COMMA: Coordinate-aware Modulated Mamba Network for 3D Dispersed Vessel Segmentation

Researchers from the Chinese Academy of Sciences and Beihang University have introduced COMMA (Coordinate-aware Modulated Mamba Network), a deep learning model designed to segment 3D dispersed vascular structures from medical images. The work, accepted by IEEE TIP, addresses a key limitation in current 3D medical segmentation models: the loss of spatial context due to patch-wise training. COMMA employs a dual-branch architecture—a global branch using a channel-compressed Mamba (ccMamba) block to encode entire images and capture long-range dependencies efficiently, and a local branch that processes cropped patches. A novel coordinate-aware modulated (CaM) block enhances interaction between these branches, allowing the local branch to retain spatial awareness.

The team also contributed the largest publicly available 3D vessel dataset, with 570 manually labeled cases covering two imaging modalities (likely CT and MRI) and five types of vascular tissues. Evaluated on six datasets, COMMA demonstrated superior performance over state-of-the-art methods, particularly in segmenting small vessels, while maintaining computational efficiency. Ablation studies confirmed the importance of both the ccMamba and CaM modules. The code and dataset will be open-sourced, enabling broader research into 3D medical image analysis.