文献类型：期刊文献

英文题名：2.5D HAU-Net with gated spatial attention for automatic hippocampus segmentation in MRI

作者：Gong, Piqiang[1,2];Li, Xue[1,2];Lin, Dongmei[3];Ding, Xiaohan[1];Liu, Zhao[1];Chen, Fuming[1]

第一作者：Gong, Piqiang

通信作者：Chen, FM[1]

机构：[1]940th Hosp Joint Logist Support Force Chinese PLA, Dept Med Engn, Lanzhou 730050, Peoples R China;[2]Gansu Univ Tradit Chinese Med, Sch Med Informat Engn, Dept Biomed Engn, Lanzhou 730050, Peoples R China;[3]Lanzhou Univ Technol, Sch Microelect Ind Educ Integrat, Lanzhou 730050, Peoples R China

第一机构：940th Hosp Joint Logist Support Force Chinese PLA, Dept Med Engn, Lanzhou 730050, Peoples R China

通信机构：[1]corresponding author), 940th Hosp Joint Logist Support Force Chinese PLA, Dept Med Engn, Lanzhou 730050, Peoples R China.

年份：2026

卷号：431

外文期刊名：JOURNAL OF NEUROSCIENCE METHODS

收录：;Scopus(收录号：2-s2.0-105033255495);WOS:【SCI-EXPANDED(收录号:WOS:001723742600001)】；

基金：This research was funded by Natural Science Foundation of China (61901515, 62361038) , Natural Science Foundation of Gansu Province (22JR5RA002) .

语种：英文

外文关键词：Hippocampal segmentation; Computer-aided diagnosis; Spatial attention mechanisms; 2.5D data processing; U -Net architecture

摘要：Background: The hippocampus is a key brain region and biomarker for Alzheimer's disease (AD). Accurate automated hippocampal segmentation is essential for anatomical and pathological analysis. However, the gap between model complexity and available computational resources hampers the clinical deployment of computeraided diagnosis (CAD) systems, often resulting in limited accuracy and generalization. New method: This study introduces a 2.5D U-Net-based framework that integrates an attention mechanism for efficient and accurate MRI hippocampal segmentation. Three consecutive slices (anterior, middle, posterior) are stacked to form 2.5D input representations, enhancing spatial context. The proposed HAU-Net incorporates a gated spatial attention module to improve feature selectivity and robustness. A hybrid Dice-BCE loss is used to address class imbalance and accelerate convergence. Results: Experiments on the MSD Task04 Hippocampus and HarP datasets demonstrate strong performance, achieving Dice scores of 91.05% and 90.62%, respectively, with stable results across datasets. Comparison with existing methods: Compared with the baseline U-Net and other widely used segmentation models, the proposed 2.5D attention-enhanced network achieves higher Dice similarity coefficients and better generalization while maintaining computational efficiency suitable for practical use. Conclusions: The attention-guided 2.5D HAU-Net provides an effective, robust, and resource-efficient solution for automated hippocampal segmentation. Its strong performance and low computational demand support its potential for realworld clinical application and broader use in neuroscience and medical imaging.