AESC-TransUnet:Attention Enhanced Selective Channel Transformer U-Net for Medical Image Segmentation

Deep learning has greatly advanced medical image segmentation, especially with the integration of U-Net and Transformer architectures. However, challenges remain in clinical practice, such as low resolution, blurred edges, and difficulty in accurately segmenting small lesions. Existing methods, including various TransUNet variants, struggle with these issues. For instance, while TransUNet leverages the Transformer to capture global context, it suffers from the inability to preserve fine-grained local details. Similarly, Swin-Unet, which uses the Swin Transformer, excels in global feature extraction but often loses precision in complex backgrounds and small structures. The classic U-Net model, despite its strong performance in extracting local features, struggles with segmentation accuracy in low-contrast or complex areas due to its upsampling and downsampling processes.To address these limitations, we propose AESC-TransUNet, a novel network combining Efficient Selective Channel Attention (ESCA) and Convolution-Transformer fusion (ConvFormer). ESCA improves feature representation by using parallel processing and selective channel enhancement, effectively addressing the low resolution and blurred edges problem while preserving local details. ConvFormer optimizes the integration of both global and local information by combining convolution with self-attention mechanisms, reducing positional loss during upsampling and improving edge segmentation. Experimental results show that AESC-TransUNet significantly outperforms existing methods, achieving higher segmentation accuracy, particularly for small lesions and complex structures.