Know your orientation: A viewpoint-aware framework for polyp segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate polyp segmentation in endoscopic images is crucial for early colorectal cancer diagnosis.
• Existing segmentation models struggle with viewpoint variations and indistinct polyp boundaries due to similar surrounding tissues.

Purpose of the Study:

• To propose a novel viewpoint-aware framework, VANet, for precise polyp segmentation in colonoscopy images.
• To enhance the robustness and accuracy of polyp segmentation algorithms in real-world clinical scenarios.

Main Methods:

• Developed VANet, a framework incorporating a viewpoint classification process to localize polyps using class activation maps.
• Introduced a viewpoint-aware Transformer (VAFormer) to improve polyp feature representation despite viewpoint changes.
• Implemented a boundary-aware Transformer (BAFormer) to refine polyp boundary segmentation by focusing on uncertain regions.

Main Results:

• VANet demonstrated state-of-the-art performance across seven public datasets and six evaluation metrics.
• The proposed VAFormer and BAFormer modules effectively addressed viewpoint variations and boundary ambiguity.
• The framework achieved significant improvements in polyp segmentation accuracy and robustness.

Conclusions:

• VANet offers a robust and accurate solution for automatic polyp segmentation in colonoscopy.
• The viewpoint-aware and boundary-aware approaches are effective in handling challenges posed by real-world endoscopic imaging.
• This method has the potential to significantly aid in the early detection of colorectal cancer.