Automated Hippocampus Segmentation from CT Scans Using Hippocampus Dual Decoder Network (HDD-Net)

Motivation: Computed tomography (CT) imaging has been widely used in clinical practice, but its low contrast for the hippocampus makes it challenging to achieve accurate segmentation of the structure. Goal(s): In this study, we developed a novel deep learning (DL) model, termed HDD-Net, for automatic hippocampal segmentation from CT images. Approach: HDD-Net features dual, parallel decoders sensitizing hippocampal regions and their boundaries, respectively, and feature fusion module and cross loss that assist inter-decoder interactions. Results: HDD-Net outperforms existing U-Net-based models in terms of Dice and IOU. Statistical analyses revealed good agreement between CT- and MRI-derived hippocampal volumes. Impact: CT-based hippocampal segmentation via HDD-Net would be a promising alternative to conventional MRI-based procedures, and is expected to find a number of applications, for example, studies on Alzheimer's disease.