Fast nonlinear regression method for CT brain perfusion analysis

Standardizing the estimation of ischemic regions can harmonize CT perfusion stroke imaging

OBJECTIVES

We aimed to evaluate the real-world variation in CT perfusion (CTP) imaging protocols among stroke centers and to explore the potential for standardizing vendor software to harmonize CTP images.

METHODS

Stroke centers participating in a nationwide multicenter healthcare evaluation were requested to share their CTP scan and processing protocol. The impact of these protocols on CTP imaging was assessed by analyzing data from an anthropomorphic phantom with center-specific vendor software with default settings from one of three vendors (A-C): IntelliSpace Portal, syngoVIA, and Vitrea. Additionally, standardized infarct maps were obtained using a logistic model.

RESULTS

Eighteen scan protocols were studied, all varying in acquisition settings. Of these protocols, seven, eight, and three were analyzed with center-specific vendor software A, B, and C respectively. The perfusion maps were visually dissimilar between the vendor software but were relatively unaffected by the acquisition settings. The median error [interquartile range] of the infarct core volumes (mL) estimated by the vendor software was - 2.5 [6.5] (A)/ - 18.2 [1.2] (B)/ - 8.0 [1.4] (C) when compared to the ground truth of the phantom (where a positive error indicates overestimation). Taken together, the median error [interquartile range] of the infarct core volumes (mL) was - 8.2 [14.6] before standardization and - 3.1 [2.5] after standardization.

CONCLUSIONS

CTP imaging protocols varied substantially across different stroke centers, with the perfusion software being the primary source of differences in CTP images. Standardizing the estimation of ischemic regions harmonized these CTP images to a degree.

CLINICAL RELEVANCE STATEMENT

The center that a stroke patient is admitted to can influence the patient's diagnosis extensively. Standardizing vendor software for CT perfusion imaging can improve the consistency and accuracy of results, enabling a more reliable diagnosis and treatment decision.

KEY POINTS

• CT perfusion imaging is widely used for stroke evaluation, but variation in the acquisition and processing protocols between centers could cause varying patient diagnoses. • Variation in CT perfusion imaging mainly arises from differences in vendor software rather than acquisition settings, but these differences can be reconciled by standardizing the estimation of ischemic regions. • Standardizing the estimation of ischemic regions can improve CT perfusion imaging for stroke evaluation by facilitating reliable evaluations independent of the admission center.

Multiphase CT angiography perfusion maps for predicting target mismatch and ischemic lesion volumes

The complexity of CT perfusion (CTP) acquisition protocols may limit the availability of target mismatch assessment at resource-limited hospitals. We compared CTP mismatch with a mismatch surrogate generated from a simplified dynamic imaging sequence comprising widely available non-contrast CT (NCCT) and multiphase CT angiography (mCTA). Consecutive patients with anterior circulation acute ischemic stroke who received NCCT, mCTA, and CTP were retrospectively included in this study. An mCTA-perfusion (mCTA-P) dynamic series was formed by co-registering NCCT and mCTA. We simulated an ideal mCTA-P study by down-sampling CTP (dCTP) dynamic images according to mCTA timing. Ischemic core and penumbra volumes were estimated by cerebral blood flow and Tmax thresholding, respectively, on perfusion maps calculated independently for CTP, dCTP, and mCTA-P by deconvolution. Concordance in target mismatch (core < 70 ml, penumbra ≥ 15 ml, mismatch ratio ≥ 1.8) determination by dCTP and mCTA-P versus CTP was assessed. Of sixty-one included patients, forty-six had a CTP target mismatch. Concordance with CTP profiles was 90% and 82% for dCTP and mCTA-P, respectively. Lower mCTA-P concordance was likely from differences in collimation width between NCCT and mCTA, which worsened perfusion map quality due to a CT number shift at mCTA. Moderate diagnostic agreement between CTP and mCTA-P was found and may improve with optimal mCTA scan parameter selection as simulated by dCTP. mCTA-P may be a pragmatic alternative where CTP is unavailable or the risks of additional radiation dose, contrast injections, and treatment delays outweigh the potential benefit of a separate CTP scan.

