Automatic white matter lesion segmentation using contrast enhanced FLAIR intensity and Markov Random Field

https://doi.org/10.1016/j.compmedimag.2015.08.005Get rights and content

Highlights

  • An automatic method for white matter lesion segmentation (WML).

  • A novel filter is used to enhance the contrast of WML.

  • A robust Markov Random Field is used to remove false positive WMLs.

  • Evaluated with respect to the grading of two expert graders.

  • Compared with state of the art methods in two datasets.

Abstract

White matter lesions (WMLs) are small groups of dead cells that clump together in the white matter of brain. In this paper, we propose a reliable method to automatically segment WMLs. Our method uses a novel filter to enhance the intensity of WMLs. Then a feature set containing enhanced intensity, anatomical and spatial information is used to train a random forest classifier for the initial segmentation of WMLs. Following that a reliable and robust edge potential function based Markov Random Field (MRF) is proposed to obtain the final segmentation by removing false positive WMLs. Quantitative evaluation of the proposed method is performed on 24 subjects of ENVISion study. The segmentation results are validated against the manual segmentation, performed under the supervision of an expert neuroradiologist. The results show a dice similarity index of 0.76 for severe lesion load, 0.73 for moderate lesion load and 0.61 for mild lesion load. In addition to that we have compared our method with three state of the art methods on 20 subjects of Medical Image Computing and Computer Aided Intervention Society's (MICCAI's) MS lesion challenge dataset, where our method shows better segmentation accuracy compare to the state of the art methods. These results indicate that the proposed method can assist the neuroradiologists in assessing the WMLs in clinical practice.

Introduction

White matter lesions (WMLs) have been associated with various diseases such as stroke [1], Alzheimer's disease [2], [3], late life depression [4] and impairment of gait [5] etc. Different studies have shown a strong relationship between increased WMLs volume and cognitive decline [6], [7]. Magnetic resonance imaging (MRI) is used for the visualization of WMLs. WMLs appear as hyperintense signal in T2-weighted and fluid attenuated inversion recovery (FLAIR) MRI images. Manual segmentation and quantification of white matter lesion is time consuming, expensive, subjective and impractical for large scale longitudinal studies. Several supervised and unsupervised algorithms are proposed in the literature for the automatic segmentation of white matter lesions [8], [9], [10], [11], [12], [13], [9]. Among the many existing segmentation methods, the accurate segmentation of white matter lesions remains a challenging task.

Low contrast and intensity inhomogeneity are the major challenges of automatic WMLs segmentation in brain MRI images. Histogram equalization [14] and adaptive histogram equalization [14] are two well known contrast enhancement methods in image processing. Since the pathology is not evenly distributed in the medical images therefore global stretching of the histogram as proposed in histogram equlization can suppress the pathology, on the contrary local stretching of the histogram as proposed in adaptive histogram equalization can over-amplify the intensity with false identification of pathology [14]. Presently in medical image processing, contrast enhancement filters are developed considering the properties of the corresponding image modality [15]. FLAIR MRI is the commonly used MRI modality for white matter lesion segmentation [16]. But a few methods have been proposed for the contrast enhancement of FLAIR intensity. Khademi et al. [9] have proposed a method to enhance the contrast of the FLAIR intensity using gradient information. In most of the cases white matter lesions have fuzzy edges thus gradient information based intensity enhancement does not work for all cases. The spatial distribution of the white matter lesion can be useful information for the contrast enhancement of FLAIR intensity. As we know, the ratio of the number of white matter lesion pixel is very low compared to the number of healthy pixel, therefore if we select a set of non-overlapping pixels from the brain area then majority of them have the high probability of being a member of healthy region. Based on this assumption we hypothesize that a lesion pixel will show high contrast with respect to a set of non-overlapping pixels.

Apart from intensity information, anatomical and textural information are also used to segment white matter lesions (WMLs) [8], [10], [11], [12], [13]. Geremia et al. [17] proposed a method where a random forest classifier [18] is used to segment WMLs. The random forest classifier is trained by using multichannel MRI intensity and long range contextual features. But in their method pixel neighbourhood information are not considered for the final segmentation. Since Markov Random Field (MRF) models spatial dependency between neighbouring pixels. Hence it is used widely in white matter lesion segmentation [19], [20]. In those methods image properties such as intensity, gradient, texture, etc. are not included in the neighbourhood model; they only consider the initial labelling information of neighbourhood pixels for final segmentation. As a result, they might show over-smoothing when the number of lesion is small. We hypothesize that incorporation of image properties into the neighbourhood structure of MRF can improve the segmentation accuracy when the number of lesions is small.

In this paper, we have proposed a novel method to enhance the contrast of WMLs in the FLAIR MRI images. Then the enhanced intensity is used with the spatial location and anatomical information to train a random forest classifier to obtain the initial segmentation. Following that a reliable and robust MRF is proposed for obtaining final segmentation by removing false positive WMLs. Our proposed MRF maximizes the segmentation accuracy by using a novel edge function.

The rest of the paper is organized as follows: in Section 2, the details of our method is described. The results and discussions are presented in Section 3. Finally, Section 4 concludes the paper.

