Structural integrity of white matter tracts as a predictor of acute ischemic stroke outcome

Abstract

Background and purpose

Clinical assessment scores in acute ischemic stroke are only moderately correlated with lesion volume since lesion location is an important confounding factor. Many studies have investigated gray matter indicators of stroke severity, but the understanding of white matter tract involvement is limited in the early phase after stroke. This study aimed to measure and model the involvement of white matter tracts with respect to 24-h post-stroke National Institutes of Health Stroke Scale (NIHSS).

Material and methods

A total of 96 patients (50 females, mean age 66.4 ± 14.0 years, median NIHSS 5, interquartile range: 2–9.5) with follow-up fluid-attenuated inversion recovery magnetic resonance imaging data sets acquired one to seven days after acute ischemic stroke onset due to proximal anterior circulation occlusion were included. Lesions were semi-automatically segmented and non-linearly registered to a common reference atlas. The lesion overlap and tract integrity were determined for each white matter tract in the AALCAT atlas and used to model NIHSS outcomes using a supervised linear-kernel support vector regression method, which was evaluated using leave-one-patient-out cross validation.

Results

The support vector regression model using the tract integrity and tract lesion overlap measurements predicted the 24-h NIHSS score with a high correlation value of r = 0.7. Using the tract overlap and tract integrity feature improved the modeling accuracy of NIHSS significantly by 6% (p < 0.05) compared to using overlap measures only.

Conclusion

White matter tract integrity and lesion load are important predictors for clinical outcome after an acute ischemic stroke as measured by the NIHSS and should be integrated for predictive modeling.

Keywords

Acute ischemic stroke computer-assisted image analysis distance mapping transform white matter support vector machines

Introduction

Lesion location is known to play a crucial role for stroke severity and recovery. It has been shown that a combination of lesion location and volume predicts stroke severity better compared to lesion volume alone.¹ However, data investigating the impact of white matter (WM) involvement on clinical outcome is sparse. Given that WM constitutes approximately half of the brain volume and recovery from WM injury has a different trajectory compared to gray matter (GM),² the importance of WM tract involvement in ischemic stroke should be investigated in more detail. Furthermore, an improved understanding of WM injury will lay the foundation to build comprehensive models of long-term outcomes that integrate the structural integrity of GM and WM structures, both.

Previous studies investigating WM tract injury post-stroke predominantly used lesion overlap measurements.^3–8 However, a WM tract may be affected by a large percentage of lesion overlap but still retain signal conduction. In contrast, a tract may have a small lesion that does not spare any fibers within it, thus incapacitating the tract (Figure 1). Analyzing the tract integrity in addition to simple overlap measurements might hold additional informative value for the prediction of the clinical outcome.

The aim of this work was to develop a computer-assisted decision support framework to quantify tract-specific structural integrity measures and estimate the utility of the proposed metric in combination with traditional lesion overlap measures to predict early stroke outcomes.

Materials and methods

Data

The data sets used in this project were pooled together from two multi-center trials: the ESCAPE⁹ and the INTERRSeCT study,¹⁰ which include patients with acute anterior circulation stroke who underwent follow-up magnetic resonance imaging (MRI) and computed tomography assessment. The ESCAPE database consists of 315 subjects with acute ischemic occlusion(s) of the middle cerebral artery (M1-, M2-, or carotid-middle cerebral artery) imaged within 12-h of last-seen-normal and a baseline National Institutes of Health Stroke Scale (NIHSS) > 5. The INTERRSeCT database consists of 100 patients, also imaged within 12-h from symptom onset.

In this study, follow-up fluid-attenuated inversion recovery (FLAIR) MRI was used as the basis for lesion delineation. Therefore, only subjects who obtained follow-up FLAIR imaging within one to seven days after acute ischemic stroke and had a measured NIHSS score at approximately 24-h after onset were included. Subjects with previous cerebral strokes, bilateral lesions, remote hemorrhagic transformations, or severe WM hyperintensities were excluded.

