Evaluating the Efficacy of Convolution Neural Networks in Age at Deat Estimation Using 3D Scans of the Pubic Symphyseal Face
Date of Award
Master of Arts
Dr. Jacqueline Eng
Dr. Michelle Machicek
Dr. Britt Hartenberger
Machine learning, forensic anthropology, pubic symphysis, convolutional neural network, age at death estimation
Masters Thesis-Open Access
The research presented assesses the utility of machine learning approaches, specifically convolutional neural networks (CNNs), to the estimation of age at death in adult decedents by analysis of the pubic symphyseal face of the os coxa rendered as a 3D image. Age at death estimation is an important duty of forensic anthropologists working in medico-legal contexts, as well as bioarcheological researchers. The purpose of this study is to evaluate the accuracy of a CNN relative to the performance of human observers using traditional methods of age estimation. To accomplish this, a CNN created for this project and expert anthropologists were tasked to provide age at death estimates for a selected population with a known age at death. CNN estimation is expected to achieve good accuracy among young adults, and to outperform humans among older adults.
A critical evaluation of the challenges associated with integration of machine learning technology to applied forensics is provided, as well as a novel strategy for interpretation of CNN age at death determinations. The results of the research indicate that, contrary to expectation, CNN age at death estimates are not accurate among young or older adults. Human estimates are superior for these age groups. However, CNN results are surprisingly highly accurate among the middle aged. A number of confounding factors, including primarily population biases in the training set, may contribute to these results. A review of future strategies for improvement of CNN performance is offered.
Brown, Melissa A., "Evaluating the Efficacy of Convolution Neural Networks in Age at Deat Estimation Using 3D Scans of the Pubic Symphyseal Face" (2018). Masters Theses. 3695.