RBS-Net: Hippocampus segmentation using multi-layer feature learning with the region, boundary and structure loss.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• The hippocampus is a critical biomarker in Alzheimer's disease (AD) research.
• Accurate hippocampus segmentation in Magnetic Resonance Imaging (MRI) is vital for clinical research on brain disorders.
• Current deep learning methods, while efficient, suffer from information loss during pooling and imprecise boundary segmentation.

Purpose of the Study:

• To develop an advanced deep learning model for accurate hippocampus segmentation.
• To overcome limitations of existing methods, specifically information loss and coarse boundary segmentation.
• To improve the performance of hippocampus segmentation for enhanced AD research.

Main Methods:

• Proposed Region-Boundary and Structure Net (RBS-Net) with a primary and auxiliary network.
• Primary net uses distance maps for boundary supervision and multi-layer feature learning to mitigate pooling-induced information loss.
• Auxiliary net focuses on structural similarity using a multi-layer feature learning module to refine encoder features.

Main Results:

• RBS-Net achieved an average Dice score of 89.76% on the HarP hippocampus dataset, outperforming state-of-the-art methods.
• Demonstrated superior performance in few-shot learning scenarios compared to existing deep learning approaches.
• Visual results showed significant improvements in boundary definition and segmentation of detailed regions.

Conclusions:

• RBS-Net effectively addresses information loss and boundary segmentation issues in hippocampus segmentation.
• The proposed model offers enhanced accuracy and robustness, particularly in limited data scenarios.
• RBS-Net shows promise for advancing clinical research in neurological disorders like Alzheimer's disease.