Elsevier

Brain and Language

Volume 94, Issue 2, August 2005, Pages 167-177
Brain and Language

Identifying lesions on structural brain images—Validation of the method and application to neuropsychological patients

https://doi.org/10.1016/j.bandl.2004.12.010Get rights and content

Abstract

The study of neuropsychological disorders has been greatly facilitated by the localization of brain lesions on MRI scans. Current popular approaches for the assessment of MRI brain scans mostly depend on the successful segmentation of the brain into grey and white matter. These methods cannot be used effectively with large lesions because lesions usually impair segmentation. We propose a novel, fully automated approach for the delineation of brain lesions on MR scans. This method involves comparing a skull stripped, smoothed, unsegmented T1 images to a control group using the general linear model. We tested this method by using images with simulated lesions of different sizes and images containing real lesions from patients with language deficits. We also tested how varying the size of the Gaussian smoothing kernel affects detection. The simulation was informed by findings of a lesion morphological study also presented here. The proposed method detected simulated lesions effectively in the range of 30–90% < normal signal. Smoothing kernels in the range of 8–12 mm resulted in the most accurate lesion detection. Both artificial and real lesions were optimally detected when the results were uncorrected for multiple comparisons at p < .001. This proposed method produced highly satisfactory results and can be used to generate reproducible detection of lesions.

Introduction

Magnetic resonance imaging (MRI) has revolutionized the in vivo examination of the brain in healthy, neurological, and psychiatric populations affected by conditions such as dementia, stroke, and brain trauma. MRI can provide both structural and functional information and is favored over other imaging techniques because it is non-invasive and uses non-ionizing radiation. The assessment and localization of brain lesions on T1 weighted scans facilitates the investigation of the brain bases of cognition by providing an important method for relating neuroanatomy and behavior. Traditionally, lesion assessment has involved using regions-of-interest (ROI) methods whereby a trained professional manually defines the healthy and/or diseased tissue (Bates et al., 2003, Damasio and Damasio, 1989), or a predefined template is superimposed on a brain region to measure the volume of healthy and/or diseased tissue within this template (Lawrie and Abukmeil, 1998, Mega et al., 2000). Defining of ROIs can be carried out using established tracing software (e.g., Analyze Biomedical Imaging Resource, Mayo Foundation). The total measured lesion and/or healthy tissue volumes are entered into a statistical analysis to calculate population deficits (Lawrie & Abukmeil, 1998) or to correlate performance on neuropsychological tests to lesion site (Adolphs et al., 2000, Bates et al., 2003, Grossman et al., 2004, Tranel et al., 2001).

However, there are distinct limitations with these methods. The use of predefined anatomical templates for the measurement of healthy and/or diseased tissue volumes can produce ambiguous results when a small lesion is assessed with a large template (resulting in a minor contribution of the lesion to the overall result and loss of region specificity), and conversely in cases where a small template is applied to a lesion larger than the template such that the full extent of that lesion is not revealed. Take for example the case where a superior temporal gyrus (STG) template is used to count lesioned voxels in a group of patients. The actual lesions could be larger than the STG template and as such their implications are not fully evaluated by the examination of lesioned voxels within the specific STG template. ROI methodologies can also be affected by issues such as operator subjectivity, and reproducibility. These types of analyses can nonetheless produce useful insights into the neural underpinnings of cognitive processes, provided they are not confined to predetermined regions, the templates used are constructed meticulously and the lesion tracing is carried out by experienced professionals (Adolphs et al., 2000, Bates et al., 2003, Dronkers and Ludy, 1998, Grossman et al., 2004, Tranel et al., 2001). ROI type analyses can additionally serve as a standard to validate automated approaches (Thompson, Rapoport, Cannon, & Toga, 2003).

Developments in image analysis algorithms have facilitated the parallel growth of fully automated brain image assessment methodologies that allow the examination of the whole brain at the voxel level. The most recent of these methodologies have been applied in the analysis of cross-sectional studies and involve the statistical comparison of images from groups with various pathologies to control groups, with complementary analyses involving correlations of signal at the voxel level with variables such as age or scores from neuropsychological tests (Ashburner and Friston, 2000, Good et al., 2001, Shapleske et al., 2002, Smith et al., 2002). The same type of fully automated framework can be used in longitudinal studies and involves aligning scans of participants obtained at different time points, and performing statistical comparisons for the detection of atrophy brought on by age or disease (Fox & Freeborough, 1997).

Automated approaches depend critically on successful preprocessing of the structural brain images. Preprocessing commonly involves normalization to a standard space, segmentation of the images to grey/white matter, and cerebrospinal fluid (CSF), and finally smoothing with a Gaussian kernel. Anatomical correspondence across all the brain images is necessary before any statistical comparisons between groups can be carried out. Correspondence is achieved with spatial alignment/normalization of brain images to a template (based commonly on the standardized atlas developed by Talairach & Tournoux, 1988) and works by reducing differences in brain position, size and shape across subjects (Fox et al., 1985, Ashburner et al., 2003). The success of the group comparisons depends largely on accurate spatial normalization of the brain images since this allows the comparison of homologous anatomical regions across subjects. Although non-linear spatial normalization can be compromised by the presence of abnormalities/lesions, this problem can often be resolved by utilizing weighting in the form of masks in order to exclude the lesion during non-linear normalization (Brett, Leff, Rorden, & Ashburner, 2001) or by penalizing unlikely deformations (Ashburner et al., 1998, Ashburner and Friston, 1999).

