Automatic eye localization for hospitalized infants and children using convolutional neural networks

Reliable localization and tracking of the eye region in pediatric hospital environment is a significant challenge for clinical decision support patient monitoring applications. Existing work achieves high performance on adult datasets but performs poorly busy environment, where face appearance varies because age, position presence medical equipment. We developed two new datasets: training dataset using public image data from internet searches, test 59 recordings patients intensive care unit. trained models, Faster R-CNN algorithm to fine-tune pre-trained ResNet base network, evaluated them images ICU. The convolutional neural network with combination child achieved an 79.7% rate, significantly higher than model alone. With additional pre-processing equalize contrast, rate rises 84%. results demonstrate potential networks ICU setting, even when limited. obtained gains by adding task-specific dataset, highlighting need custom models specialized applications like monitoring. moderate size our added shows that it feasible develop internal computer vision applications, apply transfer learning existing models.