Section snippets

Dataset

We primarily investigated a dataset of ENVISion study [21]. The dataset consist of 24 MRI images, selected from a larger cohort of elderly subjects with hypertension. In order to include a wide variety of WMLs severity, subjects are selected based on the severity of WMLs. Monash university health and research ethics committee have approved the study and all subjects gave written informed consent prior to participation. ENVISion dataset contains two modalities of MRI named as T1-weighted and

Results and discussion

White matter lesion (WML) loads are typically divided into three groups: severe lesion load (>10 ml), moderate lesion load (between 6 ml to 10 ml) and mild lesion load (0 ml to 5 ml) [35]. In our method, random forest (RF) classifier is used for the initial segmentation of white matter lesions as mentioned in Section 2.5. Then Markov Random Field (MRF) is used to remove false positive lesions and obtain the final segmentation. We have compared four state of the art classifiers which are random

Conclusion

We have developed a fully automated method for the detection of white matter lesions (WMLs) in brain MR images using enhanced intensity, anatomical and spatial feature based random forest and Markov Random Field (MRF). Our proposed MRF is based on a novel edge potential function that uses spatial and intensity interaction in a local neighbourhood to obtain the optimal segmentation outcome. The proposed method is evaluated on two distinct datasets (ENVISion dataset with WMLs and MICCAI dataset

References (39)

  • P. Aljabar et al.

    Multi-atlas based segmentation of brain images: atlas selection and its effect on accuracy

    Neuroimage

    (2009)
  • D. Yamamoto et al.

    Computer-aided detection of multiple sclerosis lesions in brain magnetic resonance images: false positive reduction scheme consisted of rule-based, level set method, and support vector machine

    Comput Med Imaging Graph

    (2010)
  • L. Shi et al.

    Automated quantification of white matter lesion in magnetic resonance imaging of patients with acute infarction

    J Neurosci Methods

    (2013)
  • T.Y. Wong et al.

    Cerebral white matter lesions, retinopathy, and incident clinical stroke

    JAMA: J Am Med Assoc

    (2002)
  • N. Hirono et al.

    Impact of white matter changes on clinical manifestation of Alzheimer's disease: a quantitative study

    Stroke

    (2000)
  • L.L. Herrmann et al.

    White matter hyperintensities in late life depression: a systematic review

    J Neurol Neurosurg Psychiatry

    (2008)
  • C.R. Guttmann et al.

    White matter abnormalities in mobility-impaired older persons

    Neurology

    (2000)
  • H. Jokinen et al.

    Longitudinal cognitive decline in subcortical ischemic vascular disease – the LADIS study

    Cerebrovasc Dis

    (2009)
  • R. Au et al.

    Association of white matter hyperintensity volume with decreased cognitive functioning: the Framingham heart study

    Arch Neurol

    (2006)
  • Cited by (24)

    • Automatic segmentation of white matter hyperintensities from brain magnetic resonance images in the era of deep learning and big data – A systematic review

      2021, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      The proposed method was cross validated with leave-one-subject-out evaluation in six studies (Stone et al., 2016; Moeskops et al., 2018; Jiang et al., 2018; Knight et al., 2018; Li et al., 2018; Wu et al., 2019a). Fourteen studies described a method for false positive removal (Sudre et al., 2017; Manjón et al., 2018; Stone et al., 2016; Moeskops et al., 2018; Rincón et al., 2017; Bowles et al., 2017; Li et al., 2018; Ghafoorian et al., 2017; Roy et al., 2015; Zhan et al., 2017; Hong et al., 2020; Fiford et al., 2020; Rachmadi et al., 2020; Ding et al., 2020). Fazekas visual rating scale was used in the validation of the results in nine studies (Rachmadi et al., 2018; Qin et al., 2018; Guerrero et al., 2017; Griffanti et al., 2016; Ling et al., 2018; Moeskops et al., 2018; Jiang et al., 2018; Bowles et al., 2017; Rachmadi et al., 2020).

    • Voxel-Wise Logistic Regression and Leave-One-Source-Out Cross Validation for white matter hyperintensity segmentation

      2018, Magnetic Resonance Imaging
      Citation Excerpt :

      Other unsupervised paradigms for segmentation of WMH are not very common, though the works by Admiraal-Behloul et al. [33], Gibson et al. [46], and Valverde et al. [75] use Fuzzy C Means, while a novel topology-driven framework for brain segmentation called TOADS was proposed in [82], and subsequently adapted for lesion segmentation by Shiee et al. [47]. WMH segmentation has also been approached using classical supervised classification methods like K-Nearest Neighbours [20,31,32,35,67,73], Support Vector Machines [34,48,54,58], and ensembles such as Random Forests and AdaBoost [38,44,50,58,65]. In such models, the employed features always include voxel intensities, in addition to some combination of spatial coordinates [31,32,73], tissue priors [20,35,50,67,65], and/or patch-based features [38,44,48,54,58,73].

    • Case-control comparison brain lesion segmentation for early infarct detection

      2018, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Unsupervised clustering techniques produce good results in finding different regions in tumour cases (Wiselin and Ganesan, 2005). The K-means clustering algorithm (Selvakumar et al., 2010; Yusoff et al., 2013; Alan et al., 2014), fuzzy c-means (FCM) clustering algorithm (Samina and Hammad, 2010; Shan et al., 2008; Umut et al., 2007; Chack and Sharma, 2015) and Markov Random fields (Pallab et al., 2015; Kabir et al., 2007) belong to unsupervised clustering techniques. On the other hand, a supervised method trains a model by using training data with target classes to segment an image into several regions of interest.

    • Segmentation of white matter hyperintensities using convolutional neural networks with global spatial information in routine clinical brain MRI with none or mild vascular pathology

      2018, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      While these approaches have been shown to be useful, we think that relying on maps that are too discrete may result in ignoring some subtle features of WMH. Hence, we propose to incorporate such coordinate values in the form of a synthetic volume (Steenwijk et al., 2013; Roy et al., 2015) as input to a CNN segmentation architecture through additional channels, which means providing spatial information at the convolutional level of the CNN. In this way, we can learn the tendency of WMH from more simple and weaker contextual data such as Cartesian coordinates and polar coordinates of the area in brain images.

    View all citing articles on Scopus
    View full text