Computed Tomography Perfusion-Based Machine Learning Model Better Predicts Follow-Up Infarction in Patients With Acute Ischemic Stroke

BACKGROUND AND PURPOSE

Prediction of infarct extent among patients with acute ischemic stroke using computed tomography perfusion is defined by predefined discrete computed tomography perfusion thresholds. Our objective is to develop a threshold-free computed tomography perfusion-based machine learning (ML) model to predict follow-up infarct in patients with acute ischemic stroke.

METHODS

Sixty-eight patients from the PRoveIT study (Measuring Collaterals With Multi-Phase CT Angiography in Patients With Ischemic Stroke) were used to derive a ML model using random forest to predict follow-up infarction voxel by voxel, and 137 patients from the HERMES study (Highly Effective Reperfusion Evaluated in Multiple Endovascular Stroke Trials) were used to test the derived ML model. Average map, T_max, cerebral blood flow, cerebral blood volume, and time variables including stroke onset-to-imaging and imaging-to-reperfusion time, were used as features to train the ML model. Spatial and volumetric agreement between the ML model predicted follow-up infarct and actual follow-up infarct were assessed. Relative cerebral blood flow <0.3 threshold using RAPID software and time-dependent T_max thresholds were compared with the ML model.

RESULTS

In the test cohort (137 patients), median follow-up infarct volume predicted by the ML model was 30.9 mL (interquartile range, 16.4-54.3 mL), compared with a median 29.6 mL (interquartile range, 11.1-70.9 mL) of actual follow-up infarct volume. The Pearson correlation coefficient between 2 measurements was 0.80 (95% CI, 0.74-0.86, P<0.001) while the volumetric difference was -3.2 mL (interquartile range, -16.7 to 6.1 mL). Volumetric difference with the ML model was smaller versus the relative cerebral blood flow <0.3 threshold and the time-dependent T_max threshold (P<0.001).

CONCLUSIONS

A ML using computed tomography perfusion data and time estimates follow-up infarction in patients with acute ischemic stroke better than current methods.

Confirmatory Study of Time-Dependent Computed Tomographic Perfusion Thresholds for Use in Acute Ischemic Stroke

Background and Purpose—

Computed tomographic perfusion (CTP) thresholds associated with follow-up brain infarction may differ by time from symptom onset to imaging and reperfusion. We confirm CTP thresholds over time to imaging and reperfusion in patients with acute ischemic stroke from the HERMES collaboration (Highly Effective Reperfusion Evaluated in Multiple Endovascular Stroke Trials) data.

Methods—

Patients with occlusion on CT angiography were acutely imaged with CTP. Noncontrast CT and magnetic resonance-diffusion weighted imaging at 24 to 48 hours defined follow-up infarction. Reperfusion was assessed on conventional angiogram. Tmax, cerebral blood flow (CBF), and cerebral blood volume maps were derived from delay-insensitive CTP postprocessing. These parameters were analyzed using receiver operator characteristics to derive optimal thresholds based on time from stroke onset-to-CTP or to reperfusion. ANOVA and linear regression were used to test whether the derived CTP thresholds were different by time.

Results—

One hundred thirty-seven patients were included. Tmax thresholds of >15.7 s and >15.8 s and absolute CBF thresholds of <8.9 and <7.5 mL·min −1 ·100 g −1 for gray matter and white matter respectively were associated with infarct if reperfusion was achieved <90 minutes from CTP with stroke onset-to-CTP <180 minutes. The discriminative ability of cerebral blood volume was modest. There were no statistically significant relationships between stroke onset-to-CTP time and Tmax, CBF, and cerebral blood volume thresholds (all P >0.05). A statistically significant relationship was observed between CTP-to-reperfusion time and the optimal thresholds for Tmax ( P <0.001) and CBF ( P <0.001). Similar but more modest relationship was noted for onset-to-reperfusion time and optimal thresholds for CBF ( P ≤0.01).

Conclusions—

CTP thresholds based on stroke onset and imaging time and taking into account time needed for reperfusion may improve infarct prediction in patients with acute ischemic stroke.

Abstract WMP29: Detection and Segmentation of Subarachnoid Hemorrhages With Deep Learning

Introduction: About 5% of all strokes are subarachnoid hemorrhages (SAHs). Accurate segmentation and detection of SAH in CT scans is important because the SAH volume is linked to delayed cerebral ischemia and poor patient outcome. SAH segmentation is a difficult task with high interobserver variability. Previous studies reported a limited average Dice of 0.66 between expert radiologists. We studied the use of deep learning for fast and accurate detection and segmentation of SAH in CT scans.

