SmartPLR: a digital solution for AI-powered smartphone pupillometry.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Smartphone-based pupillometry aims to provide a portable and accessible method for assessing pupillary light reflex (PLR).
• Existing smartphone applications often require additional hardware, limiting their commercial viability.

Purpose of the Study:

• To develop a deep learning-based smartphone application for pupillometry.
• To evaluate the accuracy of this application against a commercial pupillometer (NPi-300).

Main Methods:

• Deep learning models (Mask R-CNN with ConvNeXt V2 backbone) were trained and validated on 336 PLR exams.
• Image quality filtering was applied based on eyelid opening and blurriness.
• Pupil size difference, constriction velocity (CV), and percentage change (CP) were compared with the NPi-300, leading to the SmartPLR scoring system.

Main Results:

• The Mask R-CNN model achieved high segmentation and detection accuracy (mAP 0.8670).
• Strong Pearson correlations were found between the smartphone app and NPi-300 for pupil size difference (0.77), CV (0.77), and CP (0.74).
• The SmartPLR system was defined to classify pupil reactivity.

Conclusions:

• A novel smartphone application utilizing deep learning for pupillometry was successfully developed.
• The application demonstrated high accuracy comparable to a commercial device, without needing infrared light or add-ons.
• This technology presents a fully commercializable solution for remote and accessible PLR assessment.