DyMamba: dynamic Mamba for microscopy image semantic segmentation.

Area of Science:

• Computational Biology
• Image Analysis
• Deep Learning

Background:

• Accurate segmentation of cellular structures in microscopy is vital for biological research.
• Mamba architecture, based on State Space Models (SSMs), excels at modeling long-range dependencies but faces limitations in vision tasks due to fixed scanning strategies.
• Current Mamba scanning methods (raster, local) cause spatial discontinuities, hindering pixel-level segmentation effectiveness, especially for dense structures.

Purpose of the Study:

• To develop an advanced Mamba-based model for improved microscopy image segmentation.
• To address the limitations of static scanning strategies in Mamba architectures for vision tasks.
• To enhance the segmentation of fine details and small objects in biological images.

Main Methods:

• Propose DyMamba, a novel Mamba-based model incorporating a dynamic scanning strategy that adapts paths based on local image features and complexity.
• Introduce a local-aware module for pixel-level regional processing to improve detail and small object segmentation.
• Validate DyMamba on diverse microscopy image datasets at cell, organelle, and tissue scales.

Main Results:

• DyMamba demonstrates robust segmentation performance across various microscopy image types.
• Experiments on six datasets show DyMamba significantly outperforms state-of-the-art methods.
• Achieved an average improvement of 6.9% in mDice and 4.3% in mIoU compared to existing approaches.

Conclusions:

• DyMamba's dynamic scanning strategy effectively overcomes spatial discontinuities in Mamba-based segmentation.
• The model shows superior performance in segmenting complex microscopy images, including fine details and small objects.
• DyMamba offers a promising advancement for automated analysis in biological imaging research.