Semi-automatic hippocampus delineation algorithm using surface fairing

Background: Manual hippocampus segmentation on structural magnetic resonance images (MRI) is labor intensive and time consuming. This paper presents a semi-automatic hippocampus segmentation method that decreases segmentation time without compromising accuracy. Methods: We present a method that reconstructs sparse delineations into full hippocampal surfaces meshes with a smooth surface reconstruction technique. From fully manual segmented hippocampi in ten subjects with about 20 slice contours, we simulated sparse delineations ranging from 4-10 contours to simulate decreased contouring time by at least half. We compared the original hippocampi with reconstructed hippocampi as well as automatic segmentations obtained from each subjects’ T1 weighted MRI using FSL-FIRST and FreeSurfer. We computed Dice overlap indices, percentage volume differences (PVD) and intra-correlation coefficients (ICC) with manual hippocampus segmentations. Results: For the hippocampi reconstructed from 4 to 10 contours, we obtained high mean dice overlaps, low mean PVDs and high ICCs in the range of 81(±0.03)-91(±0.01)%, 6.85(±5.33)-1.98(±1.63)% and 0.970–0.997 respectively. Reconstructed hippocampi agreed consistently better with manual segmentations than automatic segmentation methods, even when 5 contours were used. Conclusions: We were able to reconstruct hippocampi from a minimum number of contours and maintained high accuracy results that were consistently better than automatic methods. We next need to test this method on a larger scale and validate reproducibility and robustness.