← AI Techniques Radar
Training & fine-tuning

Grouped Multi-Directional Mamba for lightweight 3D medical image segmentation

Use a grouped, multi-directional variant of Mamba inside an encoder for 3D segmentation: split input channels into G groups, flatten each group in multiple scan directions within one pass to aggregate multi-directional context, and apply Mamba only at deeper encoder stages to keep the model smaller.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

3D segmentation benefits from capturing global context, but full-width multi-directional scanning with Mamba can make the network large. Grouping channels and limiting where Mamba is used reduces model size while still providing multi-directional global context.

What changed

Replaced full-width, multi-directional Mamba scanning at every resolution with (1) grouped channel processing and (2) applying Mamba only in the deepest encoder stages.

Where it helps

  • 3D medical image segmentation
  • Memory/FLOPs-constrained segmentation models
  • Encoder-decoder architectures needing global context

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

The described approach’s effectiveness and cost trade-offs depend on the chosen grouping factor G, the set of scan directions, and where the encoder applies Mamba.

Related World Models papers

Sources