A team of researchers led by Monash University have developed a new artificial intelligence model capable of detecting skin diseases faster and with more accuracy.
In a report published in Nature Medicine, scientists unveiled one of the world’s first AI-powered tools designed to support the dermatological medical practice by analysing multiple imaging types simultaneously, delivering accurate and efficient results.
The model, known as PanDerm, can scan close-up photos, dermoscopic images, pathology slides and total body photographs in real time.
The big picture
According to Cancer Australia, there were an estimated 18, 257 new cases of skin melanoma diagnosed in 2023, accounting for a 1 in 17 risk of being diagnosed by the age of 85. While there is a wide number of different skin diseases, Australia has one of the highest rates of skin cancer in the world due to long periods of sun exposure and high UV radiation.
Major causes of skin cancer and melanoma include sunburn, tanning and solariums. According to the Cancer Council, 1 in 8 adults and 1 in 5 teenagers in Australia will report sunburn on an average summer weekend.
Supporting this new method of skin disease diagnosis is one way to help lower the prevalence of melanoma and other skin diseases in Australia.
A new frontier
To train the AI algorithm, scientists input more than two million skin images from four medical categories, alongside data from 11 institutions in multiple countries.
One of the lead co-authors of the research, Associate Professor Zongyuan Ge from Monash University’s Faculty of Information Technology, says that existing dermatology AI models are restricted to performing isolated tasks, such as diagnosing skin cancer from dermoscopic images.
“Previous AI models have struggled to integrate and process various data types and imaging methods, reducing their usefulness to doctors in different real-world settings,” Associate Professor Ge says..
“PanDerm is a tool designed to work alongside clinicians, helping them to interpret complex imaging data and make informed decisions with more confidence and support.”
Unlike the single-task approach of existing models, PanDerm can complete a wide range of clinical tasks including skin cancer screening, predicting the chances of cancer returning or spreading, skin type assessment, mole counting, lesion change tracking, lesion segmenting, and diagnosing a wide range of skin conditions.
When tested, these actions were delivered with just 5-10 per cent of the labelled data typically required for concrete answers.
Siyuan Yan, first author and PhD student from Monash University Faculty of Engineering, says training the system on a variety of image types was key to unlocking its full potential.
“By training PanDerm on diverse data from different imaging techniques, we’ve created a system that can understand skin conditions the way dermatologists do: by synthesising information from various visual sources,“ Siyuan Yan said.
“This allows for more holistic analysis of skin diseases than previous single-modality AI systems.“
Looking forward
Over a series of PanDerm evaluations, it was found that the software increases accurate skin cancer diagnosis by 11 per cent when used
by a doctor. The model also supported non-dermatologist healthcare professionals to improve diagnosis accuracy on other skin conditions by 16.5 per cent.
The AI program could detect skin cancer early, even identifying concerning lesions before clinical detection was possible.
When employed in a clinical setting, PanDerm is intended to be used as a support tool, analysing skin images that doctors routinely use to offer diagnostic probability assessments that support clinicians in interpreting visual data with greater confidence.
One lead co-author of the study and Dermatology Research Centre Director at the University of Queensland, Professor H. Peter Soyer, says the program will help combat the differences in imaging and diagnostics techniques used in clinics based on the availability of resources.
“The strength of PanDerm lies in its ability to support existing clinical workflows,” Professor Soyer says.
“It could be particularly valuable in busy or resource-limited settings, or in primary care where access to dermatologists may be limited.
“We have seen that the tool was also able to perform strongly even when trained on only a small amount of labelled data, a key advantage in diverse medical settings where standard annotated data is often limited.”
Moving forward, the team aim to create standardised protocols for cross-demographic assessments. They want to test the model in a variety of real-world clinical settings. The main focus is determining whether the program performs reliably when presented with different patient demographics or in diverse healthcare environments.
To read more about artificial intelligence, click here.