Hippocampus Substructure Segmentation Using Morphological Vision Transformer Learning.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• The hippocampus is vital for memory and cognition.
• Accurate hippocampus segmentation is crucial for hippocampal avoidance during radiotherapy to minimize toxicity.
• Existing methods struggle with the hippocampus's complex shape and small size.

Approach:

• Developed Hippo-Net, a novel model for segmenting anterior and posterior hippocampus substructures from T1-weighted MRI images.
• Employs a two-part strategy: a localization model for hippocampus volume-of-interest detection and an end-to-end morphological vision transformer for substructure segmentation.
• Integrates learning-based morphological operators to enhance feature extraction and improve segmentation accuracy of hippocampus proper and subiculum.

Key Points:

• Achieved high accuracy in segmenting hippocampus substructures, with Dice Similarity Coefficients (DSC) of 0.900±0.029 for the hippocampus proper and 0.886±0.031 for the subiculum.
• Demonstrated low surface distance errors, with Mean Surface Distances (MSD) of 0.426±0.115mm and 0.401±0.100mm for the respective substructures.
• Utilized a dataset of 260 T1w MRI scans, with rigorous validation through five-fold cross-validation and hold-out testing.

Conclusions:

• Hippo-Net shows significant promise for automated hippocampus substructure delineation on T1w MRI.
• The model's accuracy and efficiency can potentially streamline clinical workflows in radiotherapy planning.
• This automated segmentation may reduce the workload for physicians, improving patient care.