Virtual Anthropology

My research projects tend to involve analyses of anatomical variation using different forms of CT scanning, mostly of bone, but increasingly other tissues like cartilage and muscle.

Machine Learning approaches to fossil segmentation

Fossils are often damaged and can contain all kinds of inclusions such as breccia, mud, or rock. Dealing with these inclusions is a major challenge when we want to study internal bone structure with Micro CT scans. Until recently, the only way around this problem was to segment each image manually by manually labelling each pixel as either 'bone' or 'other material'. This is a major task with around 2000 images per fossil!

To overcome this issue I have developed deep learning models that can automatically separate bone from non-bone, even if they have the same pixel density values. Even better, these models can also accurately classify spongy trabecular bone from dense cortical bone.

Here is an example of the deep learning model in action.

This is a slice of a Micro CT scan of STW311, a femur assigned to Australopithecus africanus. This fossil has been damaged, parts of the cortical and trabecular bone have been broken off and there are all kinds of dark and bright non-bone inclusions.

Standard Otsu segmentation does not do a good job separating bone from other materials. The deep learning approach on the other hand accurately identifies cortical and trabecular bone and excludes any non-bone materials.

The models are regularly updated with additional fossil and extant data and the latest versions are freely available on my Github page.

Trabmap: mapping bone structure throughout whole bones

Researchers have long been interested in measuring how trabecular bone tissue is distributed within whole bones. Regional variation in bone structure can tell us a lot about the stresses that acted on a bone during life, and thus help us reconstruct the habitual behaviours of extinct animals.

Methods to calculate the distribution of trabecular bone properties throughout whole bones are either stuck behind a paywall or calculate properties incorrectly. So I decided to write my own programme, published as an R package called trabmap.

Trabmap can be used to quantify and visualise the distribution of trabecular bone properties throughout whole bones. Here is an example of the distribution of bone volume fraction in the distal humerus of Australopithecus sediba (MH2) ranging from high (red) to low (blue).

The trabmap R package is freely available on my Github page.

The evolution of speech

Vocalisations are produced by vibrating vocal folds in the larynx (voice box). While all primates can produce sounds, only humans can produce speech.

Using a combination of CT scanning, experiments with excised larynges, and mathematical modelling we were able to show that humans are able to speak in very complex and controlled combinations of sounds thanks to a simplification of our vocal anatomy.

We are the only primate that have lost their vocal membrane - a projection of tissue at the top of the vocal folds. Experiments and mathematical simulations show that having fewer moving parts in the larynx means that humans have much greater control over their vocal production.

A wonderful example of how a simplification can lead to greater complexity

Nishimura et al. (2022) Science

Pointclouds - a method to statistically compare variation in bone structure throughout whole bones

Everyone's bones are slightly different in size and shape. This is a makes it very difficult to compare the internal trabecular structure between different individuals. We have developed a new method that circumvents this issue by transforming 3D models of bones, including their internal structure, into an average shape that can be analysed statistically.

Demars et al. (2020) American Journal of Human Biology

The top represents the average of a sample of foragers while the bottom represents the average of a group of agriculturalists. The foragers appear to have denser bones but the differences vary regionally. On the right we mapped the areas where the trabecular bone is significantly denser in the foragers compared to the agriculturalists.

Bone Ratio Predictor (BRP) - an automated method for comparing low resolution pQCT and high resolution μCT

The only way we can make confident interpretations about the lifestyle and behaviours of people in the past based on the size, shape, and density of their bones, is to compare them to living people for whom we know things about their diet, behaviour, age, and sex. That is because all of these factors can influence bone structure. Sounds easy enough, but there is a problem..

..the bones of living people cannot be CT-scanned at the same resolution as dry bone. This is because greater resolution requires dangerous amounts of radiation. The picture on the left shows the same bone scanned with high resoltion uCT and low resolution pQCT. So how can we compare data that we obtain from living people, and other animals for that matter, imaged at low resolutions, to high resolution scans of dry bone? In this paper we present an approach to do just this.

This method allows us to interpret the variation in bone structure that we find in the archaeological and fossil records, in light of the variation that we see in living people today.

Saers et al. (2021) American Journal of Biological Anthropology

Page updated

Google Sites

Report abuse