CACFormer: A Hybrid CNN-Transformer Architecture Guided by Channel Attention for 3D Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Vision Transformers excel in natural image segmentation due to global modeling.
• Challenges exist in applying Vision Transformers to 3D medical imaging, including computational complexity and limited generalization on small datasets.

Purpose of the Study:

• To propose CACFormer, a hybrid CNN-Transformer architecture for 3D medical image segmentation.
• To address the limitations of Vision Transformers in medical imaging by improving efficiency and generalization.

Main Methods:

• Developed a hybrid CNN-Transformer architecture (CACFormer) incorporating a channel attention module.
• Designed a linear Transformer variant with a linear attention mechanism and tanh activation for focused segmentation.
• Guided feature fusion using adaptive channel weighting based on semantic contribution.

Main Results:

• CACFormer achieved a 89.71% Dice score on AMOS2022, outperforming 3D UX-Net and reducing inference time by 33.95%.
• On BraTS2021, CACFormer achieved a 90.20% Dice score with 28.54% fewer parameters and 15.22% faster inference than TransBTS.
• Demonstrated superior cross-dataset generalization, achieving 86.50% Dice on BraTS2021 when trained on BraTS2019, significantly outperforming TransBTS (47.90%).

Conclusions:

• CACFormer offers a favorable trade-off between performance and efficiency in 3D medical image segmentation.
• The proposed channel attention and linear Transformer variant contribute to robust segmentation and effective feature fusion.
• CACFormer shows significant potential for improving 3D medical image segmentation tasks and generalization across datasets.