The final sample size consisted of 96 patients, 39 subjects from ESCAPE and 57 from INTERRSeCT. Table 1 shows the summary of clinical data of the included patients. The majority of subjects received follow-up imaging within 1 to 2 days (n = 79), 11 subjects within 2.5 to 3 days, and 5 within 3.5 to 5.5 days. Of the 79 subjects imaged within 48-h, 67 of them were followed-up within 1 to 1.5 days (<36-h).

Table 1.

Clinical characteristics of the sample population (n = 96)

Characteristics	Left hemisphere	Right hemisphere
Frequency, n	56	40
Sex, females	31	19
Mean age, years (95% CI)	66.5 (63–70)	66 (62–70)
Median NIHSS, 24 h (IQR)	5 (9)	5 (6.25)
Median mRS, 90 day (IQR)	1 (3)	2 (2)
Treatment type, EVT (tPA)	12 (44)	12 (28)
Mean follow-up MRI scan, days (95% CI)	1.8 (1.6–2)	1.6 (1.3–1.9)

CI: confidence interval; MRI: magnetic resonance imaging; EVT: endovascular therapy; IQR: interquartile range; NIHSS: National Institutes of Health Stroke Scale.

The data that support the findings of this study can be requested from the principal investigators of the two multi-center trials. Acquisition of the data sets for the two trials was approved by the respective local ethics board at each site contributing to the two trials. All data sets used in this secondary study were made available after complete anonymization.

Preprocessing

Lesions were semi-automatically segmented in each FLAIR MRI data set using ANTONIA.^11,12 Skull stripping of the FLAIR MRI data sets was performed using the ANTs toolkit.¹³ The individual brain volumes were globally aligned to the elderly brain atlas developed by the groupe d’imagerie neurofonctionnelle (GIN)¹⁴ using an affine registration. To ensure local correspondence, this affine transform was used to initialize a secondary non-linear registration employing the Lanczos-windowed-sinc interpolation and optimization of the mutual information similarity metric with cost function masking.¹⁵ The resulting non-linear transformations were applied to the binary lesion masks using a nearest-neighbour interpolation, thereby transforming them from patient space to atlas space. No mirroring of lesions from the left to the right hemisphere (or vice versa) was performed. All registrations were visually assessed.

Feature extraction

The AALCAT WM atlas provided by the NiiStat toolbox (https://www.nitrc.org/projects/niistat/) was employed to quantify the WM tract lesion overlap and integrity as described in the following. It consists of labels for 17 WM tracts in the left and right hemisphere separately, resulting in 34 unique labels.

Tract overlap

The volume of the overlap of a lesion with a tract, normalized by the total tract volume is calculated (in atlas space) for each WM tract. This relative overlap is indicative of the extent of the insult to the WM tract.

Tract integrity

In the following, tract integrity refers only to the structural integrity of a WM tract. For calculation of this metric, a binary mask is first created for each tract and a corresponding 3D centerline representation is calculated. For each voxel of the 3D centerline, Danielsson’s distance mapping¹⁶ is used to compute the initial tract thickness (T_wm). To determine the structural tract integrity, a second binary mask is created for each lesioned tract by subtraction of the lesion from the binarized unaffected tract. Again, the tract thickness estimation is repeated for the lesioned tract (T_lesion). For each centerline voxel, the percent difference in width comparing the tract with and without lesion is computed, and the maximum value is used as the final tract integrity measure (T_spared = argmax [T_wm – T_lesion]/T_wm). Figure 1(c) juxtaposes the tract overlap and integrity metric pictorially using a sample patient data.

Predictive model

To investigate the predictive power of the tract overlap and integrity features, a linear-kernel epsilon-support vector regression (SVR) machine was trained to predict the 24-h NIHSS and evaluated using a leave-one-patient-out cross validation scheme. The algorithm aims to identify a linear function in the high-dimensional feature space such that regression errors within the tolerance of ɛ (i.e., the support vectors) are not penalized. Default parameter values of C = 1, which controls the trade-off between misclassification and error-minimization, and ɛ = 0.001 were used.

Within the cross-validation of each SVR, the ReliefF algorithm is used to rank the input features in order of their efficacy in differentiating their closest neighbors in the high-dimensional feature space.¹⁷ This iterative ReliefF-SVR algorithm results in a reduced subset of WM tract features for predicting the 24-h NIHSS outcome, which is beneficial as redundant and non-informative features often downgrade the prediction accuracy.

