BAformer: Boundary Adaptive Transformer for Medical Image Segmentation

Junlong Cheng, Jilong Chen, Lei Jiang, Chenrui Gao, Junren Chen, Ruoyu Li, Min Zhu

Abstract: Accurate automatic segmentation of medical images has long faced challenges such as significant lesion scale variations and blurred boundaries. This study proposes a Boundary-Adaptive Transformer (BAformer) specifically designed for 2D medical image segmentation. BAformer employs a hierarchical architecture that focuses on boundary-related features across varying resolutions. Additionally, we introduce a Dense Feed-forward Network (DenseFFN) to reuse features, enabling each layer to accumulate information from previous layers. Finally, we extend the benefits of the dense network to the decoder, balancing parameter efficiency and performance. Extensive experiments demonstrate that BAformer achieves state-of-the-art performance on several challenging medical segmentation tasks. Furthermore, BAformer can seamlessly integrate with existing segmentation networks, demonstrating its versatility and effectiveness.