A robust real-time deep learning based automatic polyp detection system.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer (CRC) is a leading global cancer, with colonoscopy as the gold standard for screening and polyp removal.
• Computer-aided detection (CAD) systems for colonoscopy have limitations in sensitivity and specificity.
• Deep learning offers improved polyp detection but struggles with real-time performance.

Purpose of the Study:

• To develop a novel deep learning architecture for real-time polyp detection in colonoscopy videos.
• To enhance the accuracy and efficiency of existing YOLOv4 algorithms for this task.

Main Methods:

• Modified the YOLOv4 algorithm by integrating Cross Stage Partial Networks (CSPNet) and replacing Mish activation with Leaky ReLU.
• Substituted Distance Intersection over Union (DIoU) loss with Complete Intersection over Union (CIoU) loss.
• Utilized data augmentation, ensemble learning, and NVIDIA TensorRT for optimization, validated on public datasets (ETIS-LARIB, CVC-ColonDB).

Main Results:

• The proposed method achieved state-of-the-art real-time detection accuracy.
• On the ETIS-LARIB dataset, achieved 91.62% precision, 82.55% recall, and 86.85% F1-score.
• On the CVC-ColonDB dataset, achieved 96.04% precision, 96.68% recall, and 96.36% F1-score.

Conclusions:

• The enhanced YOLOv4 architecture significantly improves real-time polyp detection in colonoscopy.
• The method demonstrates superior performance compared to existing literature, offering a promising tool for early colorectal cancer detection.