P3‐206: ACCURACY OF BMAS HIPPOCAMPUS SEGMENTATION USING THE HARMONIZED HIPPOCAMPAL PROTOCOL

Abstract

Hippocampal volume (HCV) is a key biomarker for both improving subject selection and monitoring treatment efficacy in Alzheimer's Disease (AD) studies. The main goal of the Hippocampal Protocol (http://www.hippocampal-protocol.net/) was to harmonize the multiple hippocampal protocols and define a standard protocol. Additionally, expert hippocampus segmentations for 135 ADNI cases will be made available (100 available as of January 2014).The purpose of this work was to evaluate the accuracy of the BioClinica Multi-Atlas Segmentation (BMAS) algorithm with respect to the hippocampal structure as defined by the Harmonized Hippocampal Protocol. 100 baseline three-dimensional (3D) T1 sequences from the Harmonized Hippocampal Protocol and their corresponding hippocampal segmentations were used. This dataset included 29 Normal Controls (NC), 34 Mild Cognitive Impairment (MCI) and 37 AD subjects from ADNI (http://adni.loni.ucla.edu) oriented along the AC-PC line, all used as atlases for the BMAS algorithm. Hippocampal segmentations were automatically computed with the BMAS algorithm in a “leave-one out” fashion. Segmentation accuracy was assessed using Dice and relative volume error (RVE) with the provided hippocampal segmentations as reference. Pearson's correlation between computed and reference volumes was also calculated. Finally cohort separation power was measured with the area under the ROC curve (AUC). As detailed in Tables 1 and 2, accuracy of the proposed BMAS segmentation was high (Dice= 86.59% ± 1.80) and total hippocampal volumes (right + left) were very close to the reference volumes (RVE=3.96% ± 3.31). BMAS hippocampus segmentations were highly correlated to the reference hippocampus segmentations (r=0.96 with p<0.001). Similar ability for subject cohort separation was observed for the reference and BMAS hippocampus segmentations (see Table 3). The implementation of BMAS based on the new 100 atlases from the hippocampal protocol was straightforward and successful. BMAS results were highly correlated with those of the harmonized hippocampal protocol and its ability to separate subject cohorts was also very similar. The newly released library of atlases generated based on the hippocampal protocol can be used as the new gold standard for benchmarking various segmentation methods in terms of accuracy and reproducibility and as the reference library for processing MRI scans in prospective AD trials.