Automatic gestational age estimation from newborn babies’ facial photo analysis.Tools Mosanya, Chidera Aristotle (2016) Automatic gestational age estimation from newborn babies’ facial photo analysis. [Dissertation (University of Nottingham only)]
AbstractKnowledge of the gestational age of a baby is important to determine if the baby is premature and to provide it with the necessary support and medication if any is required. According to the World Health Organization, more than 1 in 10 babies are born prematurely every year, and complications from this preterm birth are the leading causes of death among children under the age of five. Complications from preterm birth can be prevented by proper support and medication. However, accurate calculation of gestational age to determine if the baby is preterm can be problematic. It often requires the first date of the last menstruation before conception or the use of an ultrasound machine. Many pregnant women may not remember the first date of the last menstruation, and ultrasound machines may be unavailable. Health care professionals sometimes find it difficult to calculate the accurate gestational age if the menstruation history or an ultrasound machine is unavailable. They utilize the Ballard score system, a manual process to analyse the baby’s physical and neurological traits after the baby is born, but this process relies heavily on the experience of the healthcare professional and is prone to human error. This project proposes and implements a new method of predicting the gestational age of a baby within a few days or weeks of birth by analysing its facial picture. It takes two approaches. First, a traditional machine learning algorithm is used to estimate the gestational age. In this method, the face images are processed using computer vision techniques. Face detection and facial alignment techniques such as the Cascaded Continuous Regression (CCR) are used to detect the face and localize the facial points in the images. Image features such as SIFT, HOG, and HSV features are extracted from these facial points and are used with the already known gestational ages of the babies to train a random forest classifier to predict the gestational age of a new baby. In the second approach, a Convolutional Neural Network (CNN) is implemented using the Caffe library over a GPU. This CNN takes the processed images and estimates their gestational ages. The delicacy of the project and the high level of accuracy required in predicting the gestational ages influenced the choice of machine learning algorithms. Random Forest (which has been proven to be robust in handling high dimensional features), and Convolutional Neural Networks (which have shown success in image classification and computer vision in general) were chosen. These algorithms were trained and tested to predict the gestational ages, and their performances were evaluated to analyse which of them performs better. Convolutional Neural network performed slightly better than random forest. However, both algorithms did not perform as expected because of insufficient data available for this experiment.
Actions (Archive Staff Only)
|
Tools
Tools
![[thumbnail of Chidera Aristotle Mosanya 4244478.pdf]](https://eprints.nottingham.ac.uk/style/images/fileicons/application_pdf.png "Chidera Aristotle Mosanya 4244478.pdf")
