AVPDN: learning motion-robust and scale-adaptive representations for polyp detection in dynamic colonoscopy frames.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate polyp detection is crucial for early colorectal cancer diagnosis.
• Colonoscopy videos offer rich data but suffer from motion artifacts degrading image quality and increasing false positives.
• Existing methods struggle with dynamic imaging challenges like blur and scale variation.

Purpose of the Study:

• To develop a robust framework for multi-scale polyp detection in dynamic colonoscopy imagery.
• To address frame-level artifacts caused by rapid camera movement during colonoscopy.
• To improve the accuracy and reliability of polyp detection in challenging video data.

Main Methods:

• Proposed the Adaptive Video Polyp Detection Network (AVPDN) featuring Adaptive Feature Interaction and Augmentation (AFIA) and Scale-Aware Context Integration (SACI) modules.
• AFIA utilizes a dual-branch architecture with dense and sparse self-attention for enhanced feature representation.
• SACI employs dilated convolutions for multi-scale contextual information capture and denoising.

Main Results:

• AVPDN demonstrated robust performance in detecting polyps from motion-affected colonoscopy frames.
• The proposed method achieved state-of-the-art results on challenging public benchmarks.
• Experiments confirmed the effectiveness and generalization capability of the AVPDN framework.

Conclusions:

• The AVPDN framework effectively overcomes challenges posed by motion artifacts in colonoscopy videos.
• This network significantly enhances polyp detection accuracy, aiding in earlier colorectal cancer diagnosis.
• The method shows strong potential for clinical application in real-time colonoscopy analysis.