An automated template-based adaptive threshold approach for measuring ventricular volume enlargement in mouse brain MR microscopy

Authors


Lan Lin, Biomedical Research Center, College of Life Science and Bioengineering, Beijing University of Technology, Beijing, 100022, China. Tel: 86-010-67391610; Fax: 86-010-67391610; e-mail: lanlin@bjut.edu.cn

Summary

Mouse models for human diseases play an important role in developing therapeutic measures and understanding disease development and mechanisms. Ventricular enlargement is an objective and sensitive biomarker of neuropathological change associated with mild cognitive impairment (MCI) and Alzheimer's disease (AD). Thus, imaging based volumetric measures of mouse brain ventricle can be used as a biomarker for detecting and monitoring pathology. However, till now most mouse ventricular segmentation is still based on manual tracing, and current region of interests (ROIs) labelling approaches on mouse brain don't work well on small brain structure like ventricle. In this paper, an automated template-based method was developed to evaluate for the identification of a ventricular ROIs in mouse brain imaging studies. Monte Carlo simulation was applied to evaluate the efficiency of this approach in detecting computer-simulated ventricle in laboratory mice. The method demonstrated a satisfactory performance with consistent high correlations between the detected and simulated volume changes. This approach can be used to investigate the ventricular volume changing in transgenic mice and other putative animal models of AD.

Introduction

Alzheimer's disease (AD) (Alzheimer's Association, 2012), most common form of dementia, is a chronic neurodegenerative disorder, characterized by a progressive loss of cognitive function. Despite considerable progress in our understanding of the AD, a comprehensive description of the molecular pathology of this disease has yet to be elucidated. AD models of mouse (Janus & Westaway, 2001; Chuang, 2010; Hall & Roberson 2011), which were designed to reproduce various aspects of the pathological, biochemical, and/or behavioural characteristics of AD, are crucial for understanding the consequences of the pathological and biochemical changes that occur as the disease progresses and for investigating the effectiveness of potential treatment or prevention strategies. In many situations, treatment is evaluated longitudinally with or without a control group. Longitudinal design provides unique advantage that each animal serves as its own control so that the progression as well as regression of the disease can be assessed in relation to time and/or interventions. Thus, there is a growing need for noninvasive tracking techniques for monitoring mouse brain changes in normal and abnormal development.

One important indicator of dementia progression was the ventricular volume enlargements. Measurements of ventricular space volume changes from sequential MRI have been long used as a powerful noninvasive tool for tracking the progression of brain changes in human AD studies. Ventricular volume was reported to be larger in dementia compared to mild cognitive impairment (MCI), larger in MCI compared to controls, and larger in Possible-AD dementia compared to Probable-AD (Carmichael et al., 2006). Study by Wang et al. (2002) demonstrated that the brain atrophy associated with the lateral ventricular enlargement makes up about one-third of the total atrophy volume in the cerebrum in AD. Ott et al. (2010) showed that increased ventricular volume may be associated with decreased levels of CSF amyloid β-peptide (Aβ) in preclinical AD. Nestor et al. (2008) showed that ventricular enlargement represents an objective and sensitive biomarker of disease progression in subjects with MCI and subjects with AD.

The biomarker of ventricular enlargement, established in human studies, has been further transformed to animal studies to facilitate drug development for the disease. Chen et al. (2011) used high-resolution magnetic resonance microscopy (MRM) to track the longitudinal development of brain atrophy in normal aging mice. Their results showed that the ventricle size increased with age, while it began expansion during the early stage of life and continuing to old age. In another study, McDaniel et al. (2001) used MRM to track the longitudinal development of brain atrophy in mouse models of neurodegeneration (cerebral ischemia) and demonstrated ventricle enlargement postischemia. In both of these studies, measurements of ventricular volume changes were restricted to manual tracing on the MRM images by the rators. Several studies have addressed how to segment mouse brain into a group of region of interests (ROIs) (Ali et al., 2005; Kovačevićet al., 2005; Ma et al., 2005; Lebenberg et al., 2010) Those approaches are mainly based on the registration between reference brain and subject brain. The potential weakness of those approaches is that registration errors can cause structure mismatches and a postregistration correction is required. Ali et al. (2005) found that human brain labelling technique can achieve similar or even better accuracy on mouse brain labelling except ventricular system. The volumes overlap between the automated segmentation and manual tracing does exceedingly well for large structures, small and thin structures like ventricle system are penalized for their size. Ma et al. (2005) pointed out that some ROIs require post hoc manual correction due to either smaller volume, or bigger registration error. The coefficients of errors are in the range of 0.04–0.12 for most ROIs, but at ventricle (The volume of whole ventricle account for around 0.3% of total brain volume), it is 0.29. Kovačevićet al. (2005) pointed out that in the average MR Atlas, the definition of larger structures was enhanced, but smaller structures were penalized due to a local blurring effect and their inherent variability. Lebenberg et al. (2010) found that some ROIs, such as the ventricles, didn't appear similar in the two images after registration. Since manual tracing of ventricle are labour-intensive and those automated approaches don't work well on small ROIs like ventricle, it would be helpful to develop an automated method to define ventricular ROIs on mouse brain.

