SmartPLR: a digital solution for AI-powered smartphone pupillometry
摘要
To develop a smartphone-based pupillometry using deep learning and evaluate its accuracy compared to a commercial pupillometer, the NPi-300.
Methods336 pupillary light reflex (PLR) exams from 158 volunteers were analyzed using deep learning models (UNet, UNet++, DeepLabV3, DeepLabV3+, and Mask R-CNN) with different backbones (ResNet50, Swin Transformer, and ConvNeXt V2). Once the best combination was identified, image data was filtered according to the degree of eyelid opening and image blurriness. The maximum-minimum pupil size difference, constriction velocity (CV), and percentage change in pupil size (CP) were compared between our application and the NPi-300 gold standard. The kernel density estimation and Bhattacharyya Distance were used to develop a scoring method to classify pupil reactivity: SmartPLR.
ResultsMask R-CNN (ConvNeXt V2 backbone), which showed a mean intersection over union of 0.9177, segmentation mean average precision (mAP) of 0.8670, and bounding box mAP of 0.8663, was selected for our application. The Pearson correlation values comparing our application to the NPi-300 for pupil size difference, CV, and CP were 0.77, 0.77, and 0.74, respectively. The SmartPLR formula was defined as
Despite various smartphone applications developed to evaluate the PLR, they rely on additional add-ons or infrared light source. This prevents such applications from being completely commercialized. Our novel smartphone application, built on deep learning and not requiring infrared or additional devices, demonstrated high accuracy compared to the NPi-300.