High-throughput and accurate segmentation of lateral ventricles using combined single atlas propagation and tissue classification: Application to ADNI and IBSR data

Abstract

Ventricular volume measures have been proposed as a useful biomarker of AD progression as they distinguish disease from normality with a high effect size (Carmichael et al., 2007). In vivo quantification of lateral ventricular volumes is performed using automated techniques. In this work, we proposed a new high-throughput segmentation technique combining single-atlas propagation and tissue-type classification (GM, WM and CSF) for ventricular volume measurement in multicenter studies. The baseline scans of 90 subjects (30 HC, 30 MCI and 30 AD, age matched) from the ADNI database (http://adni.loni.ucla.edu) and 18 subjects from the IBSR database (http://www.cma.mgh.harvard.edu/ibsr/) were analysed. Lateral ventricular volumes were automatically estimated for all scans using the proposed segmentation strategy. An atlas including ventricular labels taken from a randomly selected reference AD patient was propagated to the subject space. In parallel, the subject's brain was classified into grey matter, white matter and CSF using a statistical technique based on Iterated Conditional Mode, Bayesian classification and Markov random field (Pachai et al., 2001). The propagated labels were then made consistent with tissue type using a classification-based nearest neighbour distance transform. The proposed scheme allowed error correction introduced by either the segmentation propagation technique or the atlas itself. The performance of the proposed technique was compared to a manual gold standard segmentation using the Dice similarity index (Dice) and D, the percent volume difference. Dice measured “anatomical precision” while D measured “volumetric accuracy”, which is the important parameter in terms of power for clinical trials. Main results (mean and SD) were summarized in Table 1. Anatomical precision was the same for ADNI and IBSR data, with high Dice value of 96.0 %. This result demonstrated the robustness of the proposed technique when processing image datasets of variable quality. Volumetric accuracy was very similar for both datasets, with low D values of 2.7 (2.5) % and 3.4 (2.0) % for ADNI and IBSR dataset, respectively. Dice and D values were very high and very low, respectively, as expected for an accurate segmentation technique. We have proposed a high-throughput automated technique for accurate segmentation of lateral ventricles in multicenter clinical trials.