Training Watson to Help Detect Melanomas Earlier and Faster

Share this post:

Kiwis and Aussies love the outdoors, especially in the summer sunshine. Unfortunately, that has resulted in the area having one of the highest rates of skin cancer in the world. With melanoma being one of the most common forms of cancer, early diagnosis is critical to improving survival rates. That’s why I teamed with IBM Research: to improve the identification of melanoma. With the work, we hope to help save the lives of more than 3,500 people each year across Australia and New Zealand.

Finding melanoma is like finding the needle in a huge haystack. We may have 10s or 100s of moles, freckles and other spots on our skin and it is most likely that none of them are melanoma. Once a melanoma reaches more than 1mm below the surface of the skin it becomes life threatening. The challenge is that these changes can be difficult to spot, as superficial melanomas generally grow slowly, spreading across the skin for months or even years before they present a serious threat.

If detected before it reaches the lymph nodes the chance of surviving for a further five years rises from 64 per cent to 95 per cent.

To save more lives our dermatologists need to see more people! Despite the widespread education campaigns to ‘slip, slop, and slap,’ early detection has not been a public health priority. Not all primary care providers are sufficiently trained to undertake increased screening and there simply aren’t enough dermatologists to cover the whole country.

I estimate one million people throughout Australia and New Zealand should be screened regularly but we only see around 50,000 each year.

The critical work we are doing with IBM Research is designed to train cognitive technology to identify melanoma from its very early stages. Rather than taking over the specialist’s role of identifying suspicious skin lesions, Watson will efficiently and effectively sift out the false alarms. Making the ‘haystack’ smaller will help our expert dermatologists to focus on, and spot, the dangerous lesions.

The first step was to train Watson to help specialists spot skin lesion changes. Watson learnt about three types of skin cancer and 12 benign disease groups from 40,000 images taken from our database of 15 million lesion images and compared that with our expert’s medical diagnosis.

Like a human, the machine’s accuracy with detecting melanoma improves with practice. My brief to IBM Research was to get to a level of accuracy of 80%, similar to what the average dermatologist (they are the experts) achieves. The results from the research so far show a level of accuracy of closer to 95%[1], which is really encouraging.

I’m really excited about the scope of what Watson can help us do. I want this Watson-enabled technology to reach the wider population, especially regional areas where people are more likely to spend their days working in the sun but are further from our specialists.

__________________________________________

[1] Z. Ge, R. Chakravorty, B. Bozorgtabar, A. Bowling, R. Garnavi, et al.”Exploiting Local and Generic Features for Accurate Skin Lesions Classification Using Clinical and Dermoscopy Imaging”, to appear in the proceeding of The 2017 IEEE International Symposium on Biomedical Imaging.