ATLAS-BASED SEGMENTATION OF ULTRAHIGH-RESOLUTION STRUCTURAL MR HEAD IMAGES ACQUIRED AT 7 TESLA

Abstract

Purpose: The purpose of this work was to find out how the existing brain atlases and segmentation algorithms perform when applied to ultrahigh-resolution MR brain images, acquired with a 7-Tesla scanner. Also to make adaptations to deal with the potential challenges and evaluate the quality of the anatomical segmentations of the 7- Tesla images. Materials: A dataset of MR brain images with various resolutions (1mm, 500 m, 250 m non averaged & 250 m averaged) shared by Lüsebrink et al. from the 7 Tesla scanner in Magdeburg was used. Methods: Two atlas-based anatomical image segmentation algorithms were applied: Pincram for brain extraction and MAPER for labelling multiple brain regions. The resulting brain masks and label maps were assessed qualitatively and quantitatively. Visual evaluation of the quality of the segmentations was made by the auther and external experts. To quantify the consistency of segmentations at the highest resolution, the Jaccard overlap coefficient were calculated. Shape base averaging (SBA) has been implemented on the MAPER-segmented atlases and applied to a 500 m resolution image to improve the appearance of the segmentation. It was then compared to Vote Rule decision Fusion (VRF) that is the standard method of fusing atlas labels in MAPER. Conclusion: MAPER and Pincram work on brain images obtained with a 7-Tesla scanner even though the algorithms have been designed for and validated on 1.5 and 3 Tesla. The data size at the highest resolution exceed available computational resources, therefore images had to be downsampled to 500 m. The segment boundaries were smoother with SBA than with VRF and they got more pleasant to look at. Some boundaries do get misplaced, so the volume estimation of the structures might not be better than with VRF.