Objective:

To develop an automated hippocampal segmentation tool using a novel network architecture.

Approach:

• Network Architecture: The Edge-aware Salient Context Fusion Refinement Network (ESCFR-Net) is built on a U-shaped encoder-decoder architecture, incorporating modules for feature enhancement and long-range spatial dependency modeling.
• Feature Enhancement: A Salient Feature Enhancer reduces background interference, while a Global Channel Context Attention (GCCA) module models long-range dependencies.
• Multi-scale Refinement: The Multi-scale Context Fusion Refinement Module (MCFRM) optimizes multi-scale feature utilization, and the Edge-Guided Refinement Attention (EGRA) module enhances edge and semantic features.

Key Findings:

• ESCFR-Net achieved a Dice coefficient of 0.9004, outperforming existing methods like SwinUNETR and PMFS-Net.
• Bilateral hippocampal volumes in patients with mild cognitive impairment (MCI) were significantly smaller than in healthy controls (p < 0.001).
• The total hippocampal volume had an area under the curve (AUC) of 0.927 for distinguishing healthy controls from MCI patients, with sensitivity of 90.11% and specificity of 83.52%.

Interpretation:

The study demonstrates that ESCFR-Net is a highly accurate tool for hippocampal segmentation.

Limitations:

• The study is based on a self-constructed dataset, which may limit generalizability.
• Further validation on larger, diverse datasets is necessary to confirm the findings.

Conclusion:

ESCFR-Net provides a robust automated segmentation method.