A second important assumption in this type of analysis is that partitioning of the T1 weighted structural images (segmentation) into grey, white matter, and CSF is successful. Statistical comparisons between controls and populations with pathologies are commonly carried out on segmented images (grey, white, and CSF) and therefore, incorrect tissue segmentation can lead to grey matter segments being compared to white matter segments and the reverse. Segmentation algorithms depend on signal differences between grey and white matter to work successfully. In the presence of abnormalities, the differences in intensity between grey and white matter decreases and consequently segmentation algorithms may not achieve correct tissue partition (Kaus et al., 2001, Stamatakis and Tyler, 2003, Warfield et al., 1995, Vinitski et al., 1997).

Given the difficulty of achieving accurate segmentation within lesioned tissue and the constraints this imposes on the fully automated methodologies discussed here, we developed a fully automated method for the assessment of T1 weighted brain scans with abnormalities without using segmentation into grey and white matter. Segmentation can be useful when assessing small to moderate changes in the brain such as atrophies, but our purpose was to assess the anatomical extent of large lesions across any type of tissue. In the first instance, we tested and fine-tuned the method using simulated lesions to control for variables such as levels of signal reduction and lesion size. Following this methodological validation, we tested the procedure on images with real lesions. We first present data to substantiate our reservations concerning existing segmentation procedures when applied to brains with large lesions. We then present data from a study we carried out to find the degree of signal drop in real lesions. The results of this analysis constituted the background for the construction of the simulated lesions. Finally, we present results from the analysis/detection of both simulated and real lesions in a sample of patients with various types of language deficits. Our research focuses on language and consequently we provide examples from patients with language deficits, but the method could be used to assess lesions on patients with any type of deficit.

Section snippets

Materials and methods

Scanning was carried out on a 3T Brucker system (Bruker BioSpin MRI) at the Wolfson Brain Imaging Center, Cambridge. Thirty-three right-handed subjects aged 19–35 years (mean = 23 years, SD = 4.4 years, 17 males, 16 females) participated in this study. All gave informed consent. Scanning was approved by Addenbrookes NHS Trust Ethical Committee. For each participant, a T1-weighted 3D MRI volume of the whole brain was acquired using a spoiled-gradient recalled (SPGR), gradient echo sequence (TR 17.7 

Comparing segmentation using different algorithms

The first step in this study involved comparing the accuracy of different methodologies to segment five T1 weighted images with lesions. Typical examples of these segmentation results are in Fig. 2. The grey matter segment images in Fig. 2 show that segmentation was not successful. Errors in segmentation in the locality of the lesion are similar across the four algorithms with more prominent misclassification of white matter as grey matter in areas with pronounced reduction in signal. The white

Discussion

This paper describes an automated method of detecting lesions in patients with various neuropsychological deficits. Existing methods for the assessment of lesions depend critically on correct image segmentation to grey/white matter and CSF. We demonstrated that in the presence of large lesions four popular segmentation algorithms (Ashburner and Friston, 1997, Ashburner and Friston, 2000, Ashburner, 2002, Van Leemput et al., 1999a, Van Leemput et al., 1999b, Zhang et al., 2001) fail to segment

Conclusions

The method described here accurately detected large lesions in the brains of brain-damaged patients, and can be used to generate highly reproducible detection of lesions. Moreover, it has the advantage of speed, and detection accuracy over manual methods. A smoothing kernel of 10 mm produced the most reliable lesion detection with maximum true positive and true negative rates and minimum false positive and false negative rates. Although the method underestimated very small lesions (within the

Acknowledgments

This research was supported by an MRC programme grant to L.K. Tyler. We thank the radiographers at the Wolfson Brain Imaging Unit Cambridge, UK, for their help with the study.

References (39)

  • J. Ashburner et al.

    Nonlinear spatial normalization using basis functions

    Human Brain Mapping

    (1999)
  • Ashburner, J. (2002). Another MRI bias correction approach. HBM02 in Neuroimage,...
  • E. Bates et al.

    Voxel-based lesion-symptom mapping

    Nature Neuroscience

    (2003)
  • E. Bunn et al.

    Category-specific deficits: The role of familiarity re-examined

    Neuropsychology

    (1998)
  • H. Damasio et al.

    Lesion analysis in neuropsychology

    (1989)
  • N.F. Dronkers et al.

    Handbook of neurolinguistics

    (1998)
  • N.C. Fox et al.

    Brain atrophy progression measured from registered serial MRI: Validation and application to Alzheimer’s disease

    Journal of Magnetic Resonance Imaging

    (1997)
  • P.T. Fox et al.

    A stereotactic method of anatomical localization for positron emission tomography

    Journal of Computer Assisted Tomography

    (1985)
  • K.J. Friston et al.

    Statistical parametric maps in functional imaging: A general linear approach

    Human Brain Mapping

    (1995)
  • Cited by (0)

    View full text