DensePVT-Net: Polyp segmentation network based on Transformer and boundary feature fusion

Polyp segmentation is essential for the early intervention of colorectal cancer. However, the morphological diversity of polyps and their indistinct boundaries with surrounding tissues pose significant challenges to accurate segmentation. This study proposes an innovative deep neural network model, termed DensePVT-Net, to tackle the difficulties associated with polyp segmentation in medical image processing, with the objective of enhancing the accuracy of early colorectal cancer diagnosis. The DensePVT-Net model directs the model to focus on polyp information by using the channel attention mechanism and the spatial attention mechanism. And a boundary feature fusion module is proposed, which uses a dense connection strategy to fuse high-frequency features with low-frequency features, thus further improving the model's ability to recognize boundary-ambiguous polyps. The experimental results show that the DensePVT-Net model outperforms the existing state-of-the-art models in several evaluation metrics, e.g., on the Kvasir-SEG dataset, the mIoU reaches 0.873 and mDice reaches 0.924; on the CVC-ClinicDB dataset, the mIoU reaches 0.895 and mDice reaches 0.940. ablation experiments further confirm the effectiveness of the boundary feature fusion module. The DensePVT-Net model proposed in this study provides an effective solution to improve the accuracy of polyp segmentation. It is of great significance for the early diagnosis and treatment of colorectal cancer.