Hippocampus substructure segmentation using morphological vision transformer learning.

Area of Science:

• Medical imaging analysis
• Neuroscience
• Artificial intelligence in medicine

Background:

• The hippocampus is critical for memory and cognition.
• Accurate hippocampus segmentation is essential for radiotherapy planning to minimize toxicity.
• Current segmentation methods struggle with the hippocampus's complex shape and small size.

Purpose of the Study:

• To develop a novel model, Hippo-Net, for accurate segmentation of hippocampus substructures.
• To improve the delineation of anterior and posterior hippocampus regions from T1-weighted MRI.
• To enhance clinical workflows in radiotherapy planning through automated segmentation.

Main Methods:

• Developed Hippo-Net, a cascaded model comprising a localization network and a morphological vision transformer.
• Integrated learning-based morphological operators into the vision transformer to enhance feature extraction.
• Utilized 260 T1-weighted MRI datasets from the medical segmentation decathlon dataset for training and validation.

Main Results:

• Achieved high accuracy in segmenting hippocampus substructures, with Dice similarity coefficients of 0.900 ± 0.029 (hippocampus proper) and 0.886 ± 0.031 (subiculum).
• Demonstrated low mean surface distances of 0.426 ± 0.115 mm and 0.401 ± 0.100 mm for the respective substructures.
• The model effectively separates the hippocampus into its distinct anterior and posterior regions.

Conclusions:

• Hippo-Net shows significant promise for automated hippocampus substructure delineation on T1w MRI.
• The proposed method can potentially streamline clinical workflows and reduce physician workload.
• Accurate segmentation of hippocampus substructures is crucial for personalized radiotherapy treatment planning.