Computer-Aided Diagnosis Using the Large-Scale Visual Language Models in Screening of Digestive Endoscopy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• High prevalence of digestive tract diseases necessitates large-scale endoscopic screening in China.
• Limited staff-to-patient ratios create a diagnostic workload challenge in endoscopic screening programs.

Purpose of the Study:

• To develop and evaluate an AI-powered computer-aided diagnosis (CAD) method for digestive endoscopy screening.
• To leverage pre-trained vision-language models for enhanced classification of endoscopic images.

Main Methods:

• A CAD method utilizing pre-trained large-scale vision-language models was developed.
• The model fused visual and textual information for classification of endoscopic images.
• Training data comprised 41,191 multicenter images across 19 disease categories and 4 normal conditions.

Main Results:

• The proposed AI model outperformed eight state-of-the-art models in comparative evaluations.
• In human-AI collaboration, the model showed superior diagnostic performance compared to experienced endoscopists.
• AI assistance led to faster decision-making times in diagnostic tasks.

Conclusions:

• The AI-driven CAD method is effective for analyzing large-scale endoscopic datasets.
• AI has the potential to significantly improve diagnostic efficiency and accuracy in digestive endoscopy.
• This study advances AI-assisted diagnostics, moving towards AI-driven clinical reality in endoscopy.