A new artificial intelligence system developed by researchers at Adelaide University could help doctors diagnose skin diseases faster and more accurately, particularly for rare conditions that can be difficult to identify.
The technology, named DermaViGNet, combines two machine learning approaches to analyse images of skin conditions while also showing clinicians why it has reached a particular diagnosis.
“Explainable AI plays a crucial role in clinical adoption to ensure model interpretability,” said Adelaide University co-researcher Dr Sabbir Ahmed, from the School of Computer Science and Information Technology.
“Many AI systems will share a recommendation or decision but cannot share the reasoning behind it, which can be an unacceptable risk in high-stakes fields such as health.
“Our technology incorporates explainable AI techniques by highlighting the specific regions of an image that influenced a diagnosis. This transparency can help clinicians better understand, verify and trust AI-generated results before incorporating them into patient care.”
The model was trained and tested using just under 10,000 dermoscopic images, covering common and underrepresented skin conditions.
Researchers then compared its performance against eight existing deep learning models, with DermaViGNet outperforming all of them.
The system achieved an impressive 98 per cent accuracy when classifying five conditions, including vitiligo, acne, nail psoriasis, hyperpigmentation and Stevens-Johnson Syndrome-Toxic Epidermal Necrolysis (SJS-TEN), a rare but potentially life-threatening skin reaction.
While the technology isn’t intended to replace doctors, researchers say it could become a valuable diagnostic aid, particularly in places where access to specialist dermatology expertise is limited.
“Our technology has the potential to support healthcare professionals by providing a fast and reliable diagnostic aid, particularly in settings where specialist dermatology expertise may not be readily available.
“It may also help reduce diagnostic errors and improve outcomes through earlier detection and treatment of serious skin diseases.”
One of the major challenges the research aims to address is the lack of rare skin conditions represented in many medical AI datasets. By incorporating a broader range of conditions, researchers say the model maintained high accuracy while improving its ability to recognise different diseases.
“This is an important step towards more equitable and clinically useful AI-powered healthcare tools,” said Dr Ahmed.
The team will now focus on expanding the dataset, incorporating additional clinical information and refining the technology before it can be considered for use in real-world clinical settings.
To read the full research paper, click here.