Two sets of input features are used for training and evaluation of the SVRs. The first set includes patient age, sex, and tract overlap measures of the WM tracts. The second set includes patient age, sex, tract overlap, and tract integrity measures. These regression models are individually evaluated using their mean absolute error (MAE), the root mean square error (RMSE), and the Spearman’s rho (ρ) correlation coefficient. The utility of using the tract integrity measure in addition to the simple tract overlap metric for predicting early stroke outcomes is assessed by statistically comparing the residuals of the regression models using the Friedman test.¹⁸

In order to investigate the features selected by the SVR in more detail, non-parametric correlation coefficients were determined for the WM tract overlap and integrity features that were selected at least once in the leave-one-patient-out cross correlations for prediction of the 24-h NIHSS (Spearman’s rho; 2-tailed). A p-value < 0.05 was considered significant.

Results

The simple SVR model using the overlap measurements as well as the patient age and sex predicted the 24-h NIHSS score with a MAE of 3.27 and a RMSE of 4.35, corresponding to a correlation of r = 0.64. Compared to this, the SVR model using the tract integrity measurements in addition to the tract overlap measurements and patient sex and age predicted the 24-h NIHSS score with a MAE of 2.92 and a RMSE of 3.95, corresponding to a correlation of r = 0.7 (Figure 2). The non-parametric Friedman chi-squared analysis of variance of the two residual distributions (Figure 3) revealed that the SVR model using the tract integrity measurements in addition to the lesion overlap measurements performs significantly better (p = 0.004).

Table 2 shows the WM tract overlap and integrity features, which were selected by the SVR model using the tract overlap as well as integrity measurements. The correlation analysis revealed that most of the selected features are significantly correlated with the 24-h NIHSS outcome variable (p < 0.05) in at least one hemisphere. Exceptions to this include the arcuate, inferior occipital fasciculus, internal capsule, posterior segment, cerebellar pedunculus, and the uncinate. Overall, it becomes apparent that the overlap measurements are more highly correlated with the 24-h NIHSS outcome compared to the corresponding integrity measurements in the majority of cases. However, the correlation values of the multi-parametric SVR models outperform the single parameter correlation values in all cases highlighting the benefit of the proposed multi-parametric machine learning approach.

Table 2.

Spearman correlation coefficients of tract overlap and integrity features used in the predictive model with 24-h NIHSS

Tract name	Overlap vs. NIHSS	Integrity vs. NIHSS
Left Anterior Commissure	0.36*	0.27*
Right Anterior Commissure	0.02	0.02
Left Arcuate	0.21*	0.19
Right Arcuate	0.09	0.1
Left Cingulum	0.36*	0.31*
Right Cingulum	0.36*	0.31*
Left Corpus Callosum	0.31*	0.27*
Right Corpus Callosum	0.17	0.14
Left Cortico Spinal	0.3*	0.24*
Right Cortico Spinal	0.07	0.03
Left Fornix	0.36*	0.29*
Right Fornix	0.15	0.11
Left Inferior Longitudinal Fasciculus	0.28*	0.34*
Right Inferior Longitudinal Fasciculus	0.04	−0.07
Left Inferior Occipito Frontal Fasciculus	0.29*	0.13
Right Inferior Occipito Frontal Fasciculus	0.07	0.12
Left Internal Capsule	0.33*	0.17
Right Internal Capsule	0.09	0.07
Left Optic Radiations	0.32*	0.25*
Left Posterior Segment	0.21*	−0.15
Right Posterior Segment	0.07	0.03
Right Superior Cerebellar Pedunculus	0.03	0.03
Left Uncinate	0.32*	0.23*
Right Uncinate	0.04	−0.08

Significant correlation (p < 0.05).

Discussion

The main finding of this study is that WM lesion overlap and integrity are important determinants of 24-h NIHSS outcome in acute ischemic stroke patients. Furthermore, our analyses showed that inclusion of the structural tract integrity metric in addition to simple tract overlap metric significantly improves the prediction of the 24-h NIHSS outcome in acute ischemic stroke patients compared to using the simple and more widely used lesion overlap metric alone.