CT perfusion stroke lesion threshold calibration between deconvolution algorithms

CTP is an important diagnostic tool in managing patients with acute ischemic stroke, but challenges persist in the agreement of stroke lesion volumes and ischemic core-penumbra mismatch profiles determined with different CTP post-processing software. We investigated a systematic method of calibrating CTP stroke lesion thresholds between deconvolution algorithms using a digital perfusion phantom to improve inter-software agreement of mismatch profiles. Deconvolution-estimated cerebral blood flow (CBF) and Tmax was compared to the phantom ground truth via linear regression for one model-independent and two model-based deconvolution algorithms. Using the clinical standard of model-independent CBF < 30% and Tmax > 6 s as reference thresholds for ischemic core and penumbra, respectively, we determined that model-based CBF < 15% and Tmax > 6 s were the corresponding calibrated thresholds after accounting for quantitative differences revealed at linear regression. Calibrated thresholds were then validated in 63 patients with large vessel stroke by evaluating agreement (concordance and Cohen's kappa, κ) between the two model-based and model-independent deconvolution methods in determining mismatch profiles used for clinical decision-making. Both model-based deconvolution methods achieved 95% concordance with model-independent assessment and Cohen's kappa was excellent (κ = 0.87; 95% confidence interval [CI] 0.72-1.00 and κ = 0.86; 95% CI 0.70-1.00). Our systematic method of calibrating CTP stroke lesion thresholds may help harmonize mismatch profiles determined by different software.

Spatio-temporal physics-informed learning: A novel approach to CT perfusion analysis in acute ischemic stroke

CT perfusion imaging is important in the imaging workup of acute ischemic stroke for evaluating affected cerebral tissue. CT perfusion analysis software produces cerebral perfusion maps from commonly noisy spatio-temporal CT perfusion data. High levels of noise can influence the results of CT perfusion analysis, necessitating software tuning. This work proposes a novel approach for CT perfusion analysis that uses physics-informed learning, an optimization framework that is robust to noise. In particular, we propose SPPINN: Spatio-temporal Perfusion Physics-Informed Neural Network and research spatio-temporal physics-informed learning. SPPINN learns implicit neural representations of contrast attenuation in CT perfusion scans using the spatio-temporal coordinates of the data and employs these representations to estimate a continuous representation of the cerebral perfusion parameters. We validate the approach on simulated data to quantify perfusion parameter estimation performance. Furthermore, we apply the method to in-house patient data and the public Ischemic Stroke Lesion Segmentation 2018 benchmark data to assess the correspondence between the perfusion maps and reference standard infarct core segmentations. Our method achieves accurate perfusion parameter estimates even with high noise levels and differentiates healthy tissue from infarcted tissue. Moreover, SPPINN perfusion maps accurately correspond with reference standard infarct core segmentations. Hence, we show that using spatio-temporal physics-informed learning for cerebral perfusion estimation is accurate, even in noisy CT perfusion data. The code for this work is available at https://github.com/lucasdevries/SPPINN.

Spatial CT perfusion data helpful in automatically locating vessel occlusions for acute ischemic stroke patients

Introduction

Locating a vessel occlusion is important for clinical decision support in stroke healthcare. The advent of endovascular thrombectomy beyond proximal large vessel occlusions spurs alternative approaches to locate vessel occlusions. We explore whether CT perfusion (CTP) data can help to automatically locate vessel occlusions.

Methods

We composed an atlas with the downstream regions of particular vessel segments. Occlusion of these segments should result in the hypoperfusion of the corresponding downstream region. We differentiated between seven-vessel occlusion locations (ICA, proximal M1, distal M1, M2, M3, ACA, and posterior circulation). We included 596 patients from the DUtch acute STroke (DUST) multicenter study. Each patient CTP data set was processed with perfusion software to determine the hypoperfused region. The downstream region with the highest overlap with the hypoperfused region was considered to indicate the vessel occlusion location. We assessed the indications from CTP against expert annotations from CTA.

Results

Our atlas-based model had a mean accuracy of 86% and could achieve substantial agreement with the annotations from CTA according to Cohen's kappa coefficient (up to 0.68). In particular, anterior large vessel occlusions and occlusions in the posterior circulation could be located with an accuracy of 80 and 92%, respectively.

Conclusion

The spatial layout of the hypoperfused region can help to automatically indicate the vessel occlusion location for acute ischemic stroke patients. However, variations in vessel architecture between patients seemed to limit the capacity of CTP data to distinguish between vessel occlusion locations more accurately.

Neural network-derived perfusion maps: A model-free approach to computed tomography perfusion in patients with acute ischemic stroke

Objective

In this study, we investigate whether a Convolutional Neural Network (CNN) can generate informative parametric maps from the pre-processed CT perfusion data in patients with acute ischemic stroke in a clinical setting.

