Fusion of Multi-Size Candidate Regions Enhances Two-Stage Hippocampus Segmentation

Abstract

The hippocampus plays an important role in the memory and cognition abilities of humans. Precise three-dimensional (3D) segmentation of the hippocampus from magnetic resonance imaging scans is of great importance in the diagnosis of neurological diseases. Conventional automatic segmentation methods poorly achieve satisfactory performance because of the irregular shape and small volume of the hippocampus. We propose a novel two-stage segmentation method, which includes a localization stage and a segmentation stage, to handle the task of the 3D segmentation of the hippocampus. In the localization stage, a novel strategy for localizing multi-size candidate regions was developed to improve the sample balance for the 3D segmentation task. In the segmentation stage, a method which fuses the multi-size candidate regions was proposed to improve the accuracy in predicting the hippocampal boundary, after which we aggregated the segmentation results from three orthogonal views to further improve the performance. Quantitative evaluation was performed on the Alzheimer's Disease Neuroimaging Initiative dataset. The experimental results achieved Dice similarity coefficients of 92.48 ± 0.61% and 92.90 ± 0.51% for the left and right hippocampus, respectively, outperforming state-of-the-art studies in hippocampus segmentation tasks.