Elsevier

NeuroImage

Volume 21, Issue 3, March 2004, Pages 1037-1044
NeuroImage

Probabilistic segmentation of white matter lesions in MR imaging

https://doi.org/10.1016/j.neuroimage.2003.10.012Get rights and content

Abstract

A new method has been developed for fully automated segmentation of white matter lesions (WMLs) in cranial MR imaging. The algorithm uses information from T1-weighted (T1-w), inversion recovery (IR), proton density-weighted (PD), T2-weighted (T2-w) and fluid attenuation inversion recovery (FLAIR) scans. It is based on the K-Nearest Neighbor (KNN) classification technique that builds a feature space from voxel intensities and spatial information. The technique generates images representing the probability per voxel being part of a WML. By application of thresholds on these probability maps, binary segmentations can be obtained. ROC curves show that the segmentations achieve both high sensitivity and specificity. A similarity index (SI), overlap fraction (OF) and extra fraction (EF) are calculated for additional quantitative analysis of the result. The SI is also used for determination of the optimal probability threshold for generation of the binary segmentation. Using probabilistic equivalents of the SI, OF and EF, the probability maps can be evaluated directly, providing a powerful tool for comparison of different classification results. This method for automated WML segmentation reaches an accuracy that is comparable to methods for multiple sclerosis (MS) lesion segmentation and is suitable for detection of WMLs in large and longitudinal population studies.

Introduction

In the last decade, many studies have focused on the prevalence of cerebral white matter lesions (WMLs) in the elderly population or in patients with cardiovascular risk factors. In both patient groups, WMLs are a common finding on cranial MR imaging. Population studies like the Cardiovascular Health Study or the Rotterdam Scan Study have shown that WMLs are associated with age, clinically silent stroke, higher systolic blood pressure, lower forced expiratory volume in 1 s, hypertension, atrial fibrillation, carotid and peripheral arterioscleroses, impaired cognition and depression De Groot et al., 2000a, De Groot et al., 2000b, Longstreth et al., 1996. Furthermore, it has been shown that stroke patients with a large WML load have an increased risk of hemorrhagic transformation, higher preoperative risk of a disabling or fatal stroke during endarterectomy or intercerebral hemorrhage during anticoagulation therapy (Briley et al., 2000). The increased interest in WML research may improve diagnosis and prognosis possibilities for patients with cardiovascular symptoms.

Since WML patterns are very heterogeneous, ranging from punctuate lesions in the deep white matter till large confluent periventricular lesions, the scoring of WMLs is complicated and it has been shown that different visual rating scales lead to inconsistencies between WML studies (Mantyla et al., 1997). Commonly used ordinal WML scoring methods, such as used in the Cardiovascular Health Study or the Rotterdam Scan Study, offer semiquantitative information on the prevalence of WMLs. Exact spatial information is useful since it has been suggested that specific WML patterns are associated with specific symptoms Benson et al., 2002, Smith et al., 2000. Moreover, for longitudinal studies and to demonstrate relatively small changes in WML patterns, accurate information of WML volume and location is essential. In this respect, the use of an automated segmentation method that detects WMLs with a high sensitivity and specificity, which are demonstrated in a quantitative and objective way, could be advantageous. Successful methods have been developed for the detection of multiple sclerosis (MS) lesions Alfano et al., 2000, Goldberg-Zimring et al., 1998, Guttman et al., 1999, Kamber et al., 1995, Van Leemput et al., 2001, Warfield et al., 1995a, Warfield et al., 1995b, Wei et al., 2002, Zijdenbos et al., 2002. For the more complicated issue of WMLs in general, also some segmentation algorithms exist Jack et al., 2001, Mohamed et al., 2001, Wei et al., 2002. However, these methods evaluate their results only by visual inspection or measurement of lesion volume. The aim of our research was to develop an automated WML segmentation algorithm, which is fully reproducible and quantitatively validated on a voxel basis.

In this study, we present a method for automatic segmentation of WMLs that is based on a supervised K-Nearest Neighbor (KNN) classification technique using information from T1-weighted (T1-w), inversion recovery (IR), proton density-weighted (PD), T2-weighted (T2-w) and fluid attenuation inversion recovery (FLAIR) scans by probability estimation of voxels being part of a lesion.

Section snippets

Patients

Twenty patients with arterial vascular disease [transient ischemic attack (TIA), n = 4; peripheral arterial disease, n = 3; coronary artery disease, n = 7; renal artery disease, n = 1; abdominal aorta aneurysm, n = 5] were included in this study. The mean age of the patients was 66 years (mean ± SD: 65.6 ± 7.7, range: 49–75). Seventeen patients were male.

MR imaging

MRI studies were performed on a Philips Gyroscan ACS-NT 1.5-T whole body system (Philips Medical Systems, Best, The Netherlands). All patients

Results

KNN classification has been performed five times per patient, according to the five different feature sets. In Fig. 2, Fig. 3, Fig. 4, example images are shown of the classification results of patients, with a small, moderate and large lesion load, with feature set Fxyz. For each patient category, the following images are shown: FLAIR, manual segmentation, probability map and a color image with segmentations generated by applying thresholds of 0.3, 0.5 and 0.8 to the probability map. The images

Discussion

The combination of spatial information and signal intensities of MR images in KNN classification provides a technique for WML segmentation with a high sensitivity and specificity for all patient categories, which is shown by ROC curves. The method generates a probability map, containing the probabilities of voxels being a lesion. The main advantage of determination of the lesion probabilities over direct classification of voxels into lesion or non-lesion is that it provides an opportunity to

References (34)

  • J.C. De Groot et al.

    Cerebral white matter lesions and depressive symptoms in elderly adults

    Arch. Gen. Psychiatry

    (2000)
  • J.C. De Groot et al.

    Cerebral white matter lesions and cognitive function: the Rotterdam Scan Study

    Ann. Neurol.

    (2000)
  • L.R. Dice

    Measures of the amount of ecologic association between species

    Ecology

    (1945)
  • R.O. Duda et al.

    Pattern Classification

    (2001)
  • C.R.G. Guttman et al.

    Quantitative follow-up of patients with multiple sclerosis using MRI: reproducibility

    J. Magn. Reson. Imaging

    (1999)
  • C.R. Jack et al.

    FLAIR histogram segmentation for measurement of leukoaraiosis volume

    J. Magn. Reson. Imaging

    (2001)
  • M. Kamber et al.

    Model-based 3-D segmentation of multiple sclerosis lesions in magnetic resonance brain images

    IEEE Trans. Med. Imag.

    (1995)
  • Cited by (300)

    View all citing articles on Scopus
    View full text