OBUSight: Clinically Aligned Generative AI for Ophthalmic Ultrasound Interpretation and Diagnosis.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Ocular B-scan ultrasonography (OBU) is crucial for diagnosing posterior segment ocular disorders.
• Interpreting OBU images presents significant challenges for ophthalmologists.
• Existing diagnostic methods require expert interpretation, leading to potential delays and variability.

Purpose of the Study:

• To develop and evaluate OBUSight, a generative AI model for comprehensive OBU image interpretation.
• To enable joint report generation and disease diagnosis for OBU images.
• To enhance diagnostic efficiency and accuracy in ophthalmic ultrasound practice.

Main Methods:

• OBUSight was trained and validated on a large multi-center OBU dataset (39,654 images, 17,586 reports, 11,381 patients).
• Model performance was assessed using natural language generation (NLG) and clinical efficacy (CE) metrics.
• Expert ratings and comparative evaluations with clinicians (ophthalmologists, residents, students) were conducted.

Main Results:

• OBUSight outperformed eight state-of-the-art models in report generation quality and clinical efficacy.
• The AI demonstrated robust performance across multi-center and multimorbidity datasets.
• Expert reviews confirmed OBUSight's ability to produce clinically aligned reports with minimal corrections.
• In evaluations, OBUSight significantly outperformed residents and students, matching ophthalmologist performance and reducing diagnostic time.

Conclusions:

• OBUSight is a promising AI tool for improving diagnostic efficiency in ophthalmic ultrasound.
• The AI assists clinicians by providing accurate diagnoses and generating reports.
• OBUSight shows particular benefit for less experienced clinicians, enhancing overall practice quality.