P2TNet: Using Pyramid Pooling Transformer and Reverse Attention for Polyp Segmentation

Convolutional neural networks (CNNs) are currently the backbone of most polyp segmentation methods, thanks to their excellent ability to extract local features. However, this reliance on CNNs presents two main challenges: their limited receptive field restricts their capacity to model long‐range dependencies, and the often indistinct boundaries between polyps and surrounding mucosa can result in incomplete segmentation. To overcome these issues, we introduce P2TNet. This network employs a pyramid pooling transformer as its encoder for enhanced feature extraction. A cascaded fusion decoder is subsequently used, which aggregates high‐level features to generate a global feature map, thereby guiding the following reverse attention (RA) module. The RA module then refines the segmentation by extracting boundary details from low‐level features and reinforcing their connection to the polyp region. We evaluated P2TNet through extensive experiments on five widely‐used polyp datasets, comparing it against existing methods across six evaluation metrics. The results confirm that P2TNet substantially improves segmentation accuracy.