Multiple studies have investigated lesion overlap of WM tracts before to model clinical outcomes (both acute and chronic). It is widely established that the greater the damage to the corticospinal tract (CST) is, the poorer the motor outcome will be.⁸ CST lesion volume overlap has been extensively studied for assessing post-stroke motor outcome ranging from finger strength and hand function⁷ to more global measure of functional ability using the Chedoke-McMaster stroke assessment (CMSA), action research arm test (ARAT), and NIHSS stroke scale.^4–6 It was reported that structural integrity of CST correlates well with long-term NIHSS outcomes (three months), complements resting state connectivity measures, and in general is better at modeling residual functional connectivity than global lesion volume measures. However, these studies measure structural integrity of the CST by only calculating the percentage of lesion volume overlap with the CST tract. We propose that the tract integrity metric is a complementary measure of the structural integrity of WM tracts and encourage future studies that quantify structure–function relationships to include tract integrity in their analysis. The tract overlap and integrity metrics measure different aspects of a WM injury (see Figure 1). Therefore, the improvement in predictive accuracy of the machine learning model can be attributed to the complementary information the tract integrity features provide the machine learning algorithm for modeling the 24-h NIHSS. This tract integrity measure is especially relevant to post-stroke disconnection syndromes. The study of conduction pathways in the brain, i.e., hodology, revealed that brain regions can retain their independent functional specializations, while having significant impairments with respect to functional outcomes that rely on the co-ordination of two disconnected brain regions.^19,20 Neuroimaging research investigating higher order mental functions has begun to juxtapose the pathophysiology of impairment from the perspective of disconnection with quantitative findings (GM eloquence) from lesion-symptom mapping studies.²¹ Recently, an aphasia study investigated the usefulness of disconnection analysis for predicting language outcomes among stroke patients.²² The authors utilized a graph-based approach to quantify the post-stroke structural connectivity changes using diffusion tensor based tractography maps and concluded that disconnection metrics do not improve the prediction of language outcomes compared to simple lesion overlap measurements. Although their results might seem to contravene our analysis, several design considerations may explain the difference. First, the stroke cohort analyzed in this study is of a broad age range (19–99 years), with varying imaging time points (few weeks to 40 years post-stroke), and long-term language assessments (58 ± 66 months). Second, in contrast to the proposed tract integrity metric that quantifies the change in cross-sectional area for each WM tract post-stroke, the disconnection metric used in the previous study is representative of regional connectivity changes and global structural network changes. The authors sought to estimate the relationship between the changes in WM connectivity with language scores assessed after the neurological recovery period (90 days post-stroke). Although their work suggests that disconnection does not improve the long-term language outcome prediction model, the broader utility of structural tract integrity measures in modeling other stroke outcomes like the 24-h NIHSS used in this work, requires further investigation.

Another study reports that in chronic stroke patients, injury of the CST is correlated with motor impairment measured by the Fugl-Meyer assessment.³ In particular, the authors compute a normalized lesion overlap metric, using a slice-by-slice approach in which at each slice, the width of the tract was used to normalize the raw overlap values. The authors suggest that higher normalized CST lesion overlaps lead to worse motor deficits (independent of the global lesion volume). While this method compensates for the tapering of the tract, it calculates a single integrity metric for each slice of the image that is affected by the orientation of the tract, instead of advanced 3D metrics as proposed in this work.

The results of this study suggest that the lesion overlap measurements are more highly correlated with the 24-h NIHSS compared to the integrity measurements in univariate analyses. At the same time, including the tract integrity measurements in addition to the tract lesion overlap measurements improved the prediction of the 24-h NIHSS significantly. For this reason, the tract integrity measurement is not a replacement for the tract lesion overlap metric but should be considered complementary.

