Automated brain computed tomographic densitometry of early ischemic changes in acute stroke

Validation of two automated ASPECTS software on non-contrast computed tomography scans of patients with acute ischemic stroke

Purpose

The Alberta Stroke Program Early Computed Tomography Score (ASPECTS) was designed for semi-quantitative assessment of early ischemic changes on non-contrast computed tomography (NCCT) for acute ischemic stroke (AIS). We evaluated two automated ASPECTS software in comparison with reference standard.

Methods

NCCT of 276 AIS patients were retrospectively reviewed (March 2018–June 2020). A three-radiologist consensus for ASPECTS was used as reference standard. Imaging data from both baseline and follow-up were evaluated for reference standard. Automated ASPECTS were calculated from baseline NCCT with 1-mm and 5-mm slice thickness, respectively. Agreement between automated ASPECTS and reference standard was assessed using intra-class correlation coefficient (ICC). Correlation of automated ASPECTS with baseline stroke severity (NIHSS) and follow-up ASPECTS were evaluated using Spearman correlation analysis.

Results

In score-based analysis, automated ASPECTS calculated from 5-mm slice thickness images agreed well with reference standard (software A: ICC = 0.77; software B: ICC = 0.65). Bland–Altman analysis revealed that the mean differences between automated ASPECTS and reference standard were ≤ 0.6. In region-based analysis, automated ASPECTS derived from 5-mm slice thickness images by software A showed higher sensitivity (0.60 vs. 0.54), lower specificity (0.91 vs. 0.94), and higher AUC (0.76 vs. 0.74) than those using 1-mm slice thickness images (p < 0.05). Automated ASPECTS derived from 5-mm slice thickness images by software B showed higher sensitivity (0.56 vs. 0.51), higher specificity (0.87 vs. 0.81), higher accuracy (0.80 vs. 0.73), and higher AUC (0.71 vs. 0.66) than those using 1-mm slice thickness images (p < 0.05). Automated ASPECTS were significantly associated with baseline NIHSS and follow-up ASPECTS.

Conclusion

Automated ASPECTS showed good reliability and 5 mm was the optimal slice thickness.

An automated ASPECTS method with atlas-based segmentation

BACKGROUND AND PURPOSE

As a simple and reliable systematic method to evaluate the early ischemic changes in the blood supply region of the middle cerebral artery of patients with ischemic stroke, the Alberta Stroke Program Early CT score (ASPECTS) can be used for rapid semi-quantitative evaluation of ischemic lesions, which is helpful to select potential candidates for intravenous and intra-arterial therapies, determine the thrombolytic effect and long-term prognosis. This method mainly relies on doctors' visual observation. However, due to different levels of doctor's experience, the poor inter-reader agreement may result in errors in the final ASPECTS. The purpose of this work was to propose an automated semi-quantitative method for the diagnosis of acute ischemic stroke based on non-contrast computed tomography (NCCT), to provide a reference for doctors in the diagnosis and evaluation.

METHODS

NCCT data from a total of 90 patients were included for auto-ASPECTS training and testing. After preprocessing CT images, the regions of interest (ROI) for ASPECTS were labeled using atlas-based segmentation. The mean difference, mean ratio and brain density shifts (BDS) of the corresponding regions of the contralateral brain were used as the standard for quantitative analysis. The auto-ASPECTS method was developed and validated to predict early ischemic changes whose performance was evaluated by the agreement (accuracy) of predictions and consensus scores of two observers.

RESULTS

A comparison was made among the results on mean difference, mean ratio, BDS and the combination of multiple parameters as the standard. The result of using BDS alone was relatively better than the result of using any other parameter alone or any combination of multiple parameters, and accuracy in the test set was 0.80. In the test set, accuracy with using different BDS thresholds increased by 6.67% compared with using the consistent BDS threshold. After dichotomy of auto-ASPECTS and consensus scores with the threshold of 7, the agreement of them was 83.3% and there was no significant difference between the two distributions (p = 0.344) in McNemar test.

