DAMamba-UNet3D: Efficient 3D Medical Image Segmentation with Dynamic Adaptive Scan
A new parameter-efficient state-space model (SSM) architecture for 3D medical image segmentation, DAMamba-UNet3D, has been proposed. The hybrid encoder-decoder integrates tri-plane 3D Dynamic Adaptive Scan (DAS) blocks at encoder stages E2-E4 while retaining convolutions elsewhere, achieving approximately 5.3 million parameters. On the BraTS 2020 dataset with five-fold cross-validation, it achieves a mean Dice score of 0.815±0.013 in full-volume per-case evaluation, at roughly 13 times lower parameter cost than SegMamba (0.824). The work addresses limitations of convolutional U-Nets (local mixing only) and transformers (quadratic cost) by leveraging Mamba's O(n) global propagation. DAS, which learns data-dependent reordering before selective scan, is applied to medical imaging and extended to 3D volumes for the first time.
Key facts
- DAMamba-UNet3D is a parameter-efficient SSM-based U-Net for 3D medical image segmentation.
- It uses tri-plane 3D Dynamic Adaptive Scan (DAS) blocks at encoder stages E2-E4.
- The model has approximately 5.3 million parameters.
- On BraTS 2020, it achieves mean Dice 0.815±0.013.
- It is about 13 times more parameter-efficient than SegMamba (0.824).
- DAS learns data-dependent reordering before selective scan.
- This is the first application of DAS to medical imaging and 3D volumes.
- The architecture combines convolutions with SSM-based blocks.
Entities
—