Skip to content

28 February 2022

Training the Algorithms

Image of Artrya Salix on a Desktop

How we developed Artrya Salix to illustrate imaging biomarkers of coronary artery disease

High-quality data was the essential factor for the computer scientists who produced Artrya’s world-class artificial intelligence solution, Artrya Salix.

Artrya Chief Technology Officer Dr Julien Flack says training a world-class algorithm relies on access to superior datasets and reporting.

“It’s the data that is the critical part,” he says. “The data is the most important part of it for us. It has been since the start.”

Expert-read Coronary Computed Tomography Angiography (CCTA) scans and assessments are the foundation of the Artrya Salix solution to non-invasively assess an individual’s coronary artery disease status.

The AI solution employs ten different deep learning algorithms to navigate CCTA scans and pinpoint coronary artery disease biomarkers, notably including the presence of high-risk plaque features such as low attenuation plaque.

To develop their cutting-edge medical technology, Artrya developed partnerships with leading medical imaging providers that allowed access to CCTA data.

A leading Australian clinic provided Artrya’s data team access to more than 20,000 CCTA scans and datasets. Further, an internationally recognised heart-health institute in Canada made available a selection of their images and expert-read reports. Ethics approval was submitted for all data used.

Over three years of development, the scans and associated data were used to train, test and validate the Artrya Salix deep learning and machine learning algorithms.

Dr Flack says the Artrya team used the medical images as a foundation, then scaled up an internal team to provide expert annotations on the medical images. They did this in order to train the tools using the highest possible quality data.

In addition, Artrya recruited Chief Medical Officer Professor Girish Dwivedi, who brings excellence in cardiology and advanced multimodality imaging training to the organisation.

Professor Dwivedi is the inaugural Wesfarmers Chair in Cardiology at The University of Western Australia’s Harry Perkins Institute of Medical Research and a Consultant Imaging Cardiologist at Fiona Stanley Hospital in Perth.

He also holds international and national accreditations in CCTA, echocardiography, cardiac magnetic resonance and nuclear cardiology.

Professor Dwivedi supervised labelling of the interior and exterior walls of the coronary arteries.

Dr Flack says having a global expert assisted the Artrya team to use the highest-quality data in the algorithms.

“That was one of the critical annotation paths we had,” Dr Flack says. “You find there is a little bit of ambiguity about that, one clinician may put the position in one place and another clinician may put it in another place.”

“Under the process we set up, not only did we have to build the tools, but we also had to build the process to allow us to have multiple people annotate the imagery. Then we found where they disagreed. We got expert clinicians to actually go and look at the specific cases and decide which was the best approach.”

“It has been a many-layered challenge to generate this very high-quality data with clinician review.”

Once the data was established and verified, the Artrya team used models to train the algorithms.

The scans, annotated by experts, were split into two or three sets, then trained, tested and refined using computer science techniques of gradient descent and back propagation.

“You get thousands of examples, you train the AI to match those examples from that data and then step back into a real-world scenario,” Dr Flack says.

The algorithms detect the aorta, then map out the coronary arteries, and subsequently their inner and outer walls. Separate AI algorithms incorporate radiodensity analysis and the location information to illustrate disease characteristics.

At this level, it is believed that AI models will achieve the equivalent accuracy as experts in the field for the identification of stenosis and calcification in coronary artery disease.

The solution is one of the few AI-based technologies to assess CCTA images, and notably one with the ability to visualise disease characteristics and guide clinical decision making.

Artrya Salix assesses the radiodensity of Hounsfield Units (HU) in combination with the location, patterning and context of plaques on the coronary artery wall to characterise coronary artery disease.

Dr Flack says the end result is an AI solution that assesses critical imaging biomarkers of coronary artery disease to support a diagnosing clinician, and this has the potential to change outcomes for health systems and patients across the world.

“Over the last couple of years, we’ve built up what we believe to be a world-class database around coronary artery disease, particularly stenosis and high-risk plaque,” he says.