By Angus Dalton
Australian scientists have created an AI tool that can quickly and accurately classify skulls to assist in the identification of remains recovered after natural disasters, mass-casualty events and crime scenes.
The tool can estimate the biological sex of a skull with 97 per cent accuracy, outperforming the 82 per cent accuracy of human assessors, and it works five times faster.
AI took a different route to human assessors when analysing skull structures.Credit: Lye et al, Nature Scientific Reports
The development of the neural network model marks the addition of AI techniques into the forensic anthropological toolkit, CSIRO research scientist Dr Hollie Min said.
“The tool may be especially useful in instances where there are large populations who need to be identified, for example, a plane crash or the discovery of mass graves,” Min, co-lead author of the study published in Nature Scientific Reports, said.
Manual identification of skulls often relies on a technique pioneered by anthropologist Phillip Walker that focuses on five regions of the skull.
Males tend to have a larger mastoid process, for example, a conical protrusion you can feel behind your ear, and a more hooked nuchal crest, which is the curve at the back of the skull where neck muscles attach.
The five structures of the skull used in the Walker method of identifying the sex of skulls.Credit: Lye et al, Nature Scientific Reports
There are also differences in the edges of the eye socket or supraorbital margin. It’s more likely to be sharp like “the edge of a dull knife” in females, according to Walker, while in males it tends to be thick and blunt like “the curvature of a pencil”.
The area of the skull between the brows, the glabella, is generally more pronounced in males too.
“We fed Walker’s identifying features into the model to train it, and we also allowed the model to learn its own traits. And they didn’t always overlap,” said study co-author Dr Jason Dowling, who leads a team within CSIRO’s Health and Biosecurity unit focused on medical imaging.
“The deep learning model has learnt different features related to the morphology of the skull.”
The AI picked up differences in the glabella and the nuchal crest, but glossed over the other three Walker traits.
The model instead focused on the overall size and shape of the skull, and potentially tuned in to more subtle structural aspects that escape human attention.
The researchers trained the AI on 200 CT scans from living patients in a hospital at Hasanuddin University in Indonesia, expanding the diversity of imaging data for Asian populations, which is currently lacking.
The new AI tool could be trained further with subsets of skulls from different populations to increase its accuracy further across different groups.
“Our goal is to provide forensic anthropologists with a reliable, interpretable tool to support their critical work, especially in cases involving individuals of unknown population backgrounds,” Min said.
Assessors using the Walker technique on south Asian skulls get it wrong about 37 per cent of the time, one study found. The manual method was based on European, African and native American skulls, which explains the discrepancy.
The scientists are seeking commercial collaborators to develop their AI tool for real-life applications.
The Examine newsletter explains and analyses science with a rigorous focus on the evidence. Sign up to get it each week.