SegMamba-V2: Long-Range Sequential Modeling Mamba for General 3-D Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Transformer models excel at 3D medical image segmentation but are computationally intensive.
• State Space Models (SSMs), like Mamba, show promise for long-range dependency modeling but are underexplored in 3D medical imaging.

Purpose of the Study:

• To introduce SegMamba-V2, a novel 3D medical image segmentation model designed to capture long-range dependencies efficiently.
• To enhance feature representation by extending global dependency modeling to three orthogonal planes.

Main Methods:

• Developed a hierarchical scale downsampling strategy to improve receptive field and reduce information loss.
• Designed a novel tri-orientated spatial Mamba block for enhanced global dependency modeling.
• Introduced and utilized the CRC-2000 dataset for fine-grained 3D colorectal cancer segmentation.

Main Results:

• SegMamba-V2 demonstrated superior performance compared to state-of-the-art methods across multiple 3D medical image segmentation datasets.
• The model showed effectiveness on diverse modalities, organs, and segmentation targets.
• The proposed hierarchical downsampling and tri-orientated Mamba block significantly improved feature representation.

Conclusions:

• SegMamba-V2 offers a computationally efficient and highly effective solution for 3D medical image segmentation.
• The model's architecture, incorporating Mamba, proves versatile and powerful for capturing long-range dependencies in volumetric data.
• The study highlights the potential of Mamba-based approaches for advancing 3D medical image analysis.