THE HIPPOCAMPAL BOUNDARY SHIFT INTEGRAL IS 70% MORE REPRODUCIBLE THAN OTHER ATROPHY ALGORITHMS

Keith S. Cover, Ronald van Schijndel, Adriaan Versteeg, Kelvin K. Leung, Emma Roxanne Mulder, Remko A. de Jong, Pieter Jelle Visser, Baptiste Grenier, Jérôme Revillard, David Manset, Alberto Redolfi, Bob W. van Dijk, Hugo Vrenken, Nick C. Fox, Giovanni Battista Frisoni, Frederik Barkhof

Research output: Contribution to conferenceAbstractAcademic

Abstract

Background: The reproducibility of algorithms for longitudinal hippocampal percentage volume change (PVC) measurement has direct repercussions on clinical trial design in diseases such as Alzheimer's (AD). The ADNI1 study provides an excellent data set on which to assess the back to back (BTB) reproducibility of hippocampal PVC as it acquired two identical MPRAGEs at each patient visit (Cover et al. Psychiatry Res. 2011;193:182-190). Methods: The BTB PVC for the baseline and 1 year MPRAGEs for 75 subjects (19 healthy controls, 38 MCI and 18 AD) was calculated using multiple algorithms. Algorithms included FreeSurfer/ ReconAll/longitudinal v5.3.0, FSL/FIRST v5.0.4, AdaBoost, manual, multiple- atlas propagation and segmentation (MAPS) and MAPS with the hippocampal boundary shift integral option (MAPS-HBSI) (Leung et al. NeuroImage 2010;51:1345-1359). The difference between the PVC's (non annualized) was calculated (BTBD). To compare the performance of any two algorithms, the absolute values of the BTBD for each subject were compared and the algorithm with the smallest absolute BTBDs was considered superior. Results: MAPS-HBSI had substantially better reproducibilities than all other algorithms. MAPS-HBSI had smaller absolute BTBD for the left hippocampus in 0.69 (FreeSurfer), 0.71 (FIRST), 0.72 (Ada- Boost), 0.68 (manual), 0.81 (MAPS) fraction of subjects, all which had a p-value smaller than 0.002. The median absolute value of the BTBD (MAVBTBD) were 2.36 (FreeSurfer), 2.19 (FIRST), 2.61 (AdaBoost), 2.75 (manual), 3.41 (MAPS) and 1.29 (MAPS-HBSI). Previous comparisons of algorithm performances have often been based on a particular disease, such as AD, that can introduce confounding factors including natural disease variation and measurement error in disease classification. For example, previous assessments of the BSI option, based on AD group size (Leung et al. 2010), is 47% - much less than the 164% found in this study. Conclusions: MAPS-HBSI has 70% better reproducibility than the nearest other hippocampal PVC algorithms due to the BSI option. The performance for BSI is 79% better than previously reported perhaps because of the disease independent assessment of the current study. The high reproducibility of MAPS-HBSI makes it much easier to compare the PVC of the left and right hippocampus for individual subjects and study the PVC variation in AD.
Original languageEnglish
PagesP390-P391
DOIs
Publication statusPublished - 6 Sep 2014

Cite this