Advancing Real-Time Polyp Detection in Colonoscopy Imaging: An Anchor-Free Deep Learning Framework with Adaptive...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp detection in colonoscopies is crucial for early colorectal cancer (CRC) prevention.
• Existing methods struggle with multi-scale polyp features, inefficient feature fusion, and reliance on complex, less generalizable anchor priors.

Purpose of the Study:

• To develop a real-time, accurate, and generalizable one-stage anchor-free detector for colonoscopic polyp detection.
• To overcome limitations in extracting multi-scale contextual cues and fusing multi-level features.

Main Methods:

• Introduced a Cross-Stage Pyramid Pooling module for efficient multi-scale context aggregation.
• Developed a Weighted Bidirectional Feature Pyramid Network for robust feature fusion.
• Implemented an anchor-free detection head with Scale-invariant Distance with Aspect-ratio IoU loss for direct point-to-boundary regression.

Main Results:

• Achieved state-of-the-art performance with 98.8% mAP@0.5 and 82.5% mAP@0.5:0.95.
• Real-time detection at 35.8 FPS on a GTX 1080-Ti GPU.
• Outperformed leading CNN and Transformer-based models on a large dataset of 103,469 colonoscopy frames.

Conclusions:

• The proposed anchor-free detector offers superior accuracy and efficiency for colonoscopic polyp detection.
• Demonstrates significant potential for clinical application in early CRC screening and prevention.