MRI white matter lesion segmentation using an ensemble of neural networks and overcomplete patch-based voting

Highlights

• A novel method to segment white matter hyperintense lesions on FLAIR images is presented.

• The proposed method uses an overcomplete patch-wise labelling within an neural network ensemble-based classifier.

• Comparison with state of the art methods show the competitive results in terms of accuracy and efficiency.

Abstract

Accurate quantification of white matter hyperintensities (WMH) from Magnetic Resonance Imaging (MRI) is a valuable tool for the analysis of normal brain ageing or neurodegeneration. Reliable automatic extraction of WMH lesions is challenging due to their heterogeneous spatial occurrence, their small size and their diffuse nature. In this paper, we present an automatic method to segment these lesions based on an ensemble of overcomplete patch-based neural networks. The proposed method successfully provides accurate and regular segmentations due to its overcomplete nature while minimizing the segmentation error by using a boosted ensemble of neural networks. The proposed method compared favourably to state of the art techniques using two different neurodegenerative datasets.

Introduction

White matter hyperintensities (WMH) are regions of increased MR signal in T2-Weighted (T2 W) and FLuid Attenuated Inversion Recovery (FLAIR) images that are distinct from cavitations (Wardlaw et al., 2013). The number, size and location of WMH can provide important information into the aetiology and progression of various diseases. This has been extensively studied in normal ageing, cerebrovascular disease, dementia (Kuo and Lipsitz, 2004; Debette and Markus, 2010) and its influence on co-morbidities (Lee et al., 2015). The presence, topography and volume of WMH is used as biomarkers for stroke (Kuller et al., 2004; Wong et al., 2002), small vessel cerebrovascular disease (CVD) (Schmidt et al., 2004), dementia (Debette and Markus, 2010) and in multiple sclerosis (MS) (Filippi and Rocca, 2011).

In clinical practice, qualitative visual rating scales have been frequently used (Scheltens et al., 1998). However, in order to use WMH volume and spatial location as a biomarker, lesions need to be accurately and precisely segmented. Some promising early-automated methods have been used in longitudinal clinical studies (Mäntylä et al., 1997), with later studies focused on improving the sensitivity, specificity and robustness of automated WMH segmentation. Manual and semi-automated segmentation of WMH is a tedious process requiring trained observers and several hours per image for manual delineation by an expert making it unsuitable for routine clinical and research usage (Udupa et al., 1997). Moreover, manual segmentation is prone to inter and intrarater variability.

With many large clinical studies investigating ageing, cerebrovascular disease, and dementia, there is a need for robust, repeatable, accurate, and automated techniques for the segmentation of WMH. In recent years, several methods have been proposed to automatically segment WMH in CVD and in MS. While the underlying pathology is different, the radiological signatures of MS and CVD are sufficiently similar that methods developed for one have good performance for the other (Caligiuri et al., 2015). Demyelinating lesions of MS and cerebrovascular disease appear as hyperintense regions on T2 W and FLAIR images. Initial approaches to segment of WMH relied on the higher intensity in lesions compared to surrounding tissue to threshold the image after correction for inhomogeneities (Jack et al., 2001; Souplet et al., 2008). The hyperintensity assumption is challenged by the natural variation in intensity found in normal tissues across the brain such as the septum pellucidum and CSF flow artefacts around the ventricles (Neema et al., 2010). Other problem includes residual intensity inhomogeneity, even after correction.

To address these issues, more complex methods have been proposed. These methods can be classified into unsupervised and supervised. Unsupervised methods rely on the natural separation of image features using clustering type approaches. For example, the lesion growth algorithm (LGA) publicly available as part of the lesion segmentation toolbox (LST) has been widely used (Schmidt et al., 2012). In this method, both T1W and FLAIR images are required to first compute a map of possible candidate lesions whose centres are then used as seeds to segment the entire lesions using region growing. Also included in the LST toolbox, a more recent method, the lesion prediction algorithm (LPA), only requires FLAIR images as input. Within the same category, (Weiss et al., 2013) proposed a dictionary learning-based approach that segments lesions as outliers from a projection of the dataset onto a normative dictionary. Similarly (Raniga et al., 2011) used a generative model to segment lesions by detecting outlier tissue. More classical unsupervised approaches have also been proposed (Admiraal-Behloul et al., 2005).

Supervised methods require training datasets where WMH lesions are manually annotated by experts. This type of methods can work on single channel (FLAIR or T2W) or multi-channel data (FLAIR or T2W and T1W and PDW). Supervised methods for WMH segmentation typically involve machine learning methods at a voxel level with pre and/or post processing steps to improve the sensitivity and specificity of the results. Such methods have used support vector machines (Lao et al., 2008), k-nearest neighbours (Steenwijk et al., 2013), random forests (Ithapu et al., 2014; Geremia et al., 2010: Jesson and Arbel, 2015), artificial neural networks (Dyrby et al., 2008), deep learning (Brosch et al., 2015; Ghafoorian et al., 2016; Valverde et al., 2017) or multiatlas patch-based label fusion methods (Guizard et al., 2015). All these methods were trained on either single or multi-channel voxel intensities jointly with some other context-related features and typically within a standardized anatomical space. Independently of the features used, these methods perform the classification step at the voxel level, and do not take into account label spatial correlations, which might affect their performance.

To overcome the lack of local consistency (i.e. each voxel is labelled independently of neighbour voxels) of the methods performing voxel-wise classification, we propose an automatic pipeline for hyperintense lesion segmentation based on the use of patch-wise neural network classifier that segments the lesions taking in consideration patch labels local context in an overcomplete manner which further reduces classification errors. This pipeline benefits from some pre-processing steps aimed to improve the image quality and to locate it in a standardized geometrical and intensity space. The proposed method which extends a previous method recently published (Manjón et al., 2016) uses a boosting based ensemble learning strategy to minimize the classification error. In the following sections, the proposed method is described and compared to manual assessment and two state-of-the-art methods. This comparison is performed on data from two datasets.