Elsevier

NeuroImage

Volume 98, September 2014, Pages 324-335
NeuroImage

Lesion segmentation from multimodal MRI using random forest following ischemic stroke

https://doi.org/10.1016/j.neuroimage.2014.04.056Get rights and content

Highlights

  • Segmentation of chronic ischemic and other lesions impacting poststroke depression.

  • Hierarchical segmentation of the probable lesion from FLAIR using Bayesian-MRF.

  • Random-forest (RF) probabilistic classification on probable lesion class.

  • Context-rich features used in RF include multimodal MRI and lesion likelihood.

Abstract

Understanding structure–function relationships in the brain after stroke is reliant not only on the accurate anatomical delineation of the focal ischemic lesion, but also on previous infarcts, remote changes and the presence of white matter hyperintensities. The robust definition of primary stroke boundaries and secondary brain lesions will have significant impact on investigation of brain–behavior relationships and lesion volume correlations with clinical measures after stroke. Here we present an automated approach to identify chronic ischemic infarcts in addition to other white matter pathologies, that may be used to aid the development of post-stroke management strategies. Our approach uses Bayesian–Markov Random Field (MRF) classification to segment probable lesion volumes present on fluid attenuated inversion recovery (FLAIR) MRI. Thereafter, a random forest classification of the information from multimodal (T1-weighted, T2-weighted, FLAIR, and apparent diffusion coefficient (ADC)) MRI images and other context-aware features (within the probable lesion areas) was used to extract areas with high likelihood of being classified as lesions. The final segmentation of the lesion was obtained by thresholding the random forest probabilistic maps. The accuracy of the automated lesion delineation method was assessed in a total of 36 patients (24 male, 12 female, mean age: 64.57 ± 14.23 yrs) at 3 months after stroke onset and compared with manually segmented lesion volumes by an expert. Accuracy assessment of the automated lesion identification method was performed using the commonly used evaluation metrics. The mean sensitivity of segmentation was measured to be 0.53 ± 0.13 with a mean positive predictive value of 0.75 ± 0.18. The mean lesion volume difference was observed to be 32.32% ± 21.643% with a high Pearson's correlation of r = 0.76 (p < 0.0001). The lesion overlap accuracy was measured in terms of Dice similarity coefficient with a mean of 0.60 ± 0.12, while the contour accuracy was observed with a mean surface distance of 3.06 mm ± 3.17 mm. The results signify that our method was successful in identifying most of the lesion areas in FLAIR with a low false positive rate.

Introduction

The ischemic stroke lesion changes over time and secondary and remote changes may occur in response to this injury. These dynamic changes are reflected in MRI tissue contrasts, especially between acute (less than 7 days) and chronic (≥ 3 months) stages (Carey et al., 2013a). It is therefore unlikely that a single MR parameter can fully characterize the complexity of the tissue changes post-stroke (Baird and Warach, 1998, Knight et al., 1994, Vannier et al., 1985, Welch et al., 1995). In clinical practice, diffusion weighted images (DWI), T1-weighted (T1W), T2-weighted (T2W) and fluid attenuated inversion recovery (FLAIR) images are often acquired to monitor the progression of stroke. In the acute stage, hyperintense signal observed on DWI provides important information about the anatomical location and extent of the infarcted territory (Carey et al., 2013a; Rivers et al., 2007). In the more chronic phase, T2W and FLAIR images are normally used to delineate the final lesion volume (Xavier et al., 2003). Chronic ischemic lesions appear as hyperintense regions in FLAIR with some heterogeneity within the lesion volume due to ongoing gliosis and demyelination (Clark et al., 1993, Cramer et al., 2006). In some patients, infarct delineation may be complicated by the presence of remote, asymptomatic lesions in addition to the primary lesion of interest. In fact silent brain infarcts and white matter (WM) lesions are common in healthy elderly people (Norrving, 2008, Vermeer et al., 2003) with an estimated 7%–28% of the patients with known stroke history and aged over 65 years having evidence of silent lesions (Bernick et al., 2001, Vernooij et al., 2007). Further, white matter changes may be observed in late stages post-stroke (Lindenberg and Seitz, 2012). Therefore, patients with clinically first-ever stroke may also have multiple lesions either in one or both hemispheres. Estimation of the ischemic lesion, secondary lesions, and remote changes are likely important to appreciate the full impact of the brain injury on functional outcome and recovery after the primary stroke (Carey et al., 2013a).