CONCLUSIONS

The proposed auto-ASPECTS method for NCCT images can provide useful information for early diagnosis and evaluation of patients with acute ischemic stroke (AIS).

The effects of the skull on CT imaging of the brain: a skull and brain phantom study

OBJECTIVE

To investigate the effects of beam hardening by the skull on the measured radiodensity of the brain. To test a hypothesis that these effects of beam hardening are decreased using a monochromatic energy source.

METHODS

Selected clinical cases were reviewed in illustration. An anthropomorphic skull and brain phantom was created and scanned in a clinical CT scanner with skull, without skull, and with hemicraniectomy. The effects of beam hardening were illustrated by scanning the phantom with mono- and poly-chromatic X-ray sources.

RESULTS

In clinical cases, the HU values of the brain were consistently lower when the X-ray beam traversed the skull than when it did not. An anthropomorphic skull-and-brain phantom further demonstrated these effects, which were evident with a polychromatic energy source and absent with a virtual monochromatic energy source.

CONCLUSIONS

Beam hardening by the skull lowers the measured HU values of the brain. The effects, which can impact quantitative imaging, may be mitigated by a virtual monochromatic energy source.

ADVANCES IN KNOWLEDGE

Beam hardening by the skull lowers the measured radiodensity of the brain. The effects may be mitigated by a virtual monochromatic energy source.

Intracranial Densitometry-Augmented Machine Learning Enhances the Prognostic Value of Brain CT in Pediatric Patients With Traumatic Brain Injury: A Retrospective Pilot Study

Background: The inter- and intrarater variability of conventional computed tomography (CT) classification systems for evaluating the extent of ischemic-edematous insult following traumatic brain injury (TBI) may hinder the robustness of TBI prognostic models. Objective: This study aimed to employ fully automated quantitative densitometric CT parameters and a cutting-edge machine learning algorithm to construct a robust prognostic model for pediatric TBI. Methods: Fifty-eight pediatric patients with TBI who underwent brain CT were retrospectively analyzed. Intracranial densitometric information was derived from the supratentorial region as a distribution representing the proportion of Hounsfield units. Furthermore, a machine learning-based prognostic model based on gradient boosting (i.e., CatBoost) was constructed with leave-one-out cross-validation. At discharge, the outcome was assessed dichotomously with the Glasgow Outcome Scale (favorability: 1-3 vs. 4-5). In-hospital mortality, length of stay (>1 week), and need for surgery were further evaluated as alternative TBI outcome measures. Results: Densitometric parameters indicating reduced brain density due to subtle global ischemic changes were significantly different among the TBI outcome groups, except for need for surgery. The skewed intracranial densitometry of the unfavorable outcome became more distinguishable in the follow-up CT within 48 h. The prognostic model augmented by intracranial densitometric information achieved adequate AUCs for various outcome measures [favorability = 0.83 (95% CI: 0.72-0.94), in-hospital mortality = 0.91 (95% CI: 0.82-1.00), length of stay = 0.83 (95% CI: 0.72-0.94), and need for surgery = 0.71 (95% CI: 0.56-0.86)], and this model showed enhanced performance compared to the conventional CRASH-CT model. Conclusion: Densitometric parameters indicative of global ischemic changes during the acute phase of TBI are predictive of a worse outcome in pediatric patients. The robustness and predictive capacity of conventional TBI prognostic models might be significantly enhanced by incorporating densitometric parameters and machine learning techniques.

