Screening and identifying hepatobiliary diseases through deep learning using ocular images: a prospective,...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Ocular changes linked to hepatobiliary diseases are often non-specific and hard to detect.
• Traditional diagnostic methods have limitations in identifying early-stage or specific hepatobiliary conditions through eye examination.

Purpose of the Study:

• To develop and validate deep learning models for automated screening and identification of hepatobiliary diseases using ocular images.
• To establish novel associations between ocular features and major hepatobiliary diseases.

Main Methods:

• A multicentre, prospective study involving slit-lamp and retinal fundus images from over 1700 participants.
• Development and external validation of seven slit-lamp and seven fundus deep learning models for disease screening and identification.
• Performance evaluation using Area Under the Receiver Operating Characteristic Curve (AUROC), sensitivity, specificity, and F1 score.

Main Results:

• Slit-lamp models achieved an AUROC of 0.74 for screening and up to 0.93 for identifying specific diseases like liver cancer.
• Fundus models showed an AUROC of 0.68 for screening and up to 0.84 for liver cancer identification.
• The models identified contributions from conjunctiva, sclera, iris, and fundus structures in disease classification.

Conclusions:

• Deep learning models demonstrate potential for non-invasive, convenient screening and identification of hepatobiliary diseases via ocular imaging.
• These models can serve as an opportunistic screening tool, complementing existing diagnostic approaches for liver conditions.