White matter lesions or hyperintensities (WMH) are observed as hyperintense regions in FLAIR when compared to the surrounding WM tissue. WM lesions are associated with a number of neurological disorders, including aging (Cavalieri et al., 2010, Debette and Markus, 2010, Launer, 2004, O'Sullivan, 2008, Park et al., 2010, Silbert et al., 2008, Teodorczuk et al., 2010, Yamauchi et al., 2002). Despite the probable differences in the underlying tissue pathologies of ischemic strokes and WM lesions, the tissue textures appear similar in T1W, T2W and FLAIR images; i.e. they are hyperintense in T2W and FLAIR and hypointense in T1W images. Adding to the complexity of delineation is the location of the infarct, for instance when the ischemic lesion occurs within deep WM territories that are anatomically coincident with the presence of WM hyperintensities. Differentiating between the two types of lesions is extremely difficult using FLAIR, and even more so in the absence of acute/sub-acute DWI images (Campbell et al., 2012). A number of examples highlighting the complex tissue contrast associated with chronic stroke are given in Fig. 1. In these cases, the primary stroke lesion exhibits both hyperintense (inflammation) and hypointense (tissue atrophy) along with the presence of other WM pathologies on FLAIR images.

Robust automated techniques that can segment the chronic ischemic lesions, WM and other secondary lesions from normal brain tissues offer a streamlined approach to investigate important structural–functional relationships following stroke (Carey et al., 2013b). A variety of automated and semi-automated methods have been proposed for ischemic lesion detection at the acute and sub-acute stages of stroke. Segmentation of intensity histograms followed by region growing and decision trees (Heinonen et al., 1998) was proposed by Dastidar et al. (2000) with extensive manual intervention. Clustering based methods like ISODATA (Iterative Self-organizing Data), mean-shift procedure, and fuzzy C-means have been proposed by Jacobs et al. (2001b,a); Hevia-Montiel et al. (2007); Shen et al. (2010); Seghier et al. (2008); Wilke et al. (2011) who either considered edge-confidence maps or prior tissue probabilities to drive the automated segmentation. Markov random field (MRF)-based automated segmentation from multimodal MRI volumes was suggested by Kabir et al. (2007). Similarly several automated methods have been proposed to segment WM lesions (Jack et al., 2001, Mohamed et al., 2001, Wei et al., 2002). Pattern classification methods involving either K-Nearest Neighbor (KNN), Support Vector Machines (SVM) and Ada Boost have been proposed by Anbeek et al. (2004b,a) and Lao et al. (2008). Kruggel et al. (2008) proposed a texture based segmentation of WM lesions using multi-sort co-occurrence matrices (Kovalev et al., 2001) and principal component analysis. Both the methods of Anbeek et al., 2004a, Anbeek et al., 2004b and Kruggel et al. (2008) provided probabilistic segmentation of the lesions. Yang et al. (2010) proposed an automated segmentation method using the method proposed by van Leemput et al. (2001) combined with random field theory. A multi-level Expectation–Maximization (EM) algorithm was proposed by Wang et al. (2012) and a fuzzy inference-based system aided with prior tissue probabilities was proposed by Admiraal-Behloul et al. (2005) to segment WM lesion loads. Recently, Ong et al. (2012) proposed an outlier detection approach based on adaptive trimmed means algorithm and Box–Whisker plot on intensity histograms.

