Degree

Doctor of Philosophy (PhD)

Department

Geography and Anthropology

Document Type

Dissertation

Abstract

This dissertation enhances age-at-death estimation by examining the impact of pelvic anomalies on age estimation methods and exploring quantitative approaches using three-dimensional (3D) surface complexity measurements and deep learning models. Paired os coxa and sacra from four skeletal collections were analyzed.

The effects of pelvic anomalies were assessed on three age-at-death methods: Hartnett (2010a), Buckberry and Chamberlain (2002), and Passalacqua (2009). Accuracy, absolute error, and bias were evaluated across the sample, where 140 individuals (27%) exhibited pelvic anomalies. The Passalacqua 95% confidence interval (CI) demonstrated the highest accuracy when anomalies were present, while the 68% CI had the lowest. Overall, accuracy decreased, and absolute error and bias increased, when anomalies were present, indicating aging methods should be applied cautiously when anomalies are present due to their influence on joint degeneration.

Next, Dirichlet normal energy (DNE) was implemented to quantify the 3D geometry of three pelvic joint surfaces: the pubic symphysis (PS) and the iliac and sacral auricular surfaces (IAS and SAS, respectively). DNE values were calculated from 3D scanned joint surfaces, and Spearman’s rank correlation analyses found low but significant correlations between DNE values and age: R² = 0.0699 (PS), R² = 0.1076 (IAS), R² = 0.0739 (SAS), and R² = 0.1190 (Combined Joint Group), all with a p-value of 0.000. Despite weak correlations, the significance suggests a relationship between DNE values and age, with a general trend of increasing DNE values with age.

Lastly, a deep learning model was implemented to classify digital images of the three pelvic joint surfaces into discrete age groups. Using an adapted ResNet50 convolutional neural network, models were trained, validated, and tested on images of the PS, IAS, and SAS. Training parameters included data augmentation, 20 epochs, and a batch size of 32 to classify images into six age groups. The SAS achieved the highest test accuracy (24%), while the IAS had the lowest (21%). After collapsing age groups into 18–69 years and 70+ years, the PS yielded the highest accuracy (59%). This study demonstrates the potential of deep learning in forensic anthropology, but highlights the need for larger training datasets to improve accuracy.

Date

3-26-2025

Committee Chair

Listi, Ginesse

Download

Available for download on Wednesday, March 24, 2032

Share

COinS