Automated search of control points in surface-based morphometry

Highlights

• Radial scanning simplifies detection of white matter voxels in gray matter images.

• An automated search of control points is integrated in the FreeSurfer pipeline.

• Automatic search of control points reduces burden and errors in brain morphometry.

Abstract

Cortical surface-based morphometry is based on a semi-automated analysis of structural MRI images. In FreeSurfer, a widespread tool for surface-based analyses, a visual check of gray-white matter borders is followed by the manual placement of control points to drive the topological correction (editing) of segmented data. A novel algorithm combining radial sampling and machine learning is presented for the automated control point search (ACPS).

Four data sets with 3 T MRI structural images were used for ACPS validation, including raw data acquired twice in 36 healthy subjects and both raw and FreeSurfer preprocessed data of 125 healthy subjects from public databases. The unedited data from a subgroup of subjects were submitted to manual control point search and editing. The ACPS algorithm was trained on manual control points and tested on new (unseen) unedited data. Cortical thickness (CT) and fractal dimensionality (FD) were estimated in three data sets by reconstructing surfaces from both unedited and edited data, and the effects of editing were compared between manual and automated editing and versus no editing.

The ACPS-based editing improved the surface reconstructions similarly to manual editing. Compared to no editing, ACPS-based and manual editing significantly reduced CT and FD in consistent regions across different data sets. Despite the extra processing of control point driven reconstructions, CT and FD estimates were highly reproducible in almost all cortical regions, albeit some problematic regions (e.g. entorhinal cortex) may benefit from different editing.

The use of control points improves the surface reconstruction and the ACPS algorithm can automate their search reducing the burden of manual editing.

Introduction

Human brain structures are not static but exhibit morphological changes, possibly related to both neurodevelopment and neurodegeneration (Salat et al., 2004). The study of brain morphology and its characterization can be performed in-vivo with high-resolution magnetic resonance imaging (MRI) combined with computational neuroanatomy methods, allowing detailed analyses of structural data in large cohorts of normative and clinical subjects (Hutton et al., 2008). Particularly, voxel-based morphometry (VBM) (Ashburner and Friston, 2000) and surface-based morphometry (SBM) (Dale et al., 1999; Fischl et al., 1999) have been widely employed for detecting gray matter (GM) changes in distinct populations of subjects across the whole brain (Clarkson et al., 2011; Matsuda, 2013).

VBM and SBM generally operate on a series of three-dimensional 1-mm isometric T1-weighted (T1w) MRI images acquired at 1.5 T or higher magnetic fields. The advantage of SBM over VBM techniques, is that the former is theoretically quantitative, as it measures and compares absolute shapes and distances and not image intensities. However, SBM requires the reconstruction of at least two three-dimensional surfaces corresponding to internal and external boundaries of the cerebral cortex, and this is typically achieved via a semi-automated pipeline requiring repeated visual inspections and manual editing actions by one or more experienced operators (McCarthy et al., 2015; Oguz et al., 2008).

An SBM pipeline typically includes several automated operations on T1w data: skull stripping, motion correction (for longitudinal studies), B1 bias field correction, GM and white matter (WM) segmentation, reconstruction of cortical surface models, labeling of regions on the cortical surface, nonlinear registration of the cortical surface of an individual with a stereotaxic model and, eventually, a statistical analysis of group morphometric differences. Particularly, this pipeline produces (triangular) meshes from which several morphological and morphometric features can be extracted and analyzed. For example, a vertex-by-vertex estimate of the cortical thickness (CT), which is dependent on the local distance between the two surfaces, is often produced to map the local amount of GM (Desikan et al., 2010; Fischl, 2012; Salat et al., 2004). Moreover, as recently showed by (Madan and Kensinger, 2016), the fractal dimensionality (FD) (Mandelbrot, 1967) of the cortical surface mesh can also be estimated in predefined regions. A fractal (structure) is a geometrical object with special properties, self-similarity and scale invariance, that cannot be easily described (or approximated) by simple Euclidean shapes. Like structures in Euclidean spaces, fractals are also characterized by a dimensionality (FD) which measures the complexity of the structure. However, unlike Euclidean dimensionality, FD is a real number that captures tiny details (and may detect changes) in the shape of an object, e. g. the cortical surface or a portion of it, when the simpler modeling with Euclidean shapes is not sensitive enough (see, e.g. (Cook et al., 2008; Di Ieva et al., 2015, 2014; Free et al., 1996; Luders et al., 2004)). For example, from a stricter biological point of view, according to previous works (Im et al., 2006; Madan and Kensinger, 2016), FD can be of great interest as it is sensitive to age-related brain atrophy and other differences in GM structures that are not indexed by CT.

