DBMamba-UNet: dual-branch mamba in vision for 3D hippocampus image segmentation

Hippocampal segmentation is a crucial task in medical image processing, playing a significant role in the diagnosis and research of neurodegenerative diseases such as Alzheimer’s disease. The hippocampus, located in the medial temporal lobe of the human brain, is a bilayer gray matter structure that protrudes into the temporal horn of the lateral ventricle and occupies its medial base. Due to its small size, irregular shape, and indistinct boundaries, accurate segmentation of the hippocampus poses considerable challenges. Currently, most hippocampal segmentation methods rely on magnetic resonance imaging (MRI). However, using structural MRI alone often fails to achieve precise segmentation. To enhance segmentation accuracy, researchers have adopted various approaches. Among these, deep learning–based methods have gained significant attention for their powerful feature extraction and classification capabilities. Nevertheless, these methods tend to confuse edge regions during segmentation, reducing overall accuracy. Transformer-based models, known for capturing long-range dependencies effectively, have also been explored, but their extensive parameter requirements often limit practical applications in segmentation tasks. To address these challenges, we propose a dual-branch Mamba-inspired U-Net (DBMamba-UNet). The dual-branch structure preserves rich detail information, particularly edge features, at high resolutions. Features of different scales from the main and auxiliary branches are fused using the FMamba module before being passed to the upsampling layer. The Mamba-inspired visual architecture, utilized as both encoder and decoder, combines the advantages of transformers and Mamba structures. This design reduces the computational resources required for segmentation tasks while enabling parallel processing. Compared with state-of-the-art algorithms, DBMamba-UNet achieves superior segmentation performance with significantly lower computational costs, delivering better results across various evaluation metrics.