PREVIEW

Scientists are using AI to unlock the mysteries of the fossils that tell us huge amounts about the Earth’s prehistoric past.

Queensland Museum and James Cook University scientists have developed an AI-based technique that speeds up the analysis of fossils from months to just days, as outlined in their recent scientific paper Accelerating segmentation of fossil CT scans through Deep Learning.

Queensland Museum palaeontologist and JCU Senior Lecturer Dr Espen Knutsen, alongside JCU Deep Learning expert Senior Lecturer Dmitry Konovalov have been working on how to rapidly analyse fossils encased in rock.

“Computed Tomography (CT) scanning provides palaeontologists with a way to look inside bone and study fragile fossil material without the need for physically removing the surrounding rock,” Dr Knutsen said.

Dr Knutsen said the CT datasets consist of a stack of x-ray image slices which are imported into a computer - but then need to be manually told what parts of each slice is fossil and what is rock before it can produce a 3D model.

“With ever-improving equipment, the size of datasets and image resolution have significantly increased, which means a significant amount of time invested in manually segmenting data,” Dr Knutsen said.

“With datasets often exceeding 2000 images per sample, this process can take months to complete.”

Instead, the scientists manually segmented around two per cent of 2000 image slices and used these to train a Deep Learning model which completed the task by itself.

“We achieved a highly precise 3D representation of a tiny Triassic reptile from Queensland that was around 240 million-years-old. This was completed in days rather than months,” Dr Knutsen said.

The research removes a huge bottleneck in modern palaeontology, allowing researchers to rapidly generate more data from fossil collections - allowing far easier study of each fossil’s secrets that will enable us to understand the dinosaurs and other animals that roamed our planet millions of years ago.

The researchers will now work to expand the capability of their Deep Learning model to be used on other and more diverse fossil material.

The paper Accelerating segmentation of fossil CT scans through Deep Learning was recently published in Scientific Reports