Methods

The CNN training was performed on a subset of 100 pre-processed perfusion CT dataset, while 15 samples were kept for testing. All the data used for the training/testing of the network and for generating ground truth (GT) maps, using a state-of-the-art deconvolution algorithm, were previously pre-processed using a pipeline for motion correction and filtering. Threefold cross validation had been used to estimate the performance of the model on unseen data, reporting Mean Squared Error (MSE). Maps accuracy had been checked through manual segmentation of infarct core and total hypo-perfused regions on both CNN-derived and GT maps. Concordance among segmented lesions was assessed using the Dice Similarity Coefficient (DSC). Correlation and agreement among different perfusion analysis methods were evaluated using mean absolute volume differences, Pearson correlation coefficients, Bland-Altman analysis, and coefficient of repeatability across lesion volumes.

Results

The MSE was very low for two out of three maps, and low in the remaining map, showing good generalizability. Mean Dice scores from two different raters and the GT maps ranged from 0.80 to 0.87. Inter-rater concordance was high, and a strong correlation was found between lesion volumes of CNN maps and GT maps (0.99, 0.98, respectively).

Conclusion

The agreement between our CNN-based perfusion maps and the state-of-the-art deconvolution-algorithm perfusion analysis maps, highlights the potential of machine learning methods applied to perfusion analysis. CNN approaches can reduce the volume of data required by deconvolution algorithms to estimate the ischemic core, and thus might allow the development of novel perfusion protocols with lower radiation dose deployed to the patient.

Probability maps classify ischemic stroke regions more accurately than CT perfusion summary maps

OBJECTIVES

To compare single parameter thresholding with multivariable probabilistic classification of ischemic stroke regions in the analysis of computed tomography perfusion (CTP) parameter maps.

METHODS

Patients were included from two multicenter trials and were divided into two groups based on their modified arterial occlusive lesion grade. CTP parameter maps were generated with three methods-a commercial method (ISP), block-circulant singular value decomposition (bSVD), and non-linear regression (NLR). Follow-up non-contrast CT defined the follow-up infarct region. Conventional thresholds for individual parameter maps were established with a receiver operating characteristic curve analysis. Probabilistic classification was carried out with a logistic regression model combining the available CTP parameters into a single probability.

RESULTS

A total of 225 CTP data sets were included, divided into a group of 166 patients with successful recanalization and 59 with persistent occlusion. The precision and recall of the CTP parameters were lower individually than when combined into a probability. The median difference [interquartile range] in mL between the estimated and follow-up infarct volume was 29/23/23 [52/50/52] (ISP/bSVD/NLR) for conventional thresholding and was 4/6/11 [31/25/30] (ISP/bSVD/NLR) for the probabilistic classification.

CONCLUSIONS

Multivariable probability maps outperform thresholded CTP parameter maps in estimating the infarct lesion as observed on follow-up non-contrast CT. A multivariable probabilistic approach may harmonize the classification of ischemic stroke regions.

KEY POINTS

• Combining CTP parameters with a logistic regression model increases the precision and recall in estimating ischemic stroke regions. • Volumes following from a probabilistic analysis predict follow-up infarct volumes better than volumes following from a threshold-based analysis. • A multivariable probabilistic approach may harmonize the classification of ischemic stroke regions.

Early detection of small volume stroke and thromboembolic sources with computed tomography: Rationale and design of the ENCLOSE study

Background

Computed tomography is the most frequently used imaging modality in acute stroke imaging protocols. Detection of small volume infarcts in the brain and cardioembolic sources of stroke is difficult with current computed tomography protocols. Furthermore, the role of computed tomography findings to predict recurrent ischemic stroke is unclear. With ENCLOSE, we aim to improve (1) the detection of small volume infarcts with thin slice computed tomography perfusion (CTP) images and thromboembolic source with cardiac computed tomography techniques in the acute stage of ischemic stroke and (2) prediction of recurrent ischemic stroke with computed tomography-derived predictors.

