Multi-modal endoscopic artificial intelligence system improving classification of colorectal subepithelial lesions: a...

Area of Science:

• Gastroenterology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Colorectal subepithelial lesions (SELs) exhibit diverse histology and prognosis despite similar endoscopic appearances, posing diagnostic challenges.
• Accurate classification of SELs is critical for determining appropriate patient management and treatment strategies.

Purpose of the Study:

• To develop and evaluate a multi-modal endoscopic artificial intelligence (AI) system, named AIOSCOPE-WE, for classifying colorectal SELs.
• To compare the diagnostic performance of the AI system against single-modal AI models and human endoscopists.

Main Methods:

• Development of single-modal (WLE, EUS) and multi-modal (AIOSCOPE-WE) AI models using retrospective data from 510 patients with confirmed colorectal SELs.
• Internal and external validation of AI models' classification performance.
• Comparison of endoscopists' diagnostic accuracy with and without AIOSCOPE-WE assistance.

Main Results:

• AIOSCOPE-WE demonstrated superior accuracy (86.0%) in classifying colorectal SELs compared to EUS experts (75.8%) and non-experts (65.4%).
• Endoscopist accuracy improved significantly from 70.6% to 83.1% with AIOSCOPE-WE support.
• AIOSCOPE-WE outperformed single-modal models (WLE, EUS) in both internal (92.1%) and external (84.9%) tests, showing consistent performance across various lesion characteristics.

Conclusions:

• The AIOSCOPE-WE system shows significant potential as an effective tool for endoscopists in diagnosing colorectal SELs.
• This AI-powered approach can enhance diagnostic accuracy and optimize clinical decision-making for patients with colorectal SELs.