Non-destructive Automated Classification of Human and Large Mammal Long Bone Fragments: A Deep Learning Approach Using Micro-CT Histomorphology and Grad-Cam Interpretation

Authors

• Kathleen Marie Kelley, University of Nebraska-LincolnFollow

First Advisor

William R. Belcher

Committee Members

Kelly Kamnikar, LuAnn Wandsnider

Date of this Version

5-2026

Document Type

Thesis

Citation

A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Arts

Major: Anthropology

Under the supervision of Professor William Belcher

Lincoln, Nebraska, May 2026

Comments

Copyright 2026, Kathleen Marie Kelley. Used by permission

Abstract

This study explores the development and validation of an automated, deep learning system designed to differentiate human from large mammal long bone fragments using micro-CT imagery. A comprehensive dataset of cross-sectional µCT images was assembled from human and animal skeletal material processed at the DPAA, Nebraska. A convolutional neural network (CNN) based on a ResNet-18 architecture, utilizing transfer learning from ImageNet weights, was trained to classify µCT images as human or animal. The model achieved a mean classification accuracy of 89.85% (± 8.27%) across five sample-level cross-validation folds, with a sensitivity of 96.21% and a 95% bootstrap confidence interval of [89.58%, 90.09%]. Model performance was evaluated using accuracy metrics, confusion matrices, and bootstrap-derived confidence intervals. Explainable Artificial Intelligence (XAI) methods were incorporated to assess the biological relevance of features identified as discriminative by the model, revealing that classification draws primarily on internal porosity and cortical remodeling texture for human bone, and on plexiform organization and diffuse low-level textural cues for animal bone, with void space composition identified as a secondary incidental signal warranting further investigation. The goal of this research is to produce a validated, non-destructive screening tool that rapidly and reliably distinguishes human from animal skeletal material while generating reproducible classifications with quantified confidence measures. With a mean accuracy of 89.85%, a sensitivity profile well suited to forensic triage, and Grad-CAM analysis suggesting biologically coherent classification strategies across much of the sample, this study demonstrates a viable proof of concept for the application of deep learning to µCT-based species discrimination. It is important to note that this tool is intended to complement, not replace, expert anthropological assessment. Instead, it offers empirically supported informed decision-making in the early stages of forensic workflows. Although the immediate application of this program is for Defense POW/MIA Accounting Agency (DPAA) casework involving fragmentary and commingled remains from combat loss incidents, this methodology has broader relevance for mass disaster victim identification, archaeological screening, and contemporary forensic investigations.

Advisor: William R. Belcher