However, before extracting measures like CT or FD, user intervention is often needed to ensure adequate quality of the reconstructed cortical surfaces as obtained via an automated SBM pipeline. In fact, any topological errors on these surfaces could result in biased or less precise estimates of the desired feature.

Most current software packages implementing SBM analyses from MRI images provide interactive graphical interfaces and computational facilities to aid the manual intervention on the data (editing) in an attempt to iteratively improving the final quality of the reconstructed surfaces. Among these, FreeSurfer (FS) (Fischl, 2012), which is the most widely used tool for CT measurements in neuroimaging MRI research, allows users to visually identify and manually edit the topological defects of the output meshes. Once the defects are detected and corrected, the user has to re-run the surface reconstruction on edited data to obtain new corrected meshes of the surfaces.

Two of the most common errors are the inclusion of non-brain tissue in the GM and the incorrect segmentation of WM voxels. Both error types are often handled by editing the data after initial segmentation and reconstruction of the cortical surfaces. The first error type is due to the failure of the skull stripping procedure (Ségonne et al., 2004) that leads to including hyper intense voxels of the skull in the GM volume. This error can only be adjusted by erasing these voxels from the segmented GM volume. The second error type can be solved in two ways: (i) Isolated non-WM voxels included in the WM volume can be manually erased by the user; (ii) Compact WM areas wrongly included in the GM volume due to an incorrect intensity normalization of the WM (Dale et al., 1999) can be manually corrected in FS by positioning a certain number of points, called control points (CPs), that would ideally modify the WM border to avoid or reduce this error. More specifically, CPs are used in the intensity normalization step of the reconstruction pipeline as follows: Starting from a list of CPs (given as input), the editing consists in locally increasing the intensity of marked voxels in such a way that these are forced to fall within the WM volume. After the editing step, all WM voxels are re-scaled in such a way that the mean intensity over all WM voxels is set to the conventional level of 110 (Dale et al., 1999). However, the exact location and the optimal number of CPs required to successfully complete the correction is jointly dependent on the image quality and the operator's experience.

The impact of data editing procedures on SBM results has been recently investigated in two works: Iscan et al. (2015) explored the impact of manually cleaning GM volumes on the statistical power and reproducibility of CT analyses in FS, reporting a higher reproducibility for the manually edited data, compared to unedited data, only after each manual correction was visually approved or rejected based on an empirical process (Iscan et al., 2015). Popescu et al. (2016) compared CT measurements in FS with histologically measured CT data. In this case, the use of data editing led to a significant correlation between FS-derived and histologically derived CT values (Popescu et al., 2016). However, neither of these works used CPs for data editing.

Another study (McCarthy et al., 2015) explored the impact of CP-based manual editing on the estimated CT differences between 22Q11.2 deletion syndrome patients and controls using both 1.5 T and 3 T MRI data sets. In this case, the number of atlas regions with significant CT differences between unedited and edited data was higher on 3T compared to 1.5 T data, suggesting that CP based editing should have a stronger impact on higher resolution data (e.g. on data acquired at 3T or higher magnetic fields).

The aim of this work is to introduce a novel algorithm for an automated CP search (ACPS) that may possibly help (or replace) the user in the manual placement of CPs. The impact of the ACPS-based editing on the surface meshes as well as on the analysis and reproducibility of two major SBM measures (CT and FD) are investigated in four different datasets.