Cortical thickness estimation accuracy with three automated algorithms

Abstract

Cortical thickness is an innovative promising marker for Alzheimer disease. So far the cortical measurement of atrophy has been largely used exclusively in research group analysis. Thus, a comparison of the gray matter loss of the three most prominent reconstruction algorithms (CIVET, FREESURFER and RIC-BRAINVISA) is needed for planning future clinical exploitations. The aim of the present study is to compare the accuracy of the three algorithms testing cortical thickness maps of the MP-RAGE scans of the US-ADNI study. We selected ADNI subjects who had MP-RAGE from month 0 to month 36. More then 7'000 ADNI scans were processed in the current analysis and processed using the neuGRID/gLite infrastructure (www.neugrid.eu). The cortical thickness maps were computed for each subject scan using CIVET (v1.1.9), FREESURFER (v5.0.0) and RIC-Brainvisa (v3.1-v3.2.1). The three tools produce data with heterogeneous formats. To address this issue a visualization tool compatible with all three was used to read the result files and display cortical surface maps in the same coordinate space and color reference system (figure 1). For every scan the cortical thickness maps generated were compared using both descriptive statistics and variance homogeneity test. The cortical reconstruction is difficult because of artifacts such as noise, partial volume effects, and intensity inhomogeneities that cause inaccuracies in cortical thickness measurement. From preliminary observations, CIVET shows the more stable and consistent thickness estimation trend compared to FREESURFER and RIC-BrainVISA (figure 2). Both CIVET and FREESURFER, unlike RIC-Brainvisa, show a progressive cortical thinning starting from Healthy Elderly to MCI and continuing to AD subjects. The three algorithms show a large difference in the mean cortical thickness calculated within the same study population (Normal, MCI and AD) as well as for different magnetic field strengths (1.5T and 3T). The homogeneity of variance in cortical thickness between the three tools is significantly different (Levene's p-value = 0.0035). These results demonstrate how the tools available today for the quantification of cortical thickness behave differently. Standard Operating Procedures might be established to enhance the applicability of the cortical thinning through a standardization of the analysis methods and developing a robust protocol for automated assessment. Maps of mean cortical thickness in a subset of 18 ADNI subjects (6 AD , 6 MCI and 6 normal elderly) obtained with CIVET, FREESURFER and RIC-BrainVISA, and displayed with the same visualisation tool. Trend of the cortical thickness average for each subject carried out with CIVET, FREESURFER and RIC-BrainVisa . The homogeneity of varience in cortical thickness between the three tools is significantly different (Levene's test p <0.05).