In general, the methods show good accuracy in terms of volume correlation and visual agreement. However, most focus on delineation of either the primary stroke lesion or the WM lesions, despite evidence and clinical significance of secondary abnormalities. A common reported difficulty associated with lesion segmentation was the challenge to reduce false positives from the FLAIR and T1W MRI due to heterogeneity in tissue contrasts. For ischemic lesion detection, most of the available methods utilize DWI images to help define the anatomical location of the initial infarcted territory acutely post-stroke. These hyperintense regions easily target and isolate the ischemic stroke from other brain pathologies, they are however not always available, especially in the late stages of recovery.

In this paper we investigate the challenges of intensity heterogeneities within the lesion areas, and aim to develop a method to segment stroke with minimized detection of WMH due to normal-aging relying more on statistical concepts of randomness and information theory that allow the amount of heuristic decisions to be minimized without requiring acute DWI MRI. As outlined in Fig. 2, we have employed the methods of Expectation Maximization (EM) likelihood estimation, Bayesian–Markov random field (MRF) segmentation and random forest classification to segment the lesion from the multimodal MRI (FLAIR, T1W, T2W) and apparent diffusion coefficient (ADC) maps obtained from the DWI of the chronic stage. Since the primary lesions are best identified in FLAIR with hyperintense regions, we apply the initial Bayesian–MRF classification on the FLAIR images. Having the probable lesion areas identified from the Bayesian–MRF classification, we refine the segmentation using random forest on the multimodal data with spatial and context-rich features. Random forest classification has been successfully used by Geremia et al. (2011) for multiple sclerosis (MS) lesion segmentation and by Zikic et al. (2012) for identifying regions of high-grade glioma.

The key areas of novelty in our method are:

  • the multilevel splitting and merging of Gaussian intensity classes using Bayesian/MRF classification to extract probable lesion areas,

  • probabilistic classification using random forest of the probable lesion areas, and

In this study, we developed our segmentation framework using MRI data acquired from 36 chronic stroke patients and compared the accuracy of the technique against lesion volumes manually defined by a stroke expert. As our technique is also based, in part on random forest classification as in (Geremia et al., 2011), we further assessed the performance of our method using the data available from the MS segmentation lesion challenge 2008 (www.ia.unc.edu/MSseg/).

Section snippets

Patients

MRI data from 36 patients was used to develop our automated lesion segmentation strategy. Thirty-four patients had radiological evidence of a primary ischemic infarct consistent with their clinical history and presentation, while two other patients did not have evidence of a primary stroke, as determined by an expert neurologist, at 3 months after stroke. Most patients (n = 30) contained additional WM lesions, while 7 subjects also exhibited evidence of a previous lesion. The patient demographics

Parameters and implementation

During the random forest training process, randomly-sampled training samples were used to train the forest to avoid over-fitting (Criminisi et al., 2011). The edge length D for extracting context-rich features was fixed at 3 mm and the offset υ was chosen as 8 mm empirically. The random forest was trained with 100 trees to a maximum depth of 25 restricting the number of samples per leaf node to be 2 samples and information gain of 10 6. The validation for random forest was done using a 4-fold

Discussions

In this paper, we have proposed an automated segmentation method to identify ischemic, WM and other secondary lesions in chronic stroke from multimodal brain images using random forest after initial screening of probable lesion areas from the FLAIR images. The proposed method aimed to segment lesion areas from the FLAIR images obtained at 3-months post-stroke using a Bayesian and MRF classification. The assumption of 5 Gaussian classes and modeling FLAIR lesion intensities as an extra class has

Conclusions

In this paper, we presented an automated method to segment ischemic, WM and other secondary lesions, that involves strategic combination of state-of-the-art methods including Bayesian and MRF segmentation; and random forest classification. We have shown how the different stages of the method aid in segmenting the lesion areas out of the heterogeneous intensities of the multichannel MRI. Incremental improvements have been observed through the subsequent stages and qualitative results have

Acknowledgments

We would like to acknowledge the Stroke Imaging Prevention and Treatment (START) program of research which is supported in part by the CSIRO of Australia through the Preventative Health Flagship Cluster, the National Health and Medical Research Council of Australia, and a Victorian Government Operational Infrastructure Support Grant. In particular, we wish to acknowledge the stroke patients, radiologists and START researchers who contributed to the data collected for this study. LC is supported

References (67)

  • F.B. Mohamed et al.