Our overall goal in this study is to develop and assess a template-based adaptive threshold strategy for quantitative studies of mouse brain ventricular space volume enlargement and potentially shape analysis of the ventricle. Two approaches were used to validate this proposed technique. First, the automated approach was compared to the commonly used manual drawing method. Second, Monte Carlo simulation was used to evaluate the sensitivity and specificity of this approach in detecting computer-simulated ventricle volume changes introduced to the MR data in wildtype C57BL/6J laboratory mice.

Materials and methods

Animals and scans

All experiments were performed in compliance with relevant laws and institutional guidelines in Beijing University of Technology. Animal experiments were approved by local authorities. The study was carried out with six C57Bl/6J on a 7T horizontal bore Bruker Biospec 70/20 USR Preclinical MRM System using a 72 mm ID birdcage coil for excitation and a four-element phased array surface coil for reception. Animals were imaged in an animal bed restraint system with ear bars and bite bar for head fixation. The MRM data sets were obtained with a standard 3D Cartesian Fast Spin Echo (FSE) T2 weighting using following imaging parameters: TR = 1800 ms, TE = 40 ms, ETL = 8, FA = 90°, echo spacing = 10 ms. The matrix size was 300 × 170 × 96 for a field of view of 3.0 × 1.7 × 0.96 cm yielding an isotropic resolution of 100 microns.

For computational purposes, the voxel size of MRM was scaled to [1 1 1] millimetres (mm). While the scaling of voxel size is technical and will not affect the accuracy and feasibility of our approach, the units described below reflect scaled dimensions.

Template based adaptive Otsu approach

The images were first corrected for bias field using the Nonparametric Intensity Nonuniformity Normalization (N3) algorithm (Sled et al., 1998), with spline distance 25 mm, maximum iterations 1000, end tolerance 0.0001 and kernel FWHM 0.15. In addition, images were preprocessed to exclude the nonbrain tissues from the brain tissue (i.e., grey matter, white matter, and CSF) by an automated approach develop in our group. Images were then rigidly aligned to the template image using automated image registration software (http://bishopw.loni.ucla.edu/AIR5/, AIR). The template construction for mouse brain was detailed in (Lin et al., 2003). After rigid body alignment, images had the same orientation as the template image.

Our proposed approach (Fig. 1) consisted of four steps to optimally minimize the effects of spatial normalization inaccuracy. First, the ROIs were manually defined in template space using MRICro software (Neuropsychology Lab, Columbia SC, USA). This step was only performed once. Second, the template was normalized by discrete cosine functions (Ashburner & Friston, 1999) to individual mouse brain space and the deformation field was applied on the ROIs to transform them from the template space to an individual space using SPM2 software (http://www.fil.ion.ucl.ac.uk/spm). The spatial normalization settings were 2 mm FWHM for smoothness, and the default settings for the rest parameters. Third, the transformed ROIs were further refined to reduce the inclusion of nonventricular space using Otsu algorithm (Otsu, 1979) with adaptive threshold. The ventricular ROIs were first expanded by dilating operation. The voxel intensity within the dilated ROIs are then divided into two classes, high intensity class (the intensity of ventricle) and low intensity class (intensity of the surrounding grey and white matter). Smaller local windows have been divided along the ventricular boundary mask. The size of local window is fixed at 3*3*3 voxels. For each window, a threshold is obtained using the Otsu's method. The Otsu's threshold is applied to classify tissue into ventricular and nonventricular tissue in local window. If a voxel has been classified multiple times in different local windows (m times as ventricular tissues and n times as nonventricular tissues), then this voxel would be classified as ventricular tissue if m > n, or nonventricular tissue if m < n. Fourth, 3D morphological technique was applied to fill the ‘holes’ (misclassified as nonventricle) within the ventricle ROIs and discard the isolated nonventricle ‘islands’ (misclassified as ventricle) outside the ROIs.

