Real-World Colonoscopy Video Integration to Improve Artificial Intelligence Polyp Detection Performance and Reduce...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Artificial intelligence (AI) for colon polyp detection shows high sensitivity but low specificity in real-world applications.
• Current AI models often rely solely on public datasets, limiting their practical performance.
• Manual annotation of colonoscopy videos is labor-intensive for expert endoscopists.

Purpose of the Study:

• To develop and evaluate a semi-automatic annotation method using real colonoscopy videos to improve AI colon polyp detection.
• To enhance AI model performance by incorporating real-world data.
• To reduce the manual labeling burden in AI model development.

Main Methods:

• An integrated AI model was trained and validated on a large dataset (86,258 training, 17,616 validation images).
• Two models were compared: Model 1 (public datasets only) and Model 2 (public datasets plus semi-automatically annotated real colonoscopy videos).
• Semi-automatic annotation significantly reduced the need for expert endoscopist labeling.

Main Results:

• Model 2, incorporating real-world data, significantly outperformed Model 1.
• At epoch 35, Model 2 achieved 90.6% sensitivity, 96.0% specificity, 94.5% accuracy, and 89.9% F1 score.
• All polyps in test videos were detected by Model 2, showing enhanced performance.

Conclusions:

• Semi-automatic annotation of real colonoscopy videos markedly improves AI diagnostic accuracy for colon polyp detection.
• This approach can potentially decrease the reliance on extensive manual annotation by experts.
• Further validation using multicenter external datasets is necessary to confirm generalizability.