A novel application of deep learning to forensic hair analysis methodology

• Published in: Australian journal of forensic sciences Vol. 56; no. 3; pp. 311 - 322
• Main Authors: Airlie, Melissa, Robertson, James, Ma, Wanli, Airlie, David, Brooks, Elizabeth
• Format: Journal Article
• Language: English
• Published: Sydney Taylor & Francis 03.05.2024; Australian Academy of Forensic Sciences; Copyright Agency Limited (Distributor)
• Subjects: Accuracy; Classification; Datasets; Deep learning; Deoxyribonucleic acid; DNA; forensic hair analysis; Forensic science; Forensic sciences; Hair; Hair analysis; Machine learning; Methodology; objective; Social media; supervised; Training; AUSTRALIA
• ISSN: 0045-0618; 1834-562X
• DOI: 10.1080/00450618.2022.2159064

A deep learning model called HairNet was developed to conduct forensic hair analysis, including the classification of hair as human and suitability for nuclear DNA (nDNA) analysis. The training and testing data used were microscopic images of hair features including the medulla and the hair root. The final model iterations obtained 100% accuracy on the medulla dataset to classify hair as human or non-human and between 96% and 100% accuracy on the hair root dataset to classify human hair as suitable for nDNA analysis depending on the grouping of root types. The greatest impact on accuracy was the quantity and quality of the training and testing data and therefore the critical step in model development. The application of ML to forensic methodology is a novel and innovative approach and a means to improve objectivity; however, the creation of training and testing data initially requires expert human judgement and therefore collaboration is essential in the development of benchmark datasets. This research demonstrates how deep learning can be successfully applied to forensic methodology and the possibilities for other forensic disciplines.