Figure 1.

Flow chart of the image processing protocol.

Comparison with manual ROI approach

The ventricular volumes of six mice were traced manually by MRICro on its native space. To ensure that ROIs were consistent and sufficiently smooth for accurate computation of structural information, ROIs of ventricle were first outlined in coronal plane (posterior to anterior), and later corrections were performed in sagittal view (lateral to medial) to achieve satisfactory results from a neuroanatomical standpoint. Intrarater reliability of manually defined ROIs was also assessed by repeating the procedure twice by the same operator 2 weeks apart. And interrater reliability was accessed by repeating the procedure by two operators.

For evaluating the performance of our methods, we make use of the concept of Jaccard similarity index (JSI) and boundary concordance ratio (BCR). The JSI measures similarity between the volume of ROIs, and is defined as the volume of the intersection divided by the volume of the union of the ROIs. The BCR measures the degree of the boundary overlap, and is defined as the number of voxels common to both boundaries divided by the number of voxels in one of the boundaries. In calculating the BCR, two voxels, one from each of the two boundaries lying within 1.41 voxels of one another were accepted as an overlap.

Validation of the proposed approach using Monte Carlo simulated ventricle volume changes

In addition to its comparison to the manual approach, Monte-Carlo simulation was used to further validate the proposed approach in terms of detecting ventricular space volume changes between baseline and follow-up scans. The baseline scans from two randomly chosen MRM data (mouse 1 and mouse 2) were used for the Monte Carlo simulation procedure. The follow-up scans were simply the duplicate of the baseline scans with or without some artificial enlargement/shrinkage of the ventricle space volume in the presence or absence of the added noise. Both detection specificity and sensitivity of proposed method in detecting ventricle volume changes were assessed.

Specificity Misclassification of nonventricle voxels as ventricle ones (or vice versa) can occur simply due to the presence of noise alone (i.e. the ventricle space is identical on both the baseline and the follow-up scans, but the presence of noise leads to nonzero difference calculation). Consequently, the misclassification will result in inaccuracy of the estimated volume changes. The MRM data of Mouse 1 was used to evaluate the specificity in detecting brain ventricle volume change. To do that, the baseline and follow-up images were simply obtained by independently adding measurement noise to the original baseline image from Mouse 1. (Thus, the true volume change is known to be none, but both baseline and follow-up scans were with added noise independently.) The added noises were Gaussian with zero-mean and standard deviation (SD) proportional to the voxel intensities. Two noise levels were tested: high (100% SD of the voxel intensity) and moderate (50% SD of the voxel intensity). The ventricular volume changes between the baseline and the follow-up were then calculated using the proposed approach. The deviation of the volume change for its true value (0) is an indication of the specificity of the volume change detection relative to the noise.

Sensitivity The sensitivity of detecting volume change by the proposed approach can be examined by introducing additional varying amounts of ventricle to the follow-up image in the presence of added noise (while keeping the baseline ventricle un-touched) to determine the smallest detectable volume change.

Artificial ventricular ROIs of varying sizes were manually introduced along the original ventricle boundary. The procedure is as follows:

First, ventricle ROIs were carefully defined over the baseline image. Second, the mean and SD of the voxel intensities within this ROI were calculated. Third, Artificial ventricular ROIs of varying sizes were manually defined on the boundary of grey matter and ventricle. Fourth, the voxel intensities inside artificial ventricular ROIs were replaced by Gaussian random numbers with the ventricle mean and SD obtained in step 2. Note that ventricle ROIs on the baseline scan were unchanged during the process, and that the artificially introduced amounts of ventricle volume increase are known, which can be used in assessing the ventricle volume changes estimated by the proposed approach. In this study, the sizes of added ventricle ROI volumes ranged from 1.4% to 17.6%.

