GMD-Mamba: Lightweight 3D Medical Image Segmentation Using a Grouped Multi-Directional Mamba
Organizations: School of Computing, University of Glasgow, Glasgow, United Kingdom · James Watt School of Engineering, University of Glasgow, Glasgow, United Kingdom · School of Computing, Data and Mathematical Sciences, University of Stirling, Stirling, United Kingdom
Abstract
Mamba has the capability to capture global context in linear time for 3D medical image segmentation. However, existing methods employ Mamba with full-width directional scans at every resolution, keeping the model large (64-138M parameters and high FLOPs). We propose Grouped Multi-Directional (GMD) Mamba, a lightweight network based on the GMD scan that divides input channels into groups and flattens each group in different directions in one pass. This gives multi-directional context at instead of , where is the number of scan directions. Additionally, we employ Mamba only at the deepest encoder stages, resulting in 19.2M parameters. On BraTS2023, it achieves the best average Dice of 92.0% and the lowest HD95 of 2.7 at 1/7 the parameters of SegMamba-V2. For model validation, we also ran experiments on the MSD Spleen dataset, achieving competitive performance. The implementation details and code are publicly available at https://github.com/2802990u/GMD-Mamba.
Figures & tables
| WT | TC | ET | Average | |||||||
| Family | Method | Params (M) | Dice | HD95 | Dice | HD95 | Dice | HD95 | Dice | HD95 |
| CNN | SegResNet [ 11 ] | 18.80 | 92.02 6.62 | 4.07 8.34 | 89.10 17.00 | 4.08 7.95 | 83.66 19.05 | 3.88 9.25 | 88.26 14.22 | 4.01 8.51 |
| nnU-Net [ 10 ] | 31.4 | 93.50 6.31 | 3.54 7.58 | 91.20 14.34 | 3.94 4.51 | 86.70 18.15 | 3.55 8.78 | 90.46 12.93 | 3.67 6.95 | |
| MedNeXt [ 12 ] | 17.6 | 92.41 7.01 | 4.98 8.77 | 87.75 19.19 | 4.67 6.75 | 83.96 19.02 | 4.51 9.90 | 88.04 15.07 | 4.72 8.47 | |
| Transformer | UNETR [ 8 ] | 93.0 | 92.19 9.83 | 6.17 11.77 | 86.39 21.60 | 5.29 11.93 | 84.48 21.35 | 5.03 11.82 | 87.68 17.59 | 5.49 11.84 |
| SwinUNETR [ 7 ] | 62.2 | 92.71 7.91 | 5.22 9.58 | 87.79 19.95 | 4.42 7.63 | 84.21 20.24 | 4.48 10.17 | 88.23 16.03 | 4.70 9.12 | |
| Configuration | Params (M) | Dice (%) |
|---|---|---|
| Backbone, no Mamba | 19.10 | 91.31 |
| + Bidirectional scan (deep) | 19.65 | 91.17 |
| + GMD scan (shallow+deep) | 19.24 | 89.85 |
| + GMD scan (deep) [ours] | 19.23 | 91.47 |
| Method | Params (M) | Dice (%) |
|---|---|---|
| SegResNet (backbone) | 18.80 | 79.4 |
| GMD-Mamba | 19.20 | 91.9 2.5 |
| 3D U-Net [ 1 ] | 4.9 | 81.95 |
| UNETR [ 8 ] | 93.0 | 86.42 |
| SegMamba [ 15 ] | 67.0 | 90.04 |
| SwinUNETR [ 7 ] | 62.0 | 91.26 |
Medical Image Segmentation
Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.
See all 839 papers