An edge-aware salient context fusion and refinement network for hippocampal segmentation in MR images and its diagnostic value for mild cognitive impairment.

Background: Accurate assessment of hippocampal volume is of significant clinical value for the early diagnosis and disease monitoring of Alzheimer's disease (AD). However, automatic segmentation of the hippocampus in MR images remains challenging due to its elongated and irregular morphology, blurred boundaries, low contrast with surrounding tissues, and substantial inter-individual anatomical variability. Methods: We propose an Edge-aware Salient Context Fusion Refinement Network (ESCFR-Net). Built upon a classic U-shaped encoder-decoder architecture, the proposed network employs a Salient Feature Enhancer to suppress background interference and enhance weak feature responses of the hippocampus. A Global Channel Context Attention (GCCA) module is introduced to model long-range spatial dependencies, while a Multi-scale Context Fusion Refinement Module (MCFRM) improves the utilization of multi-scale features. Furthermore, an Edge-Guided Refinement Attention (EGRA) module synergistically enhances edge and semantic features to precisely delineate weak boundaries. Results: Experimental results on a self-constructed dataset comprising 225 3D-T1 MRI scans demonstrate that ESCFR-Net achieves a Dice coefficient of 0.9004, outperforming state-of-the-art methods such as SwinUNETR and PMFS-Net. Clinical association analysis, conducted on 91 healthy controls (HCs) and 91 patients with mild cognitive impairment (MCI), reveals that bilateral hippocampal volumes in MCI group are significantly smaller than those in HCs ( Conclusion: This study provides a highly accurate and robust automated hippocampal segmentation tool for early diagnosis, disease monitoring, and clinical decision-making in Alzheimer's disease.