Furthermore, the results of this study suggest that left hemispheric tracts are generally more important for modeling the clinical outcome in acute ischemic stroke patients. This finding is generally well in line with previous findings. For example, Ernst et al.²³ investigated the eloquence with respect to motor function in each hemisphere independently with a sample of comparable lesion loads in both hemispheres and reported a left hemispheric dominance after accounting for age, sex, and lesion volume in their univariate voxel-wise analysis. Similar results are also reported in other studies that model NIHSS as the outcome variable.⁵ Aside from this, a majority of patients included in this study had left-hemispheric strokes (see Table 1), which could also contribute to the dominance of eloquent WM tracts found in the left hemisphere. Most importantly, the NIHSS outcome score assesses language, which is a lateralized function. Therefore, the predilection to left hemispheric tracts is likely also due to the nature of the NIHSS scale itself.²⁴

There are a few limitations of the current work that should be acknowledged. First, the sample considered in this study includes mild to moderate strokes only (Table 1). Therefore, the top-ranked WM tracts are affected by a sample bias. The degree of involvement of a tract might change given a different stroke population and these estimates may differ. The extent of tract integrity also depends on the initial width of the tract as marked by the atlas. Therefore, the ratio of tract integrity can vary based on the fidelity of the WM atlas used.²⁵ Furthermore, the tract integrity measure does not consider compensatory routes that may support the function of a remote brain region. The severity of the clinical deficit results mainly from the inability of the entire functional network to perform a certain task. Given the possibility of perilesional neuroplasticity (remapping signal conduction pathways around the site of injury), a disruption in a single WM tract may not disrupt the entire functional network. Lastly, the tolerance of structural tract integrity for each WM tract in the brain is unknown. In other words, there is no evidence in literature how the extent of tract dissection relates to clinical deficits and if there are any specific thresholds. Although the tract integrity measure offers a new perspective of measuring tract involvement, the clinical significance of this metric needs to be investigated in future studies. Finally, it should be highlighted that this work focused solely on WM tract involvement for predicting stroke outcome. For this reason, the overall lesion volume was not used as an additional feature for the prediction as this would have included GM involvement, which would have made it difficult to investigate the true importance of the WM involvement. Thus, future work should focus on including WM and GM lesion metrics at the same time for predicting clinical outcome.

Despite these limitations, it is envisaged that the proposed framework for quantifying WM involvement is a step in the right direction towards modeling stroke outcomes. Future research to improve the prediction accuracy of long-term clinical outcome may benefit from including tract-based and atlas-based features such as regional lesion overlaps. Previous research in this area, however, has mostly focussed on univariate statistics and GM involvement.^26,27 Given the results of this study, it appears beneficial to combine features that quantify involvement and integrity of GM and WM at the same time along with patient clinical data within a multi-parametric machine learning framework to predict clinical outcome.

Conclusion

In conclusion, the results of this study suggest that the WM tract integrity and lesion load are important predictors for clinical outcome as measured by the NIHSS. In the future, both tract-based and gray matter lesion features could be utilized to model clinical outcomes.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is supported by the Heart and Stroke Foundation of Canada Grant in aid (G-17-0018368).

ORCID iDs

Deepthi Rajashekar

Michael D Hill

References

Menezes

Wang Zhu

, et al. The real estate factor: quantifying the impact of infarct location on stroke severity. Stroke 2007; 38: 194–197.

Helenius

Tatlisumak

MR imaging of white matter changes. In: von Kummer

Back

(eds). Magnetic resonance imaging in ischemic stroke, Berlin: Springer, 2006, pp. 149–158.

Zhu

Lindenberg

Alexander

Schlaug

. Lesion load of the corticospinal tract predicts motor impairment in chronic stroke. Stroke 2010; 41: 910–915.

Lam

Binns

Honjo

, et al. Variability in stroke motor outcome is explained by structural and functional integrity of the motor system. Sci Rep 2018; 8: 9480.

Habegger

Wiest

Weder

, et al. Relating acute lesion loads to chronic outcome in ischemic stroke-an exploratory comparison of mismatch patterns and predictive modeling. Front Neurol 2018; 9: 737.

Sterr

Shen

Szameitat

Herron

. The role of corticospinal tract damage in chronic motor recovery and neurorehabilitation: a pilot study. Neurorehabil Neural Repair 2010; 24: 413–419.

