A volume-rendered CT scan featuring the five cranial traits used in the study. Credit: CSIRO
Researchers at CSIRO, Australia’s national science agency, have developed an advanced artificial intelligence (AI) tool that can more accurately and more quickly estimate biological sex from skulls compared with human forensic anthropologists.
The study authors say the tool can support forensic anthropologists when results are needed rapidly, such as in a mass casualty event. It also can better account for a more diverse population.
In traditional forensic anthropology, one of the most popular methods for morphoscopic cranial sex assessment is the Walker method, derived from English/American and Native American population groups. While the method, which involves the manual assessment of five dimorphic cranial traits, is still functional, the study authors argue that modern technologies can help facilitate new, more diverse approaches. For example, virtual collections imaged with computer tomography (CT) allow researchers to obtain skeletal datasets that are representative of a contemporary population.
“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,” said Hollie Min, CSIRO research scientist and first author of the new study.
In the study, published in Scientific Reports, Min and her team trained multiple deep learning frameworks on biological sex estimation using 200 cranial CT scans of Indonesian individuals. According to the study results, the most accurate deep learning network, which learned to estimate sex and cranial traits as an auxiliary task, achieved a classification accuracy of 97%. Humans estimating biological sex using the Walker standard achieved an accuracy of 82%.
Data indicated that the deep learning model focused on certain cranial traits—like the Walker standard— but also considered overall size, shape and other subtle variations in cranial morphology. Additionally, all AI networks reported higher accuracies when compared with a human observer when using just the five Walker traits on the test dataset.
“The inclusion of the general morphology of the skull in the networks is likely one of the contributing factors to the enhanced performance over the human observer who was restricted to the Walker traits in this study,” the authors write in their study. “The general size and shape of the skull are key features that reflect sexual dimorphism in human populations, with male skulls being overall larger and more robust than females.”
Interestingly, both the human observer using Walker traits and the deep learning networks using the skull as input showed a bias toward misclassifying males in the Indonesian dataset.
The deep learning models were also found to be faster when compared to their human counterparts.
“Our AI tool produces its results approximately five times faster than humans can, meaning families waiting for results of investigations can receive news about their loved ones more quickly,” said Min.
While the study helped address some of the perceived limitations of traditional methods while better accounting for diversity in forensic data, Min says future research is needed “around expanding datasets to include diverse populations, enhancing the robustness and generalizability of the AI framework.”
The CSIRO team is now looking for industry collaborators to develop and translate this technology for real-life applications.
The AI tool was developed in collaboration with The University of Western Australia, whose forensic anthropology experts provided labeled data and domain knowledge to support model development. The CT database was collected at Dr Wahidin Sudirohusodo General Hospital at Hasanuddin University in Indonesia.