57) | 77 (53) | 0.56 |
Baseline NIHSS | 17 (7–23) | 14 (6–18) | 0.12 |
Baseline ASPECTS | 9 (8–10) | 9 (8–10) | 0.15 |
Onset-to-imaging time (min) | 131 (94–226) | 139 (88–294) | 0.35 |
Imaging-to-reperfusion time (min) | 90 (68–115) | 87 (64–125) | 0.97 |
Onset-to-reperfusion time (min) | 245 (172–330) | 240 (181–377) | 0.71 |
Follow-up infarct volume (mL) | 22.2 (10.3–59.4) | 25.9 (10.1–60.6) | 0.60 |
Site of occlusion | |||
ICA | 22 (16) | 26 (18) | 0.76 |
MCA:M1 | 73 (52) | 70 (48) | 0.64 |
Distal M2, M3, M4, P2, P3, A2, A3, vertebral artery, basilar artery | 45 (32) | 48 (33) | 0.63 |
Values are presented as median (interquartile range) or number (%).
NIHSS, National Institutes of Health Stroke Scale; ASPECTS, Alberta Stroke Program Early CT score; ICA, internal carotid artery; MCA, middle cerebral artery.
Statistical comparison between infarct volumes predicted by the mCTA machine learning models versus those by CTP (time-dependent Tmax thresholds as per literature [
Variable | mCTA core and penumbra model | mCTA tissue perfusion model | CTP Tmax thresholded model (CTP) [ |
|
---|---|---|---|---|
Predicted volume (median [IQR], mL) | 37.3 (21.3 to 57.8) | 40.5 (22.9 to 63) | 38.3 (15.0 to 65.5) | 0.67 |
Volume difference |
21.7 (–44 to 86.3) | 20.4 (–51.3 to 92.1) | 22.3 (–42.6 to 87.2) | 0.45 |
DSC (median [IQR], %) | 22.5 (13.8 to 30.4) | 21.7 (10.9 to 31.2) | 23.2 (13.9 to 33) | 0.55 |
CCC (95% CI) | 0.43 (0.18 to 0.58) | 0.41 (0.16 to 0.62) | 0.45 (0.32 to 0.54) | NA |
ICC (95% CI) | 0.5 (0.29 to 0.64) | 0.47 (0.3 to 0.56) | 0.54 (0.3 to 0.64) | NA |
mCTA, multiphase computed tomographic angiography; CTP, computed tomographic perfusion; IQR, interquartile range; LoA, limit of agreement; DSC, Dice similarity coefficient; CCC, concordance correlation coefficient; CI, confidence interval; NA, not applicable; ICC, intra-class correlation coefficient.
Volume difference is defined as follow-up infarct volume minus model prediction, generated from Bland-Altman analysis.