Objective:

To develop and evaluate a physician-in-the-loop AI system (RetinaAIM) for automated panretinal photocoagulation (PRP) planning using ultrawidefield (UWF) fundus images. The system is designed to improve consistency and safety in PRP delivery, particularly in settings with limited access to experienced retinal specialists, by integrating instance segmentation, patient-specific geometric modeling, and automated power modulation to generate reproducible, safety-aware laser treatment patterns.

Methods:

The training dataset consisted of 300 manually annotated UWF images, expanded to 700 images through structured augmentation. A Roboflow 3.0 Instance Segmentation (Fast) model was trained using a 70%/20%/10% train/validation/test split. Images were normalized to 640×640 resolution, and annotations included retinal field (RF), optic disc (OD), and posterior pole (PP). The model first identified the treatable RF, after which anatomical no-go regions (PP, OD) were geometrically subtracted. Remaining regions underwent Poisson-disc sampling with an adjustable pixel radius to generate evenly spaced laser coordinates. Real-time optic disc diameter measurements produced standardized, patient-specific safety margins, which could be further customized by physicians based on pathology and clinical gestalt. Physicians retained full control over spacing, burn density, and power presets.

Results:

The model achieved high segmentation performance (mAP@50 98.0%; mAP@50–95 73.0%; precision 98.7%; recall 98.3%; F1 98.5%). Two PP false negatives were observed at the optimal threshold (68%); however, applying the system’s 5% operating confidence level preserved all PP and OD territories, yielding 100% protection of no-go anatomical structures with no unsafe exposures. The planner generated ~700–2,500 evenly spaced candidate burns per image while maintaining full physician oversight. Average planning time was 8.4 seconds on local M-series hardware.

Conclusion:

This prototype demonstrates a high-accuracy, geometry-aware PRP planning system capable of producing standardized, safety-preserving laser maps in near–real time. By unifying robust segmentation, anatomical subtraction, spatial optimization, and physician supervision, RetinaAIM provides a foundation for consistent, patient-specific PRP planning. Such tools may help reduce provider-to-provider variability and expand access to high-quality diabetic retinopathy care in settings with limited retinal specialist availability, including rural and underserved regions globally.