Purpose: Hippocampal segmentation can aid the diagnosis of neurodegenerative disorders and through avoidance can reduce cognitive deficits following whole brain radiotherapy. We trained a novel Hippo-Net using a mutual enhancement strategy for accurate hippocampus substructure segmentation from T1-weighted (T1w) MR images.
Methods: Two hundred and sixty T1w MRI cases from the Medical Segmentation Decathlon were used in this study. The Hippo-Net consists of a localization module, classification module and segmentation module. The training and the inference steps for a new case follow a feed-forward path in the Hippo-Net. The Hippo-Net is trained by several supervision mechanisms that optimize the learnable parameters to segment the anterior and posterior hippocampal substructures from a T1w MRI. In the Hippo-Net, a localization module is used to detect the substructure regions of interest (ROIs). Then, a classification module uses feature enhancement to improve mixed boundary accuracy. Finally, a segmentation module takes the feature map derived from localization and classification modules to derive the semantic segmentation of each substructure within its respective ROI. During inference, a new patient’s MRI is fed into the trained Hippo-Net to segment hippocampal substructures. The manual contours from the Medical Segmentation Decathlon dataset served as the ground truth to quantify the Hippo-Net segmentation accuracy.
Results: The Dice similarity coefficients (DSC) are 0.84±0.03 and 0.82±0.03 for the anterior and posterior hippocampal substructures, respectively. The volume difference (VD) is 0.13±0.11 cc and 0.19±0.11 cc for the anterior and posterior of hippocampus, respectively.
Conclusion: We have developed a novel Hippo-Net to accurately segment the anterior and posterior of hippocampus. Our results showed good performance in terms of DSC and VD between the segmentation result and the ground truth manual contours. The proposed method offers a strategic solution for hippocampal delineation that may significantly improve the efficiency of workflows in radiology and radiation oncology.