InnerEye-HS: a disease-agnostic clinical tool for hippocampal segmentation

The hippocampus is subject to atrophy in both Alzheimer's disease and temporal lobe epilepsy. Hippocampal volumes thus provide an early biomarker for these diseases. However, automated segmentation models typically lack robustness to disease-related changes in the hippocampus. In this work, we present the InnerEye hippocampal segmentation tool (InnerEye-HS). This deep learning tool was trained on MRI scans across the Alzheimer's disease spectrum, providing exposure to varying hippocampal size and topology. We validate the model against manually segmented hippocampi on both clinical dementia and epilepsy datasets collected in clinical settings and compare our model's performance to four other freely available tools (Automatic Segmentation of Hippocampal Subfields (ASHS), FreeSurfer, FastSurfer and HIPPOSEG). When compared to other freely available tools, the InnerEye-HS model provides the best Dice scores in our hospital dementia dataset (mean = 0.85 ± 0.02, P ≤ 0.0125), and InnerEye-HS and ASHS provided the best Dice scores in our epilepsy dataset (InnerEye-HS mean = 0.85 ± 0.02, ASHS mean = 0.84 ± 0.03). Furthermore, we found a high correlation (R² = 0.85) between hippocampal volumes extracted from ground-truth segmentations and those extracted from InnerEye-HS segmentations, demonstrating the model's ability to robustly segment the hippocampus throughout the disease time course. In summary, we present the InnerEye-HS model and demonstrate its advantage over currently available tools. These advantages highlight the clinical utility of our tool.