Methodology: A convolutional neural network (CNN) was designed for segmentation and detection of SAH in CT scans. This CNN was trained to classify each voxel as “SAH” or “no SAH”. The input of this CNN was the local information around the voxel and its output was the probability that this voxel is “SAH” or “no SAH”. We used active learning for optimal selection of training samples. By doing so, we boosted the learning performance while keeping a balanced representation of both classes of voxels.

Results: The CNN segmented 469 scans from SAH patients and achieved an average Dice of 0.62 ± 0.16 [0.19, 0.92] (avg. ± std. [min., max.]). The CNN also segmented 383 scans from patients with repeated bleeding after an initial SAH. Because most of these patients were treated with endovascular coiling, strong metal artefacts were present in these 383 scans. Regardless of these metal artefacts, an average Dice of 0.66 ± 0.19 [0.05, 0.92] was achieved. Finally, the CNN was used for detection of SAH in 35 scans from SAH patients and 35 images from ischemic stroke patients. The proposed CNN achieved an accuracy of 96% for detection of SAH in these 70 images. The CNN took less than 40 seconds to analyze a CT scan.

Conclusion: The proposed CNN can produce SAH segmentations with a similar accuracy as expert radiologists. This CNN was able to reach state-of-the-art SAH segmentation results and SAH detection accuracy.

Abstract WP70: Artificial Intelligence Based Detection of Large Vessel Occlusion on Non-Contrast Computed Tomography in Stroke

Introduction: Rapid detection of a large vessel occlusion (LVO) in acute ischemic stroke is beneficial as it may accelerate therapeutic management. The hyperdense artery sign (HAS) on non-contrast CT (NCCT) is result of an acute thrombolytic occlusion and is known as the earliest radiological marker to identify LVO. However, HAS is not apparent for all LVOs and detection by a less-trained eye can be sub-optimal. Artificial intelligence (AI) may support the non-trained radiologist in swift and accurate LVO detection. We evaluated the LVO detection accuracy of AI, and compared this to human expert performance.

Materials and Methods: We used a convolutional neural network (CNN) developed by Nico.lab (www.nico-lab.com) that automatically detects and segments thrombi on NCCT using a patch-based approach. The CNN was trained using thrombus and non-thrombus patches, combined with the contralateral side. In a retrospective analysis of thin-slice NCCTs, obtained from Amsterdam UMC patients, with (ICA-T, M1, or M2) and without occlusions (stroke mimics), the CNN and two observers (>15y and >4y of experience) assessed LVO-presence and location. Ground truth was established by consensus between two different experts with >5y of experience, using both CTA and NCCT. Thrombus segmentations were considered accurate if segmentations overlapped with the ground truth. Sensitivity and specificity of the LVO detection were assessed.

Results: We included 107 patients, of which 59 had proven LVOs. Nico.lab CNN showed a thrombus segmentation accuracy of 81%, vs. 81% and 77% of human experts. Sensitivity of LVO detection by CNN was 0.86, vs. 0.95 and 0.79 for human observers. Specificity was 0.65 for the CNN vs. 0.58 and 0.82 for human experts respectively.

Conclusion: AI-based LVO detection on NCCT showed comparable results to expert observers. This supportive tool could lead to earlier detection of LVO in acute stroke situations. Fig 1A. HAS; B. CNN output (red); C. clot on CTA (blue arrow)

Abstract TP80: Deep Learning Based Prediction of Tissue Status From Native CT Perfusion Images

Introduction: CT Perfusion (CTP) imaging allows estimations of the ischemic core and penumbra. We hypothesize that a deep learning approach improves upon classic deconvolution analysis.

Method: We trained a deep neural network to predict the final infarct volume based on the native CTP images, time to reperfusion and reperfusion status in a derivation cohort (MR CLEAN study). Given a CTP scan, prediction of the final infarct with simulated treatment parameters (instant reperfusion or no reperfusion at all), allows estimation of core and perfusion lesion. The model was validated in a five-fold cross-validation on the derivation cohort and, without retraining, on an external dataset (CRISP study). We calculated the mean absolute difference (MAD) between the volumes of the predicted core/perfusion lesion and true final infarct in reperfusers and non-reperfusers (based on 24h MRA and mTICI at the end of the procedure: 0 vs 2B-3) and compared these to the volume estimations of CTP processing by IschemaView RAPID. We assumed the measured final infarct to correspond to the core at baseline in reperfused patients and to the perfusion lesion in non-reperfused patients.

Results: In the derivation cohort the deep learning prediction improved core prediction in reperfusers (MAD of 34 ml vs 66 ml, p<0.001) and provided equivalent results in non-reperfusers for perfusion lesion prediction (both MAD of 35ml, p=0.5). Validation on the CRISP dataset showed improved core prediction (MAD of 32 ml vs 37 ml, p<0.001) and was inconclusive for non-reperfusers (n=3).