In-Silico Trials for Treatment of Acute Ischemic Stroke

Despite improved treatment, a large portion of patients with acute ischemic stroke due to a large vessel occlusion have poor functional outcome. Further research exploring novel treatments and better patient selection has therefore been initiated. The feasibility of new treatments and optimized patient selection are commonly tested in extensive and expensive randomized clinical trials. in-silico trials, computer-based simulation of randomized clinical trials, have been proposed to aid clinical trials. In this white paper, we present our vision and approach to set up in-silico trials focusing on treatment and selection of patients with an acute ischemic stroke. The INSIST project (IN-Silico trials for treatment of acute Ischemic STroke, www.insist-h2020.eu) is a collaboration of multiple experts in computational science, cardiovascular biology, biophysics, biomedical engineering, epidemiology, radiology, and neurology. INSIST will generate virtual populations of acute ischemic stroke patients based on anonymized data from the recent stroke trials and registry, and build on the existing and emerging in-silico models for acute ischemic stroke, its treatment (thrombolysis and thrombectomy) and the resulting perfusion changes. These models will be used to design a platform for in-silico trials that will be validated with existing data and be used to provide a proof of concept of the potential efficacy of this emerging technology. The platform will be used for preliminary evaluation of the potential suitability and safety of medication, new thrombectomy device configurations and methods to select patient subpopulations for better treatment outcome. This could allow generating, exploring and refining relavant hypotheses on potential causal pathways (which may follow from the evidence obtained from clinical trials) and improving clinical trial design. Importantly, the findings of the in-silico trials will require validation under the controlled settings of randomized clinical trials.

Validation of automated Alberta Stroke Program Early CT Score (ASPECTS) software for detection of early ischemic changes on non-contrast brain CT scans

Purpose

In ASPECTS, 10 brain regions are scored visually for presence of acute ischemic stroke damage. We evaluated automated ASPECTS in comparison to expert readers.

Methods

Consecutive, baseline non-contrast CT-scans (5-mm slice thickness) from the prospective MR CLEAN trial (n = 459, MR CLEAN Netherlands Trial Registry number: NTR1804) were evaluated. A two-observer consensus for ASPECTS regions (normal/abnormal) was used as reference standard for training and testing (0.2/0.8 division). Two other observers provided individual ASPECTS-region scores. The Automated ASPECTS software was applied. A region score specificity of ≥ 90% was used to determine the software threshold for detection of an affected region based on relative density difference between affected and contralateral region. Sensitivity, specificity, and receiver-operating characteristic curves were calculated. Additionally, we assessed intraclass correlation coefficients (ICCs) for automated ASPECTS and observers in comparison to the reference standard in the test set.

Results

In the training set (n = 104), with software thresholds for a specificity of ≥ 90%, we found a sensitivity of 33–49% and an area under the curve (AUC) of 0.741–0.785 for detection of an affected ASPECTS region. In the test set (n = 355), the results for the found software thresholds were 89–89% (specificity), 41–57% (sensitivity), and 0.750–0.795 (AUC). Comparison of automated ASPECTS with the reference standard resulted in an ICC of 0.526. Comparison of observers with the reference standard resulted in an ICC of 0.383–0.464.

Conclusion

The performance of automated ASPECTS is comparable to expert readers and could support readers in the detection of early ischemic changes.

Electronic supplementary material

The online version of this article (10.1007/s00234-020-02533-6) contains supplementary material, which is available to authorized users.

Data-efficient deep learning of radiological image data for outcome prediction after endovascular treatment of patients with acute ischemic stroke

Treatment selection is becoming increasingly more important in acute ischemic stroke patient care. Clinical variables and radiological image biomarkers (old age, pre-stroke mRS, NIHSS, occlusion location, ASPECTS, among others) have an important role in treatment selection and prognosis. Radiological biomarkers require expert annotation and are subject to inter-observer variability. Recently, Deep Learning has been introduced to reproduce these radiological image biomarkers. Instead of reproducing these biomarkers, in this work, we investigated Deep Learning techniques for building models to directly predict good reperfusion after endovascular treatment (EVT) and good functional outcome using CT angiography images. These models do not require image annotation and are fast to compute. We compare the Deep Learning models to Machine Learning models using traditional radiological image biomarkers. We explored Residual Neural Network (ResNet) architectures, adapted them with Structured Receptive Fields (RFNN) and auto-encoders (AE) for network weight initialization. We further included model visualization techniques to provide insight into the network's decision-making process. We applied the methods on the MR CLEAN Registry dataset with 1301 patients. The Deep Learning models outperformed the models using traditional radiological image biomarkers in three out of four cross-validation folds for functional outcome (average AUC of 0.71) and for all folds for reperfusion (average AUC of 0.65). Model visualization showed that the arteries were relevant features for functional outcome prediction. The best results were obtained for the ResNet models with RFNN. Auto-encoder initialization often improved the results. We concluded that, in our dataset, automated image analysis with Deep Learning methods outperforms radiological image biomarkers for stroke outcome prediction and has the potential to improve treatment selection.

