Bishop, Courtney A.Jenkinson, MarkDeclerck, JeromeMerhof, DoritDirk Bartz and Charl Botha and Joachim Hornegger and Raghu Machiraju and Alexander Wiebel and Bernhard Preim2014-01-292014-01-292010978-3-905674-28-62070-5786https://doi.org/10.2312/VCBM/VCBM10/017-024Hippocampal atrophy is a clinical biomarker of Alzheimer's disease (AD) and is implicated in many other neurological and psychiatric diseases. For this reason, there is much interest in the accurate, reproducible delineation of this region of interest (ROI) in structural MR images. Here, both current and novel MR hippocampal segmentation methods are presented and evaluated: Two versions of FMRIB's Integrated Registration and Segmentation Tool (FIRST and FIRSTv2), Freesurfer's Aseg (FS), Classifier Fusion (CF) and a Fast Marching approach (FMClose). Segmentation performance on two clinical datasets is assessed according to three common measures: Dice coefficient, false positive rate (FPR) and false negative rate (FNR). The first clinical dataset contains 9 normal controls (NC) and 8 highly-atrophied AD patients, whilst the second is a collection of 16 NC and 16 bipolar (BP) patients. Results show that CF outperforms all other methods on the BPSA data, whilst FIRST and FIRSTv2 perform best on the CMA data, with average Dice coefficients of 0.81+-0.01, 0.85+-0.00 and 0.85+-0.01, respectively. This work brings to light several strengths and weaknesses of the evaluated hippocampal segmentation methods, of utmost importance for robust and accurate segmentation in the presence of specific and substantial pathology.Categories and Subject Descriptors (according to ACM CCS): I.4.6 [Image processing and computer vision]: Segmentation - Edge and feature detection, Region growing, partitioning