Conclusion: We developed and validated a deep learning based method which improved volume estimations compared to classic CTP processing.

Abstract WP5: Reperfusion Reduces Brain Edema in Patients With Acute Ischemic Stroke. A MR CLEAN Substudy

Introduction: The main objective of this study was to elucidate the relation between reperfusion and brain edema. The secondary objective was to evaluate if brain edema partially mediated worse outcome in patients without reperfusion, or in patients with delayed reperfusion or a lower ASPECTS.

Methods: MR CLEAN was a prospective, randomized, multicenter clinical trial of endovascular treatment (EVT) compared to conventional care. In this exploratory post hoc study, 462 of 500 patients were included. Brain edema was assessed by midline shift (MLS) measured on follow-up CT scans. Reperfusion status was assessed by the modified thrombolysis in cerebral infarction (mTICI) score in the EVT arm. The modified Arterial Occlusive Lesion (mAOL) score was used to evaluate the recanalization status in both arms. The associations between MLS and modified Rankin Scale (mRS), and MLS with reperfusion or recanalization were performed in univariable and multivariable analyses. Mediation analyses were performed to determine whether MLS was partially responsible for worse functional outcome in the groups without reperfusion or recanalization, and in the groups with delayed reperfusion or lower ASPECTS.

Results: Reperfusion and recanalization were associated with reduced likelihood of having MLS (acOR 0.26, 95%CI 0.13-0.51, p < 0.001; acOR 0.38, 95%CI 0.24-0.60, p < 0.001 respectively). Accordingly, the presence of MLS was associated with an increased risk of worse 90-day outcome (acOR 3.89, CI 2.71-5.61, p < 0.001). In mediation analysis, MLS was partially responsible for worse mRS scores in patients without reperfusion or recanalization, and in those with delayed reperfusion or lower ASPECTS (MLS changed the logistic regression coefficients by 30.3%, 12.6%, 33.3%, and 64.2%, respectively).

Conclusion: Successful reperfusion or recanalization reduces brain edema, reflected by MLS. The presence of MLS mediated part of the worse functional outcomes in patients without reperfusion or in those patients with delayed reperfusion or a lower ASPECTS. This study confirms a pleiotropic benefit of reperfusion that extends beyond the concept of rescuing ischemic brain tissue at risk for infarction.

Associations of Ischemic Lesion Volume With Functional Outcome in Patients With Acute Ischemic Stroke

Background and Purpose—

Ischemic lesion volume (ILV) on noncontrast computed tomography at 1 week can be used as a secondary outcome measure in patients with acute ischemic stroke. Twenty-four–hour ILV on noncontrast computed tomography has greater availability and potentially allows earlier estimation of functional outcome. We aimed to assess lesion growth 24 hours after stroke onset and compare the associations of 24-hour and 1-week ILV with functional outcome.

Methods—

We included 228 patients from MR CLEAN trial (Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands), who received noncontrast computed tomography at 24-hour and 1-week follow-up on which ILV was measured. Relative and absolute lesion growth was determined. Logistic regression models were constructed either including the 24-hour or including the 1-week ILV. Ordinal and dichotomous (0–2 and 3–6) modified Rankin scale scores were, respectively, used as primary and secondary outcome measures.

Results—

Median ILV was 42 mL (interquartile range, 21–95 mL) and 64 mL (interquartile range: 30–120 mL) at 24 hours and 1 week, respectively. Relative lesion growth exceeding 30% occurred in 121 patients (53%) and absolute lesion growth exceeding 20 mL occurred in 83 patients (36%). Both the 24-hour and 1-week ILVs were similarly significantly associated with functional outcome (both P <0.001). In the logistic analyses, the areas under the curve of the receiver–operator characteristic curves were similar: 0.85 (95% confidence interval, 0.80–0.90) and 0.87 (95% confidence interval, 0.82–0.91) for including the 24-hour and 1-week ILV, respectively.

Conclusions—

Growth of ILV is common 24-hour poststroke onset. Nevertheless, the 24-hour ILV proved to be a valuable secondary outcome measure as it is equally strongly associated with functional outcome as the 1-week ILV.

Clinical Trial Registration—

URL: http://www.isrctn.com . Unique identifier: ISRCTN10888758.

Association of Automatically Quantified Total Blood Volume after Aneurysmal Subarachnoid Hemorrhage with Delayed Cerebral Ischemia

BACKGROUND AND PURPOSE

The total amount of extravasated blood after aneurysmal subarachnoid hemorrhage, assessed with semiquantitative methods such as the modified Fisher and Hijdra scales, is known to be a predictor of delayed cerebral ischemia. However, prediction rates of delayed cerebral ischemia are moderate, which may be caused by the rough and observer-dependent blood volume estimation used in the prediction models. We therefore assessed the association between automatically quantified total blood volume on NCCT and delayed cerebral ischemia.

