Automatic retinoblastoma screening and surveillance using deep learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Pediatric Oncology

Background:

• Retinoblastoma is the most common childhood intraocular cancer.
• Improved survival rates present challenges in long-term patient surveillance and offspring screening.
• Deep learning offers a potential solution to reduce the burden of follow-up and screening.

Purpose of the Study:

• To develop and validate a deep learning algorithm for retinoblastoma monitoring.
• To assess the algorithm's accuracy in distinguishing normal, stable, and active retinoblastoma.
• To evaluate the cost-effectiveness of an AI-based approach for retinoblastoma management.

Main Methods:

• A cohort study involving retinoblastoma patients from Beijing Tongren Hospital (March 2018 - June 2022).
• Development of the Deep Learning Assistant for Retinoblastoma Monitoring (DLA-RB) algorithm using 36,623 fundus images.
• Prospective validation of DLA-RB on 139 eyes, comparing its performance with ophthalmologists.

Main Results:

• DLA-RB achieved high Area Under the Curve (AUC) values: 0.998 for normal vs. active and 0.940 for stable vs. active retinoblastoma in internal validation.
• Prospective validation showed AUCs of 0.991 and 0.962 for identifying active retinoblastoma.
• Combination of DLA-RB with ophthalmologists significantly improved diagnostic accuracy; the AI-based mode proved cost-effective.

Conclusions:

• DLA-RB demonstrates high accuracy and sensitivity in identifying active retinoblastoma.
• The algorithm can effectively aid in retinoblastoma surveillance and high-risk offspring screening.
• DLA-RB offers a cost-effective, telemedicine-compatible solution for retinoblastoma diagnosis and monitoring.