A Robust Hippocampus Segmentation Method for Brain MRI Based on Multi-View Fusion and Edge Optimization

Abstract

: With the rapid aging of the global population, the prevalence of neurodegenerative disorders such as Alzheimer's disease (AD) is increasing, posing a substantial burden on healthcare systems and society. Hippocampal atrophy is recognized as one of the earliest and most reliable biomarkers for AD, making its accurate and automated segmentation from Magnetic Resonance Imaging (MRI) scans a critical task for early diagnosis and disease monitoring. This paper proposes a comprehensive and robust method for automated hippocampus segmentation, leveraging a synergistic combination of multi-view information and targeted boundary refinement. The process of this article begins with meticulous data preprocessing, including the extraction of a standardized Region of Interest (ROI) and skull stripping, to enhance model focus. Then, developing an improved U-Net architecture, which is significantly enhanced by increasing network depth and integrating three key modules: a deep supervision mechanism (DSN) to combat gradient vanishing and improve feature learning, a Generative Adversarial Network (GAN) framework to enforce structural plausibility, and a Convolutional Block Attention Mechanism (CBAM) to refine feature representation. To fully exploit the rich spatial information available in volumetric MRI data. The article applies this improved architecture to both 2D and 3D U-Net models. A hierarchical multi-view fusion strategy is then employed to integrate the predictions from 2D anatomical views (axial, coronal, sagittal) and the 3D volumetric view, capitalizing on their complementary strengths. Finally, a dedicated two-stage edge optimization technique is introduced to specifically refine the segmentation boundaries, which are often ambiguous. Extensive experiments were conducted on a private dataset. The results demonstrate the efficacy of the approach, with the final model achieving a Dice Similarity Coefficient of 0.925, a Jaccard Index of 0.860, and a 95% Hausdorff Distance of 1.000 mm on the test set. This work presents an efficient, accurate, and highly promising solution for automated hippocampus segmentation, with strong potential for clinical translation.