MATERIALS AND METHODS

We retrospectively studied clinical and radiologic data of consecutive patients with aneurysmal SAH admitted to 2 academic hospitals between January 2009 and December 2011. Adjusted ORs with associated 95% confidence intervals were calculated for the association between automatically quantified total blood volume on NCCT and delayed cerebral ischemia (clinical, radiologic, and both). The calculations were also performed for the presence of an intraparenchymal hematoma and/or an intraventricular hematoma and clinical delayed cerebral ischemia.

RESULTS

We included 333 patients. The adjusted OR of total blood volume for delayed cerebral ischemia (clinical, radiologic, and both) was 1.02 (95% CI, 1.01-1.03) per milliliter of blood. The adjusted OR for the presence of an intraparenchymal hematoma for clinical delayed cerebral ischemia was 0.47 (95% CI, 0.24-0.95) and of the presence of an intraventricular hematoma, 2.66 (95% CI, 1.37-5.17).

CONCLUSIONS

A higher total blood volume measured with our automated quantification method is significantly associated with delayed cerebral ischemia. The results of this study encourage the use of rater-independent quantification methods in future multicenter studies on delayed cerebral ischemia prevention and prediction.

Automatic quantification of subarachnoid hemorrhage on noncontrast CT

BACKGROUND AND PURPOSE

Quantification of blood after SAH on initial NCCT is an important radiologic measure to predict patient outcome and guide treatment decisions. In current scales, hemorrhage volume and density are not accounted for. The purpose of this study was to develop and validate a fully automatic method for SAH volume and density quantification.

MATERIALS AND METHODS

The automatic method is based on a relative density increase due to the presence of blood from different brain structures in NCCT. The method incorporates density variation due to partial volume effect, beam-hardening, and patient-specific characteristics. For validation, automatic volume and density measurements were compared with manual delineation on NCCT images of 30 patients by 2 radiologists. The agreement with the manual reference was compared with interobserver agreement by using the intraclass correlation coefficient and Bland-Altman analysis for volume and density.

RESULTS

The automatic measurement successfully segmented the hemorrhage of all 30 patients and showed high correlation with the manual reference standard for hemorrhage volume (intraclass correlation coefficient = 0.98 [95% CI, 0.96-0.99]) and hemorrhage density (intraclass correlation coefficient = 0.80 [95% CI, 0.62-0.90]) compared with intraclass correlation coefficient = 0.97 (95% CI, 0.77-0.99) and 0.98 (95% CI, 0.89-0.99) for manual interobserver agreement. Mean SAH volume and density were, respectively, 39.3 ± 31.5 mL and 62.2 ± 5.9 Hounsfield units for automatic measurement versus 39.7 ± 32.8 mL and 61.4 ± 7.3 Hounsfield units for manual measurement. The accuracy of the automatic method was excellent, with limits of agreement of -12.9-12.1 mL and -7.6-9.2 Hounsfield units.

CONCLUSIONS

The automatic volume and density quantification is very accurate compared with manual assessment. As such, it has the potential to provide important determinants in clinical practice and research.

Automated cerebral infarct volume measurement in follow-up noncontrast CT scans of patients with acute ischemic stroke

BACKGROUND AND PURPOSE

Cerebral infarct volume as observed in follow-up CT is an important radiologic outcome measure of the effectiveness of treatment of patients with acute ischemic stroke. However, manual measurement of CIV is time-consuming and operator-dependent. The purpose of this study was to develop and evaluate a robust automated measurement of the CIV.

MATERIALS AND METHODS

The CIV in early follow-up CT images of 34 consecutive patients with acute ischemic stroke was segmented with an automated intensity-based region-growing algorithm, which includes partial volume effect correction near the skull, midline determination, and ventricle and hemorrhage exclusion. Two observers manually delineated the CIV. Interobserver variability of the manual assessments and the accuracy of the automated method were evaluated by using the Pearson correlation, Bland-Altman analysis, and Dice coefficients. The accuracy was defined as the correlation with the manual assessment as a reference standard.

RESULTS

The Pearson correlation for the automated method compared with the reference standard was similar to the manual correlation (R = 0.98). The accuracy of the automated method was excellent with a mean difference of 0.5 mL with limits of agreement of -38.0-39.1 mL, which were more consistent than the interobserver variability of the 2 observers (-40.9-44.1 mL). However, the Dice coefficients were higher for the manual delineation.

CONCLUSIONS

The automated method showed a strong correlation and accuracy with the manual reference measurement. This approach has the potential to become the standard in assessing the infarct volume as a secondary outcome measure for evaluating the effectiveness of treatment.