AI framework for multidisease detection via retinal imaging

Area of Science:

• Ophthalmology
• Endocrinology
• Metabolic Diseases
• Artificial Intelligence
• Medical Imaging

Background:

• Endocrine and metabolic diseases pose a growing global health challenge.
• Current screening methods can be resource-intensive and time-consuming.
• There is a need for scalable and accessible screening solutions.

Purpose of the Study:

• To develop Reti-Pioneer, an AI-powered framework for multitask retinal imaging.
• To enable efficient, multidisease detection using color fundus photographs.
• To assess the framework's performance and generalizability across diverse populations.

Main Methods:

• Developed a multitask retinal imaging framework (Reti-Pioneer) integrating quality-aware modules and pre-trained foundation models.
• Trained the framework on 107,730 color fundus photographs from community and hospital cohorts.
• Validated performance on internal test data and six external cohorts, including resource-limited settings.

Main Results:

• Achieved high AUC values for detecting type 2 diabetes mellitus (0.833), gout (0.832), osteoporosis (0.787), hypertension (0.740), hyperlipidemia (0.736), and thyroid disease (0.699).
• Demonstrated strong generalization across diverse external cohorts.
• Screening time averaged 30.6 ± 6.0 seconds per case, significantly faster than standard workflows.
• Clinical pilot for type 2 diabetes mellitus showed AUC of 0.776 and NPV of 0.966, outperforming the Finnish Diabetes Risk Score.

Conclusions:

• Reti-Pioneer offers an efficient, low-cost, and translatable screening pathway using retinal imaging (oculomics).
• The framework demonstrates potential for early detection and management of multiple endocrine and metabolic diseases.
• High acceptance by clinicians and patients suggests clinical utility in primary care settings.