Multi-scale dual-channel feature embedding decoder for biomedical image segmentation.

Area of Science:

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

Background:

• Accurate biomedical image segmentation requires capturing both global context and local dependencies.
• Transformer-based models show promise but struggle with limited context capture and high computational complexity.
• Existing methods often require large datasets and face challenges in achieving optimal segmentation accuracy.

Purpose of the Study:

• To develop a novel deep learning model for enhanced biomedical image segmentation.
• To address the limitations of existing transformer-based models in capturing multi-scale contexts.
• To improve segmentation accuracy while managing computational complexity.

Main Methods:

• A novel multi-scale dual-channel decoder architecture is proposed, utilizing parallel convolutional encoders.
• The decoder employs hierarchical Attention-gated Swin Transformers for multi-scale feature embedding and capturing dependencies.
• A fine-tuning strategy is implemented to refine features and reduce over-segmentation.

Main Results:

• The proposed model demonstrates superior performance on public (LiTS, 3DIRCADb, spleen) and private datasets.
• Outperforms state-of-the-art models in liver tumor and spleen segmentation accuracy.
• Achieves competitive results with comparable computational cost.

Conclusions:

• The novel dual-channel decoder effectively integrates multi-scale features for efficient context representation.
• The fine-tuning strategy refines features, selecting only necessary information for improved segmentation.
• The proposed method achieves superior segmentation performance compared to existing state-of-the-art models.