Three-stage polyp segmentation network based on reverse attention feature purification with Pyramid Vision...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal polyps are precursors to colorectal cancer.
• Accurate polyp segmentation is crucial for early diagnosis and preventing missed cases.
• Current segmentation models struggle with polyp similarity to surrounding tissues.

Purpose of the Study:

• To develop an advanced AI model for accurate colorectal polyp segmentation.
• To overcome limitations of existing models in differentiating polyps from similar-looking tissues.
• To enhance the reliability of automated polyp detection in colonoscopy.

Main Methods:

• Proposed a novel three-stage network: Reverse Attention Feature Purification with Pyramid Vision Transformer (RAFPNet).
• Introduced a Multi-Scale Feature Aggregation (MSFA) module for initial saliency map generation.
• Implemented a Reverse Attention Feature Purification (RAFP) module to refine features using a UNet architecture.

Main Results:

• RAFPNet demonstrated superior performance compared to state-of-the-art models.
• The model achieved high accuracy across five static and three video polyp segmentation datasets.
• The iterative feedback UNet architecture effectively refined polyp saliency maps.

Conclusions:

• RAFPNet offers a significant advancement in automated polyp segmentation.
• The proposed method enhances the accuracy of identifying potential colorectal cancer precursors.
• This technology has the potential to reduce misdiagnoses and improve patient outcomes.