    Increased differentiation of intracranial white matter lesions by multispectral 3D-tissue segmentation: preliminary results

    Magn. Reson. Imaging

    (2001)
  • B. Norrving

    Leucoaraiosis and silent subcortical infarcts

    Rev. Neurol. (Paris)

    (2008)
  • K.H. Ong et al.

    Automatic white matter lesion segmentation using an adaptive outlier detection method

    Magn. Reson. Imaging

    (2012)
  • C.S. Rivers et al.

    Acute ischemic stroke lesion measurement on diffusion-weighted imaging-important considerations in designing acute stroke trials with magnetic resonance imaging

    J. Stroke Cerebrovasc. Dis.

    (2007)
  • M.L. Seghier et al.

    Lesion identification using unified segmentation–normalisation models and fuzzy clustering

    NeuroImage

    (2008)
  • S. Shen et al.

    An improved lesion detection approach based on similarity measurement between fuzzy intensity segmentation and spatial probability maps

    Magn. Reson. Imaging

    (2010)
  • Y. Wang et al.

    Multi-stage segmentation of white matter hyperintensity, cortical and lacunar infarcts

    NeuroImage

    (2012)
  • M. Wilke et al.

    Manual, semi-automated, and automated delineation of chronic brain lesions: a comparison of methods

    NeuroImage

    (2011)
  • F. Yang et al.

    White matter lesion segmentation based on feature joint occurrence probability and χ2 random field theory from magnetic resonance (MR) images

    Pattern Recogn. Lett.

    (2010)
  • L.D. Alexander et al.

    Correlating lesion size and location to deficits after ischemic stroke: the influence of accounting for altered peri-necrotic tissue and incidental silent infarcts

    Behav. Brain Funct.

    (2010)
  • A.E. Baird et al.

    Magnetic resonance imaging of acute stroke

    J. Cereb. Blood Flow Metal.

    (1998)
  • M.S. Bartlett

    Properties of sufficiency and statistical tests

  • T. Bayes et al.

    An essay towards solving a problem in the doctrine of chances. By the late Rev. Mr. Bayes, communicated by Mr. Price, in a letter to John Canton, M.A. and F.R.S.

    Philos. Trans. R. Soc. Lond.

    (1763)
  • C. Bernick et al.

    Silent MRI infarcts and the risk of future stroke: the cardiovascular health study

    Neurology

    (2001)
  • J. Besag

    On the statistical analysis of dirty pictures

    J. R. Stat. Soc. Ser. B

    (1986)
  • L. Breiman et al.

    Classification and Regression Trees

    (1984)
  • B.C. Campbell et al.

    Assessing response to stroke thrombolysis: validation of 24-hour multimodal magnetic resonance imaging

    Arch. Neurol.

    (2012)
  • L.M. Carey et al.

    Beyond the lesion — neuroimaging foundations for poststroke recovery

    Future Neurol.

    (2013)
  • L. Carey et al.

    START (STroke imAging pRevention and Treatment): a longitudinal stroke cohort study: Clinical Trials Protocol

    Int. J. Stroke

    (2013)
  • M. Cavalieri et al.

    Vascular dementia and Alzheimer's disease — are we in a dead-end road?

    Neurodegener. Dis.

    (2010)
  • S.C. Cramer et al.

    Activity in the peri-infarct rim in relation to recovery from stroke

    Stroke

    (2006)
  • A. Criminisi et al.

    Decision forests for classification, regression, density estimation, manifold learning and semi-supervised learning

  • S. Debette et al.

    The clinical importance of white matter hyperintensities on brain magnetic resonance imaging: systematic review and meta-analysis

    Br. Med. J.

    (2010)
  • Cited by (144)

    • Magnetic Resonance Imaging: Recording, Reconstruction and Assessment

      2022, Magnetic Resonance Imaging: Recording, Reconstruction and Assessment
    View all citing articles on Scopus

    The START Research Team (www.START.csiro.au).

    View full text