Wolbrecht

Rowe

Chan

Ingemanson

Cramer

Reinkensmeyer

. Finger strength, individuation, and their interaction: relationship to hand function and corticospinal tract injury after stroke. Clin Neurophysiol 2018; 129: 797–808.

Stinear

Barber

Smale

, et al. Functional potential in chronic stroke patients depends on corticospinal tract integrity. Brain 2007; 130: 170–180.

Demchuk

Goyal

Menon

, et al. Endovascular treatment for small core and anterior circulation proximal occlusion with emphasis on minimizing CT to recanalization times (ESCAPE) trial: methodology. Int J Stroke 2015; 10: 429–438.

10.

Menon

d’Esterre

Qazi

, et al. Multiphase CT angiography: a new tool for the imaging triage of patients with acute ischemic stroke. Radiology 2015; 275: 510–520.

11.

Forkert

Cheng

Kemmling

Thomalla

Fiehler

. ANTONIA perfusion and stroke. Methods Inf Med 2014; 53: 469–481.

12.

Cheng

Brinkmann

Forkert

, et al. Quantitative measurements of relative fluid-attenuated inversion recovery (FLAIR) signal intensities in acute stroke for the prediction of time from symptom onset. J Cereb Blood Flow Metab 2013; 33: 76–84.

13.

Avants

Tustison

Song

Cook

Klein

Gee

. A reproducible evaluation of ANTs similarity metric performance in brain image registration. Neuroimage 2011; 54: 2033–2044.

14.

Lemaître

Crivello

Grassiot

Alpérovitch

Tzourio

Mazoyer

. Age- and sex-related effects on the neuroanatomy of healthy elderly. NeuroImage 2005; 26: 900–911.

15.

Brett

Leff

Rorden

Ashburner

. Spatial normalization of brain images with focal lesions using cost function masking. NeuroImage 2001; 14: 486–500.

16.

Danielsson

P-E

. Euclidean distance mapping. Comp Graph Image Process 1980; 14: 227–248.

17.

Robnik-Šikonja M and Kononenko I. An adaptation of relief for attribute estimation in regression. In: Fisher DH (ed.) Machine learning: Proceedings of the fourteenth international conference (ICML’97). Vol. 5. Burlington: Morgan Kaufmann, 1997, pp. 296–304.

18.

Friedman

. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. J Am Stat Assoc 1937; 32: 675–701.

19.

Catani

Mesulam

. What is a disconnection syndrome? Cortex 2008; 44: 911–913.

20.

Geschwind

. Disconnexion syndromes in animals and man: part I. Neuropsychol Rev 2010; 20: 128–157.

21.

Doricchi

Thiebaut de Schotten

Tomaiuolo

Bartolomeo

. White matter (dis)connections and gray matter (dys)functions in visual neglect: gaining insights into the brain networks of spatial awareness. Cortex 2008; 44: 983–995.

22.

Hope

TMH

Leff

Price

. Predicting language outcomes after stroke: is structural disconnection a useful predictor? NeuroImage: Clinical 2018; 19: 22–29.

23.

Ernst

Boers

Forkert

, et al. Impact of ischemic lesion location on the mRS score in patients with ischemic stroke: a voxel-based approach. Am J Neuroradiol 2018; 39: 1989–1994.

24.

Anemaet

. Using standardized measures to meet the challenge of stroke assessment. Top Geriatr Rehabil 2002; 18: 47–62.

25.

de Haan

Karnath

H-O

. ‘Whose atlas I use, his song I sing?’ – the impact of anatomical atlases on fiber tract contributions to cognitive deficits after stroke. NeuroImage 2017; 163: 301–309.

26.

Beume

Martin

Kaller

, et al. Visual neglect after left-hemispheric lesions: a voxel-based lesion–symptom mapping study in 121 acute stroke patients. Exp Brain Res 2017; 235: 83–95.

27.

Tarhan

Watson

Buxbaum

. Shared and distinct neuroanatomic regions critical for tool-related action production and recognition: evidence from 131 left-hemisphere stroke patients. J Cognit Neurosci 2015; 27: 2491–2511.