MLAgg-UNet: Advancing Medical Image Segmentation With Efficient Transformer and Mamba-Inspired Multi-Scale Sequence.

Area of Science:

• Medical image analysis
• Deep learning for computer vision
• Biomedical engineering

Background:

• Transformers and state space sequence models (SSMs) show promise in biomedical image segmentation for long-range dependency capture.
• Traditional visual state space (VSS) models face challenges with image token compatibility and autoregressive assumptions.
• Transformer models, while not requiring autoregressive assumptions, incur high computational costs limiting channel-wise information use.

Purpose of the Study:

• To introduce a novel deep learning architecture, MLAgg-UNet, for improved biomedical image segmentation.
• To address limitations of existing methods, including computational cost and feature representation.
• To enhance the utilization of channel-wise information and multi-scale feature integration.

Main Methods:

• Proposed the Mamba-Like Aggregated UNet (MLAgg-UNet) incorporating Mamba-inspired mechanisms.
• Developed the Mamba-Like Aggregated Attention (MLAgg) block for efficient token dependency modeling.
• Introduced the Multi-Scale Mamba Module with Implicit Causality (MSMM) within skip connections for cross-scale feature optimization.

Main Results:

• MLAgg-UNet demonstrated superior performance across four benchmark datasets (AbdomenMRI, ACDC, BTCV, EndoVis17) covering MRI, CT, and endoscopy.
• Achieved significant Dice Similarity Coefficient (DSC) score improvements: 1.24% (AbdomenMRI), 0.20% (ACDC), 0.33% (BTCV), and 0.39% (EndoVis17).
• Outperformed state-of-the-art CNN, Transformer, and Mamba-based segmentation methods.

Conclusions:

• MLAgg-UNet effectively captures feature correlations and integrates complementary multi-scale information for robust medical image segmentation.
• The proposed architecture offers a balance between representational ability and computational efficiency.
• The study provides a promising solution for various medical imaging modalities, with publicly available implementation.