Vanessa Prinsen1, Philippe Jouvet2, Sally Al Omar3, Gabriel Masson4, Armelle Bridier3, Rita Noumeir5. 1. École de technologie supérieure, 1100 Notre-Dame St W, Montréal, Québec H3C 1K3 Canada; CHU Sainte-Justine, 3175 Chemin de la Côte-Sainte-Catherine, Montréal, QC H3T 1C5. Electronic address: vanessa.prinsen.1@ens.etsmtl.ca. 2. CHU Sainte-Justine, 3175 Chemin de la Côte-Sainte-Catherine, Montréal, QC H3T 1C5. 3. Université de Montréal, 2900 Edouard Montpetit Blvd, Montréal, Québec H3T 1J4 Canada. 4. CHU Lille, Pôle Anesthésie Réanimation, Lille, France. 5. École de technologie supérieure, 1100 Notre-Dame St W, Montréal, Québec H3C 1K3 Canada.
Abstract
BACKGROUND: Reliable localization and tracking of the eye region in the pediatric hospital environment is a significant challenge for clinical decision support and patient monitoring applications. Existing work in eye localization achieves high performance on adult datasets but performs poorly in the busy pediatric hospital environment, where face appearance varies because of age, position and the presence of medical equipment. METHODS: We developed two new datasets: a training dataset using public image data from internet searches, and a test dataset using 59 recordings of patients in a pediatric intensive care unit. We trained two eye localization models, using the Faster R-CNN algorithm to fine-tune a pre-trained ResNet base network, and evaluated them using the images from the pediatric ICU. RESULTS: The convolutional neural network trained with a combination of adult and child data achieved an 79.7% eye localization rate, significantly higher than the model trained on adult data alone. With additional pre-processing to equalize image contrast, the localization rate rises to 84%. CONCLUSION: The results demonstrate the potential of convolutional neural networks for eye localization and tracking in a pediatric ICU setting, even when training data is limited. We obtained significant performance gains by adding task-specific images to the training dataset, highlighting the need for custom models and datasets for specialized applications like pediatric patient monitoring. The moderate size of our added training dataset shows that it is feasible to develop an internal training dataset for clinical computer vision applications, and apply it with transfer learning to fine-tune existing pre-trained models.
BACKGROUND: Reliable localization and tracking of the eye region in the pediatric hospital environment is a significant challenge for clinical decision support and patient monitoring applications. Existing work in eye localization achieves high performance on adult datasets but performs poorly in the busy pediatric hospital environment, where face appearance varies because of age, position and the presence of medical equipment. METHODS: We developed two new datasets: a training dataset using public image data from internet searches, and a test dataset using 59 recordings of patients in a pediatric intensive care unit. We trained two eye localization models, using the Faster R-CNN algorithm to fine-tune a pre-trained ResNet base network, and evaluated them using the images from the pediatric ICU. RESULTS: The convolutional neural network trained with a combination of adult and child data achieved an 79.7% eye localization rate, significantly higher than the model trained on adult data alone. With additional pre-processing to equalize image contrast, the localization rate rises to 84%. CONCLUSION: The results demonstrate the potential of convolutional neural networks for eye localization and tracking in a pediatric ICU setting, even when training data is limited. We obtained significant performance gains by adding task-specific images to the training dataset, highlighting the need for custom models and datasets for specialized applications like pediatric patient monitoring. The moderate size of our added training dataset shows that it is feasible to develop an internal training dataset for clinical computer vision applications, and apply it with transfer learning to fine-tune existing pre-trained models.
Authors: Monisha Shcherbakova; Rita Noumeir; Michael Levy; Armelle Bridier; Victor Lestrade; Philippe Jouvet Journal: IEEE Open J Eng Med Biol Date: 2021-12-17