Changes in the gray and white matter of patients with ischemic-edematous insults after traumatic brain injury

OBJECTIVE

Gray matter (GM) and white matter (WM) are vulnerable to ischemic-edematous insults after traumatic brain injury (TBI). The extent of secondary insult after brain injury is quantifiable using quantitative CT analysis. One conventional quantitative CT measure, the gray-white matter ratio (GWR), and a more recently proposed densitometric analysis are used to assess the extent of these insults. However, the prognostic capacity of the GWR in patients with TBI has not yet been validated. This study aims to test the prognostic value of the GWR and evaluate the alternative parameters derived from the densitometric analysis acquired during the acute phase of TBI. In addition, the prognostic ability of the conventional TBI prognostic models (i.e., IMPACT [International Mission for Prognosis and Analysis of Clinical Trials in TBI] and CRASH [Corticosteroid Randomisation After Significant Head Injury] models) were compared to that of the quantitative CT measures.

METHODS

Three hundred patients with TBI of varying ages (92 pediatric, 94 adult, and 114 geriatric patients) and admitted between 2008 and 2013 were included in this retrospective cohort study. The normality of the density of the deep GM and whole WM was evaluated as the proportion of CT pixels with Hounsfield unit values of 31-35 for GM and 26-30 for WM on CT images of the entire supratentorial brain. The outcome was evaluated using the Glasgow Outcome Scale (GOS) at discharge (GOS score ≤ 3, n = 100).

RESULTS

Lower proportions of normal densities in the deep GM and whole WM indicated worse outcomes. The proportion of normal WM exhibited a significant prognostic capacity (area under the curve [AUC] = 0.844). The association between the outcome and the normality of the WM density was significant in adult (AUC = 0.792), pediatric (AUC = 0.814), and geriatric (AUC = 0.885) patients. In pediatric patients, the normality of the overall density and the density of the GM were indicative of the outcome (AUC = 0.751). The average GWR was not associated with the outcome (AUC = 0.511). IMPACT and CRASH models showed adequate and reliable performance in the pediatric and geriatric groups but not in the adult group. The highest overall predictive performance was achieved by the densitometry-augmented IMPACT model (AUC = 0.881).

CONCLUSIONS

Both deep GM and WM are susceptible to ischemic-edematous insults during the early phase of TBI. The extent of the secondary injury was better evaluated by analyzing the normality of the deep GM and WM rather than by calculating the GWR.

Abilities of a Densitometric Analysis of Computed Tomography Images and Hemorrhagic Parameters to Predict Outcome Favorability in Patients With Intracerebral Hemorrhage

BACKGROUND

Intracerebral hemorrhage (ICH) is one of the most devastating subtypes of stroke. A rapid assessment of ICH severity involves the use of computed tomography (CT) and derivation of the hemorrhage volume, which is often estimated using the ABC/2 method. However, these estimates are highly inaccurate and may not be feasible for anticipating outcome favorability.

OBJECTIVE

To predict patient outcomes via a quantitative, densitometric analysis of CT images, and to compare the predictive power of these densitometric parameters with the conventional ABC/2 volumetric parameter and segmented hemorrhage volumes.

METHODS

Noncontrast CT images of 87 adult patients with ICH (favorable outcomes = 69, unfavorable outcomes = 12, and deceased = 6) were analyzed. In-house software was used to calculate the segmented hemorrhage volumes, ABC/2 and densitometric parameters, including the skewness and kurtosis of the density distribution, interquartile ranges, and proportions of specific pixels in sets of CT images. Nonparametric statistical analyses were conducted.