Results

Compared with manual approach

Compared to the intrarater JSI and BCR which were 0.967 ± 0.019 and 0.982 ± 0.009 and interrater JSI and BCR which were 0.962 ± 0.021 and 0.978 ± 0.015, we found high degree of overlapping (JSI 0.961 ± 0.020 and BCR 0.972 ± 0.014) between the ventricle ROIs computed by the proposed approach and the ROIs by the manual method. As seen in Figure 2 visually, the manually defined ventricle ROIs are not distinguishable from the one by the automated approach. Comparison between the automated and manual segmentation results showed excellent agreement. Automated measures were slightly smaller than hand tracing ones (mean difference 2.9%).

Figure 2.

From left column to right column is mouse MRM, the ROIs from this method, the ROIs from manual way and the overlapping of two ROIs (here red means the interaction of two ROIs and purple means the difference of two ROIs).

Monte Carlo simulations

The ventricle volume changes computed by the proposed approach when comparing the image pairs with only added measurement noise at different levels (50 trials) were –0.0%± 0.2% for moderate noise and –0.1%± 0.2% for high noise. The maximum detected false volume changes were 0.44% and 0.51%, respectively for the moderate noise and high noise level. Differences between moderate noise and high noise were evaluated using the paired sample two-tailed t-test. Values of p < 0.05 were considered to indicate statistical significance. There was little differences in the ROIs volume change across the two noise levels tested (p= 0.135 over 50 trials). We have also looked at the paired t-test p values for the hypothesis that the detected ROIs with noise is equal to zero (p= 0.072 for moderate noise and p= 0.001 for high noise). Although the detected false volume changes at high noise are statistically different from zero, their numerical values (–0.1%± 0.2%) are relatively invariantly much smaller than the smallest changes detectable in our simulations (see later), suggesting a satisfactory level of protecting false positive, and the performance robustness of this method in terms of the noise levels in the images.

The detected simulated ventricular ROI volume enlargement at two levels of noise, and six levels of simulated sizes (from 1.4% to 17.6% of ventricle volume) were presented in Table 1. For the all simulated ROIs, the detected volume changes were above the ROIs volume change detected in the absence of the true change (in simulation 1 earlier) and for all the noise levels examined (see more in Discussion section). Moreover, the detected volume changes were strongly and positively correlated with the amount of introduced ventricle volume, R= 0.999, 0.999, 0.999 for no noise, moderate noise and high noise.

Table 1.  Detected ventricle volume change for simulated data (50 trials).
Introduced atrophyDetected change (no noise)Detected change (moderate noise)Detected change (high noise)
 1.4%1.1%1.0%± 0.3%0.9%± 0.3%
 3.2%2.7%2.5%± 0.2%2.5%± 0.3%
 5.1%4.7%4.6%± 0.2%4.6%± 0.3%
 9.7%8.8%8.6%± 0.2%8.5%± 0.4%
15.3%13.9%13.7%± 0.4%13.6%± 0.3%
17.6%16.1%15.9%± 0.3%15.8%± 0.5%

Discussion

The proposed approach exhibited significant sensitivity in detecting simulated ventricular volume changes. Comparison with manual tracing approach confirmed excellent accuracy of the automated approach. When comparing the MRM pair simulated from a single scan, we found that the proposed method can detect the volume change as small as 1.4% of the whole ventricle volume. This is about 85% smaller than the detected ventricle volume enlargement in an empirical mannitol study (He et al., 2004).

The method takes advantages of the high contrast between the ventricle volume and its other surrounding brain tissues over T2 weighted MRM images. With the contrast quality available, its use for other brain regions where the contrast is not of such high quality, especially in light of the refinement of the template ROI using Otsu procedure, which may not as reliable as the ventricle ROI demonstrated in this study. On the other hand, we expect the proposed method will find its potential uses for non-T2-weighted images, as long as the interested region has high contrast comparing to its neighbouring volumes. The adequacy of our proposed approach was demonstrated for the wildtype C57BL/6J mice. For strains that are with greater variability of this type, further studies are needed to validate our approach.

In summary, this study evaluated the template based ROI approach as an automated and objective procedure for the noninvasive and high-resolution MRM studies for tracking longitudinal ventricular volume change in laboratory mice. The reliability and practicability of this template-based automated method make it a useful alternative for the manual determination of ROIs, which could be affected by intra- and interoperator differences. Results from this study suggest that this approach can be used to investigate the increases in ventricular volume in mouse studies. Further investigations are needed for its feasible use in detecting early changes and predicting onset of other symptoms in the research of transgenic mice and other putative animal models of AD.

Acknowledgements

This work was supported by grants from Natural Science Foundation of Beijing (3112005).

Ancillary