Polyp Detection Using YOLOv9 on Real and Synthetic Colonoscopy Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Early detection of colorectal polyps is crucial for preventing cancer mortality.
• Current colonoscopy methods can miss polyps, necessitating improved detection technologies.

Purpose of the Study:

• To develop and assess a YOLOv9-based model for enhanced colorectal polyp detection.
• To evaluate the impact of synthetic polyp image augmentation on model performance.

Main Methods:

• Fine-tuning a pretrained YOLOv9-C model on the LDPolypVideo dataset.
• Generating synthetic polyp images using a diffusion model to augment the training data.
• Evaluating model performance using precision, recall, F1 score, and mean Average Precision (mAP) on internal and external datasets.

Main Results:

• The YOLOv9 model trained with synthetic images outperformed the model trained without synthetic images on the LDPolypVideo dataset across all metrics.
• The best-performing model achieved high precision (0.878), recall (0.861), F1 score (0.869), and mAP 50 (0.897) on the external Kvasir-SEG dataset.
• The model demonstrated strong generalization capabilities, performing well on unseen data.

Conclusions:

• Integrating synthetic polyp images effectively enhances the performance of YOLOv9 for colorectal polyp detection.
• This AI-driven approach addresses limitations in training data diversity and shows significant potential for real-world colonoscopy applications.