RESULTS

The densitometric parameter interquartile range exhibited greatest accuracy (82.7%) in predicting favorable outcomes. The combination of skewness and the interquartile range effectively predicted mortality (accuracy = 83.3%). The actual volume of the ICH exhibited good coherence with ABC/2 (R = 0.79). Both parameters predicted mortality with moderate accuracy (<78%) but were less effective in predicting unfavorable outcomes.

CONCLUSION

Hemorrhage volume was rapidly estimated and effectively predicted mortality in patients with ICH; however, this value may not be useful for predicting favorable outcomes. The densitometric analysis exhibited significantly higher power in predicting mortality and favorable outcomes in patients with ICH.

A Relative Noncontrast CT Map to Detect Early Ischemic Changes in Acute Stroke

BACKGROUND AND PURPOSE

Early ischemic changes on noncontrast computed tomography (NCCT) are often subtle. We developed a novel postprocessing technique that aids in detecting such changes.

METHODS

NCCT maps were generated that display the relative density difference between corresponding voxels in contralateral hemispheres (ratio maps of the NCCT [rNCCT]). Voxels with a relative density difference below .95 were designated as infarct. We pilot tested the rNCCT for infarct segmentation on 6 consecutive subjects enrolled in the CT Perfusion to predict Response in Ischemic Stroke Project (CRISP) study and applied the inclusion criteria of an adequate quality NCCT and successful endovascular reperfusion. rNCCT infarct segmentation was compared to baseline NCCT, baseline CTP, and day-5 follow-up fluid-attenuated inversion recovery (FLAIR).

RESULTS

Five of the six selected cases met the inclusion criteria. Their median time from symptom onset to CT was 4.95 hours (standard deviation [SD], ±3.5; range, 1.05-10.45), and median NIHSS was 13. Early ischemic changes were identified on the rNCCT in all five cases and on the standard NCCT in three of the five cases. Lesions outlined by the rNCCT maps trended toward a better estimation of the day-5 FLAIR volume (median difference = 6.2 mL) than the ischemic core volumes assessed on baseline CTP (median difference = 51.7 mL) in the four cases with a day-5 FLAIR (P = .1).

CONCLUSION

In this proof-of-concept study, the rNCCT appears promising for detecting and quantifying early ischemic changes. These findings should be confirmed in a larger cohort.

Automated ASPECTS on Noncontrast CT Scans in Patients with Acute Ischemic Stroke Using Machine Learning

BACKGROUND AND PURPOSE

Alberta Stroke Program Early CT Score (ASPECTS) was devised as a systematic method to assess the extent of early ischemic change on noncontrast CT (NCCT) in patients with acute ischemic stroke (AIS). Our aim was to automate ASPECTS to objectively score NCCT of AIS patients.

MATERIALS AND METHODS

We collected NCCT images with a 5-mm thickness of 257 patients with acute ischemic stroke (<8 hours from onset to scans) followed by a diffusion-weighted imaging acquisition within 1 hour. Expert ASPECTS readings on DWI were used as ground truth. Texture features were extracted from each ASPECTS region of the 157 training patient images to train a random forest classifier. The unseen 100 testing patient images were used to evaluate the performance of the trained classifier. Statistical analyses on the total ASPECTS and region-level ASPECTS were conducted.

RESULTS

For the total ASPECTS of the unseen 100 patients, the intraclass correlation coefficient between the automated ASPECTS method and DWI ASPECTS scores of expert readings was 0.76 (95% confidence interval, 0.67-0.83) and the mean ASPECTS difference in the Bland-Altman plot was 0.3 (limits of agreement, -3.3, 2.6). Individual ASPECTS region-level analysis showed that our method yielded κ = 0.60, sensitivity of 66.2%, specificity of 91.8%, and area under curve of 0.79 for 100 × 10 ASPECTS regions. Additionally, when ASPECTS was dichotomized (>4 and ≤4), κ = 0.78, sensitivity of 97.8%, specificity of 80%, and area under the curve of 0.89 were generated between the proposed method and expert readings on DWI.

CONCLUSIONS

The proposed automated ASPECTS scoring approach shows reasonable ability to determine ASPECTS on NCCT images in patients presenting with acute ischemic stroke.