Reproduce “Polyp-PVT: Polyp Segmentation with Pyramid Vision Transformers” on OpenResearch

About this paper

Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, Ling Shao

Researchers developed Polyp-PVT, a polyp segmentation framework integrating a Pyramid Vision Transformer encoder with specialized fusion and identification modules. The framework demonstrates enhanced accuracy and generalization across various clinical datasets, achieving a mean Dice score of 0.917 on Kvasir-SEG and 0.808 on the unseen ColonDB dataset.

Key insights

• Pyramid Vision Transformers, with their multi-head self-attention, provide more robust and powerful feature extraction for polyp segmentation compared to CNNs, especially when dealing with input disturbances.
• Separately addressing high-level semantic integration via a cascaded module (CFM) and low-level detail extraction using an attention-based module (CIM) is effective for handling polyp camouflage and refining features.
• Graph-convolution-based similarity aggregation (SAM) facilitates the effective fusion of multi-level features, enabling the transfer of fine-grained information across the entire polyp area for improved boundary delineation.

Summary by alphaXiv.