SA-UMamba: Spatial attention convolutional neural networks for medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Network (CNN) and Transformer models dominate medical image segmentation but have limitations.
• CNNs suffer from locality, while Transformers face quadratic complexity in attention computations.
• State-space models like Mamba offer linear complexity for global modeling but struggle with local feature extraction.

Purpose of the Study:

• To address the limitations of existing models in medical image segmentation.
• To propose a novel framework that effectively integrates global and local feature extraction.
• To enhance spatial expressiveness and segmentation performance in medical imaging.

Main Methods:

• Developed a novel Residual Spatial State-Space (RSSS) block integrating Mamba for global dependencies and Receptive Field Attention Convolution (RFAC) for local patterns.
• Introduced a residual adjust strategy for dynamic fusion of global and local information.
• Designed a U-shaped SA-UMamba segmentation framework utilizing the RSSS block for multi-scale spatial context capture.

Main Results:

• The proposed SA-UMamba framework demonstrated effective segmentation performance across multiple datasets.
• The RSSS block successfully enhanced spatial feature extraction by combining global and local representations.
• Experimental validation on Synapse, ISIC17, ISIC18, and CVC-ClinicDB datasets confirmed the framework's efficacy.

Conclusions:

• The SA-UMamba framework offers a promising advancement in medical image segmentation.
• Integrating global and local feature extraction via the RSSS block improves spatial expressiveness.
• The model shows significant potential for improving medical diagnosis and treatment planning.