TSCA-Net: Transformer based spatial-channel attention segmentation network for medical images

Area of Science:

• Medical imaging analysis
• Computer vision
• Deep learning

Background:

• Convolutional neural networks (CNNs) and Transformers excel in medical image segmentation.
• U-Net and similar encoder-decoder models are widely used.
• Challenges include low contrast, varied object scales, and complex backgrounds, hindering accurate segmentation.

Purpose of the Study:

• To propose a novel encoder-decoder architecture for medical image segmentation.
• To enhance feature extraction using spatial and channel attention modules based on Transformers.
• To improve segmentation performance by effectively learning details at different scales.

Main Methods:

• Developed an encoder-decoder architecture incorporating Transformer-based spatial and channel attention modules.
• Utilized these modules to extract global complementary information across different network layers.
• Designed a spatial and channel feature fusion block for the decoder to integrate multi-scale features.

Main Results:

• The proposed network combines CNN's local feature representation with Transformer's long-range dependency.
• Achieved superior performance compared to eight state-of-the-art methods on five public datasets.
• Demonstrated high Dice values (e.g., 80.23% on MoNuSeg, 93.56% on CHAOS-CT) and IoU values (e.g., 67.13% on MoNuSeg, 88.94% on CHAOS-CT).

Conclusions:

• The proposed Transformer-based attention mechanism effectively addresses challenges in medical image segmentation.
• The integrated approach enhances the learning of detailed features and improves segmentation accuracy.
• The method shows significant potential for various medical imaging applications.