Combining CNN and Self-attention-Free Transformer Using Local-Global Attention Fusion for Lung Cancer Segmentation

Segmentation of lung cancer lesions can assist physicians in diagnosing the condition. Transformer typically has high computational complexity, and how to effectively integrate Convolutional Neural Network (CNN) and Transformer for lung cancer segmentation remains an unresolved issue. Therefore, this paper introduces a novel hybrid CNN and Transformer network for lung cancer segmentation. The designed method consists of a CNN encoder, a Transformer encoder, a fusion module, and a CNN decoder. The CNN encoder is employed to extract local features, while the Transformer encoder captures global features. We propose a new self-attention-free Transformer block, which replaces the self-attention mechanism with a combination of shifting operations and depth-wise convolution to reduce computational complexity while maintaining performance. In addition, to enhance the interaction between CNN features and Transformer features, we introduce an innovative local-global attention fusion module, which performs feature fusion from both local and global views, resulting in better feature representations. The proposed approach is evaluated on a lung cancer segmentation dataset. Experimental results show that the proposed method achieves a Dice of 0.714 for lung cancer segmentation, outperforming several state-of-the-art methods.