Methods/design: ENCLOSE is a prospective multicenter observational cohort study, which will be conducted in three Dutch stroke centers (ClinicalTrials.gov Identifier: NCT04019483). Patients (≥18 years) with suspected acute ischemic stroke who undergo computed tomography imaging within 9 h after symptom onset are eligible. Computed tomography imaging includes non-contrast CT, CTP, and computed tomography angiography (CTA) from base of the heart to the top of the brain. Dual-energy CT data will be acquired when possible, and thin-slice CTP reconstructions will be obtained in addition to standard 5 mm CTP data. CTP data will be processed with commercially available software and locally developed model-based methods. The post-processed thin-slice CTP images will be compared to the standard CTP images and to magnetic resonance diffusion-weighted imaging performed within 48 h after admission. Detection of cardioembolic sources of stroke will be evaluated on the CTA images. Recurrence will be evaluated 90 days and two years after the index event. The added value of imaging findings to prognostic models for recurrent ischemic stroke will be evaluated.

Conclusion

The aim of ENCLOSE is to improve early detection of small volume stroke and thromboembolic sources and to improve prediction of recurrence in patients with acute ischemic stroke.

Improving the Quality of Cerebral Perfusion Maps With Monoenergetic Dual-Energy Computed Tomography Reconstructions

OBJECTIVE

We compared 40- to 70-keV virtual monoenergetic to conventional computed tomography (CT) perfusion reconstructions with respect to quality of perfusion maps.

METHODS

Conventional CT perfusion (CTP) images were acquired at 80 kVp in 25 patients, and 40- to 70-keV images were acquired with a dual-layer CT at 120 kVp in 25 patients. First, time-attenuation-curve contrast-to-noise ratio was assessed. Second, the perfusion maps of both groups were qualitatively analyzed by observers. Last, the monoenergetic reconstruction with the highest quality was compared with the clinical standard 80-kVp CTP acquisitions.

RESULTS

Contrast-to-noise ratio was significantly better for 40 to 60 keV as compared with 70 keV and conventional images (P < 0.001). Visually, the difference between the blood volume maps among reconstructions was minimal. The 50-keV perfusion maps had the highest quality compared with the other monoenergetic and conventional maps (P < 0.002).

CONCLUSIONS

The quality of 50-keV CTP images is superior to the quality of conventional 80- and 120-kVp images.

Effect of prolonged acquisition intervals for CT-perfusion analysis methods in patients with ischemic stroke

Introduction

The limited axial coverage of many computed tomography (CT) scanners poses a high risk on false negative findings in cerebral CT‐perfusion (CTP) imaging. Axial coverage may be increased by moving the table back and forth during image acquisition. However, this method often increases the acquisition interval between CT frames, which may influence the CTP analysis. In this study, we evaluated the influence of different acquisition intervals on quantitative perfusion maps and infarct volumes by analyzing patient data with three CTP analysis methods.

Methods

CT‐perfusion data from 25 patients with ischemic stroke were used for this study. The acquisition interval was synthetically reduced from 1 to 5 s before calculating perfusion values, which included cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT). The color scaling of the perfusion was scaled such that the mean perfusion value had the same color‐coding as the mean perfusion in the 1 s reference. Also, infarct core and penumbra volumes (summary map) were calculated using default thresholds of CBV and relative MTT (rMTT). The original, 1 s acquisition interval scan served as the reference standard. A commercial block‐circulant singular value decomposition (bSVD) based method (ISP; Philips Healthcare), a non‐commercial bSVD method, and a non‐linear regression (NLR) model‐based method were evaluated.

Results

Cerebral blood volume values generated with bSVD and NLR were not significantly different from the reference standard, while ISP showed significant differences for acquisition intervals of 3 and 4 s. MTT and CBF values generated with bSVD and ISP were significantly different for all acquisition intervals, whereas NLR did not show any significant differences. Calibrated perfusion maps were able to distinguish healthy from infarcted tissue up to an acquisition interval of 5 s for all methods. The infarct core volumes were significantly different for acquisition intervals of 2 (NLR) and 3 s (bSVD and ISP) or greater. For the penumbra volumes, NLR showed no significant differences, while bSVD and ISP showed significant differences for the 5 s interval and for all intervals, respectively. Visual inspection of the summary maps indicated minor differences between the reference standard and acquisition intervals of 4 s or less (ISP) and 5 s or less (bSVD and NLR).

Conclusion

Altering the acquisition interval may introduce a bias in the perfusion parameters. Calibration of the visualization of the perfusion maps with increasing acquisition intervals allowed distinction between healthy and infarcted tissue. Infarct volumes based on relative MTT can be influenced by the acquisition interval, but visual inspection of the summary maps indicated minor differences between the reference standard and acquisition intervals up to 4 (ISP) and 5 s (bSVD and NLR). Taken together, axial coverage can be increased by prolonging the acquisition interval up to 5 s depending on the perfusion analysis.