Automatic eye localization for hospitalized infants and children using convolutional neural networks
| Autor: | Vanessa Prinsen, Rita Noumeir, Sally Al Omar, Gabriel Masson, Armelle Bridier, Philippe Jouvet |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Adult
020205 medical informatics Computer science Remote patient monitoring Hospital bed Health Informatics Image processing 02 engineering and technology Machine learning computer.software_genre Convolutional neural network Clinical decision support system 03 medical and health sciences 0302 clinical medicine Image Processing Computer-Assisted 0202 electrical engineering electronic engineering information engineering Humans 030212 general & internal medicine Child Pediatric intensive care unit business.industry Infant 3. Good health The Internet Neural Networks Computer Artificial intelligence Performance improvement business computer Algorithms |
| Zdroj: | International Journal of Medical Informatics. 146:104344 |
| ISSN: | 1386-5056 |
| DOI: | 10.1016/j.ijmedinf.2020.104344 |
| Popis: | The appearance and behaviour of the eye region are important windows into a patient’s condition and level of consciousness, particularly for patients too young to speak. Unfortunately, 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. The overall aim of this research project is to develop a clinical decision support system that uses bedside cameras to detect signs of consciousness and distress due to pain. This work focuses on the first problem to be solved, namely how to locate the eyes in an image of a pediatric patient in a hospital bed. 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. Few studies have examined the application of computer vision and facial analysis techniques to young children in a hospital environment. To develop an appropriate solution for eye localization, a new training dataset, formed of images of young children from internet searches, is added to adult facial images to train cascade classifiers and convolutional neural networks. Another novel dataset, consisting of 59 recordings of patients in a pediatric intensive care unit, is used to evaluate the performance of these models. This dataset will also serve future work on this and other research projects in pediatric computer vision. The convolutional neural network trained with the added image data of young children achieves a 79.7% eye localization rate, much higher than models trained on adult data alone. This model also outperforms the cascade models. The dramatic performance improvement gained from adding task-specific images to the training data highlights the need for custom-trained models for specialized applications like pediatric patient monitoring. Existing models and datasets are not sufficient, but the moderate size of the task-specific training dataset used here suggests that developing an internal training dataset is within reach of a typical large hospital. The effectiveness of the convolutional neural network, given the challenges of this setting, makes it a powerful approach for eye localization and tracking in the hospital environment. The convolutional neural network’s ability to learn unique features allows it to adapt to the challenges of eye localization in an atypical setting where usual assumptions about facial appearance do not necessarily apply. The present weaknesses of the model, like poor recognition of uncommon eye appearances and slow image processing times, will improve with larger training datasets and technological improvements. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect