Multimodal artificial intelligence for retinal detachment diagnosis using fundus imaging and patient questionnaires

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Retinal detachment (RD) diagnosis relies on expert interpretation of fundus images and patient data.
• Current diagnostic methods can be time-consuming and may miss subtle cases.
• Developing automated diagnostic tools is crucial for early detection and treatment.

Purpose of the Study:

• To develop a multimodal artificial intelligence (AI) system for diagnosing retinal detachment (RD).
• To integrate ultra-widefield fundus imaging and patient questionnaire data for enhanced diagnostic accuracy.
• To compare the performance of the multimodal AI model against single-modal AI and human clinicians.

Main Methods:

• Collected ultra-widefield fundus images and patient questionnaires from RD patients and controls.
• Developed a multimodal AI model using the Contrastive Language-Image Pretraining framework.
• Conducted per-image and per-subject analyses to evaluate diagnostic performance.

Main Results:

• The multimodal AI model demonstrated superior performance compared to image-only or questionnaire-only models.
• Achieved high accuracy (0.899±0.054 per-image, 0.893±0.071 per-subject) and recall (0.902±0.043 per-image, 0.949±0.044 per-subject).
• The AI model showed a higher recall rate than human clinicians, improving detection of true RD cases.

Conclusions:

• Integrating fundus imaging and patient questionnaire data significantly enhances AI-based RD diagnosis.
• The multimodal approach shows promise for achieving clinician-level diagnostic accuracy.
• Future work should focus on larger datasets and refined questionnaires to further optimize AI performance.