In vivo computer-aided diagnosis of colorectal polyps using white light endoscopy

Area of Science:

• Gastroenterology
• Medical Artificial Intelligence
• Endoscopic Imaging

Background:

• Existing AI systems for polyp histology prediction often require complex optical technologies and lack in vivo validation.
• High-definition white light endoscopy (WLE) is a standard tool, but accurate real-time optical diagnosis remains challenging.

Purpose of the Study:

• To evaluate the efficacy of the ATENEA deep learning system for in vivo optical diagnosis of colorectal polyps using only WLE.
• To compare the diagnostic performance of ATENEA against experienced endoscopists in a clinical setting.

Main Methods:

• Prospective testing of the ATENEA system on consecutive colorectal polyps during screening colonoscopies.
• Comparison of ATENEA's WLE-based predictions (adenoma vs. non-adenoma) with those of four staff endoscopists blinded to the AI output.
• Histological analysis served as the gold standard for diagnosis.

Main Results:

• ATENEA achieved 91.3% accuracy in predicting adenomas and 57.1% for non-adenomas.
• Endoscopists demonstrated 75.4% accuracy for adenomas and 95.2% for non-adenomas.
• Overall accuracy was 83.3% for ATENEA and 80% for endoscopists, with ATENEA showing similar global performance but lower accuracy for non-adenomatous lesions.

Conclusions:

• The ATENEA system shows potential for accurate in vivo characterization of colorectal polyps using standard WLE.
• ATENEA's performance is comparable to endoscopists, potentially aiding real-time clinical decision-making.
• Further refinement is needed to improve ATENEA's accuracy in diagnosing non-adenomatous colorectal lesions.