Artificial intelligence in colonoscopy: from detection to diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopy is crucial for detecting gastrointestinal diseases.
• Advancements in artificial intelligence (AI) offer potential improvements in colonoscopy procedures.
• Deep learning (DL) techniques are increasingly applied to medical image analysis.

Purpose of the Study:

• To review recent progress in AI for colonoscopy, focusing on detection and diagnosis.
• To identify effective deep learning models for specific colonoscopy tasks.
• To summarize the performance metrics of AI in colonoscopy from recent studies.

Main Methods:

• Systematic review of 27 original studies from PubMed published in 2021 or later.
• Search terms: "colonoscopy" (title) and "deep learning" (abstract).
• Inclusion criteria focused on deep learning for classification, detection, and segmentation in colonoscopy, with specific outcome measures.

Main Results:

• Specific deep learning models demonstrated high efficacy: Efficientnet for classification (AUC 99.8%), You Only Look Once for detection (F1 96.3%), and Unet for segmentation (Dice 97.3%).
• Performance metrics varied across studies, including accuracy (74.0-95.0%), sensitivity (60.0-93.0%), specificity (60.0-100.0%), and AUC (71.0-99.8%).
• Inference speeds (FPS) ranged from 66 to 182.

Conclusions:

• AI, particularly deep learning, serves as an effective decision support system for colonoscopy.
• AI aids in the detection and diagnosis of gastrointestinal diseases during colonoscopy.
• The findings highlight the potential of AI to enhance colonoscopy outcomes and patient care.