Weakly supervised end-to-end artificial intelligence in gastrointestinal endoscopy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Artificial intelligence (AI) is crucial for analyzing gastrointestinal (GI) endoscopy images, with approved algorithms for polyp detection.
• The high cost of manual annotations limits AI applications beyond polyp detection.
• Weakly supervised learning offers a potential solution to overcome annotation barriers.

Purpose of the Study:

• To investigate the efficacy of a weakly supervised AI model trained on routine, non-annotated GI endoscopy data.
• To assess the AI's ability to learn visual patterns for diagnosing a wide range of GI diseases.
• To evaluate the AI's performance in both internal and external validation datasets.

Main Methods:

• A deep neural network was trained on a large dataset (N=29,506 gastroscopy, N=18,942 colonoscopy) using only routine diagnosis data.
• The model was trained on 42 common GI diseases without manual image labeling or annotation.
• Performance was evaluated using cross-validated area under the receiver operating curve (AUROC) and external validation.

Main Results:

• The AI system achieved high performance across diverse GI diseases, including inflammatory, degenerative, infectious, and neoplastic conditions.
• Cross-validated AUROC exceeded 0.70 for 13 diseases and 0.80 for two diseases.
• External validation showed significant prediction of diverticulosis, candidiasis, and colon/rectal cancer (AUROC > 0.76).

Conclusions:

• Weakly supervised AI can generate high-performing diagnostic classifiers from non-annotated routine GI endoscopy data.
• The AI successfully identified clinically relevant visual patterns associated with various GI diseases.
• This approach holds promise for expanding AI applications in GI endoscopy and other clinical imaging fields.