P2‐025: Accurate intracranial cavity volume estimation using multiatlas segmentation

Abstract

Intracranial cavity volume (ICV) is often used to correct brain structure volumes for inter-individual differences in magnetic resonance imaging (MRI) studies. ICV is frequently used as a correction factor, among other factors such as age, gender or genetics cross-sectional analyses of MRI biomarkers such as hippocampal and brain volumes in Alzheimer's Disease studies. Most existing skull stripping methods extract the brain surface, which can significantly differ from ICV in patients with cerebral atrophy, and is thus unsuitable for volume normalization. In this study, an automated method for accurate ICV extraction and estimation is proposed. Baseline scans of 276 MCI patients from the ADNI database (http://adni.loni.ucla.edu) were analyzed. HCV measurements were performed on three-dimensional T1 sequences. 20 additional MCI patients from the ADNI database were selected as atlases. For each atlas-subject, the ICV was semi-automatically segmented and manually quality-controlled by an experienced rater and used as “gold standard”. The ICV protocol included the area enclosing the brain and CSF, and the contour was delimited by the arachnoid when visible, and the dura mater when not. The atlases and their corresponding ICV labels were propagated to the patient using diffeomorphic demons registration. The 20 propagated ICV labels were then combined using the STAPLE algorithm to get the final patient ICV. The accuracy of the proposed technique was quantitatively compared to our “gold standard” and to the labels obtained with the cross-sectional and longitudinal versions of FreeSurfer (http://surfer.nmr.mgh.harvard.edu/), available on the ADNI database, using the Dice similarity index (“anatomical precision”) and the percent volume difference, D (“volumetric accuracy”), both expressed in percent. Results (mean and SD) are summarized in Table 1. The best Dice and D values were obtained with the proposed technique (98.4%-0.4%) when compared to the cross-sectional (95.0%-0.8%) or longitudinal (93.9%-1.2%) versions of FreeSurfer. A novel ICV quantification method was proposed. The multi-atlas strategy led to significantly better results than a single-atlas approach. The resulting accuracy was higher than cross-sectional and longitudinal implementations of FreeSurfer which appeared to be detecting a “brain envelope,” rather than the ICV, therefore missing a significant portion of external CSF.