A unified multi-scale probabilistic atlas of the human gray matter

FreeSurfer provides us some of the most used parcellation methods for segmenting human brain gray matter (GM) structures. These methods produce different output images which need to be gathered in order to generate a finer parcellation of the GM. The aim of this work is to create an unified multi-scale probabilistic atlas of the human GM by combining different cortical, subcortical and brainstem parcellation methods.

Two hundred and forty subjects from the Human Connectome Project (HCP) database (80 subjects, 40 males, for each age range: 22-25, 26-30 and 31-35 years old) were used to build the probabilistic atlas. The MRI acquisition protocols are described in Van Essen et al, 2012. The HCP Test-Retest dataset (45 subjects with baseline and follow-up acquisitions) were used to assess the performance of an atlas-based segmentation approach by using the developed atlas. These subjects were not used to build the atlas. The HCP database provides also the FreeSurfer outputs computed by a custom processing pipeline (Glasser et al, 2016). For each subject (285 baseline and 45 follow-up) the methods proposed by Iglesias et al (Iglesias et al, 2015a; Iglesias et al, 2015b) were used to segment the hippocampus subfields and brainstem. An atlas-based segmentation approach, by using the Advanced Normalization Tools (ANTS) and a thalamic nuclei atlas (Nadjenovska et al, 2018), was used to subvide the thalamus. Cortical and subcortical structures were obtained from FreeSurfer outputs. The method developed by Cammoun (Cammoun et al, 2012) was used to perform a multi-scale parcellation of the cortical surface, starting by the 34 regions (per hemisphere) described in Desikan et al, 2006. The resulting parcellations were gathered to obtain five individual GM parcellations for each subject.The 240 skull stripped T1 and T2 images were used to build a customized template by using the symmetric group-wise diffeomorphic normalization (SyN) method (Avants et al, 2010). This template was non-linearly warped to MNI space with the non-linear ICBM152 atlas (Evans et al, 2012) as target template. Both spatial transformations were used to resample the individual parcellations into the MNI space. The probability map for each GM structure, in each anatomical scale, was obtained by averaging across the subjects (Figure. 1).Statistics
An atlas-based segmentation approach, by using the developed atlas, was performed over the HCP Test-Retest.
For each GM structure, the Dice Similarity Index (DSI) (Dice et al, 1945) was used to assess the similarity between both segmentation approaches, the atlas-based segmentation and the FreeSurfer-based. The relative volume change between time points ((Volfup- Volbas/ Volbas)*100) for both parcellation approaches was also computed. Mann-Whitney U tests were performed to test for DSI differences between time points and to test for volume change differences between parcellation approaches. FDR correction was used to control for type I errors (q=0.05).

The DSI showed a mean overlap (for both time points) higher than 0.7 for all the structures, except for the hippocampus subfields, in scale 1 (Figure. 2A). Regarding the size of the regions, these findings could be considered as a fair match. DSI decreases with the decrease of the size of the structures, reaching mean values of 0.5 for scale 5 (Figure. 2C). The relative volume change showed a size-mismatch of ~5% between regions obtained from both segmentation methods (Figure. 2B). This value increases as the volume of the regions decreases (Figure. 2D). No significant differences in DSI or volume change were found. Thus, both segmentation methods behave the same for both time points.