June 27, 2016 | Written by: Rahil Garnavi
Categorized: IBM Research | IBM Watson
Share this post:
Australia has the highest incidence of melanoma in the world, accounting for 80 percent of all newly diagnosed cancers in the country. According to the Melanoma Institute Australia, an Australian will die from melanoma every six hours. With these figures it’s not surprising that melanoma is often called Australia’s national cancer.
Clinicians do not always have the skills or the instruments to accurately recognize in situ melanoma, and people often need to wait weeks before seeing experienced dermatologists. With the early-detection window being vital, there is a real need to develop new technologies to help early detection of melanoma. In my work, back to my PhD research at The University of Melbourne to develop a computer-aided melanoma detection system, I’ve seen the severe, detrimental consequences when it’s detected too late.
When detected at an early stage, treatment can often result in a complete recovery. The longer the disease goes untreated, the survival rates drop significantly, which is why early detection is vital.
Over the past few years, my team at IBM Research – Australia, along with colleagues at our Thomas J Watson Research Center, have collaborated with world-leading cancer centers in the US to apply machine learning techniques to medical imaging, and in particular, images of skin lesions. We have created algorithms that can analyze dermatology images and determine, to a given degree of confidence, if there is a presence of melanoma within the image.
Mapping moles with data
Our newly-signed separate research agreements with two leading melanoma organizations, MoleMap and Melanoma Institute Australia, to apply cognitive computing to dermatology images, have the potential to help clinicians with earlier and more-accurate identification of skin cancer. This can lead to earlier treatment, and potentially save more lives.
But machine learning needs data. The more training data ingested by our deep learning algorithms, the more accurate the analytics will be. The agreements mean that our team will be able to refine and further develop our analytics; algorithms, that were built on tens of thousands of medical images, will be improved by ingesting more than a million new dermoscopy and clinical skin images. The algorithms will be trained to identify specific dermatological patterns in images and to identify various skin disease states.
We have had promising early results so far, and will continue the focus on improving and further developing our analytics. This means training our algorithms to more quickly and accurately identify the presence of melanoma and other types of skin cancer through combining images of different modalities (dermoscopy and clinical), as well as textual information. The end goal of this work is to create advanced cognitive visual analytics that can support clinicians to make more informed diagnoses and save lives. I am confident that these collaborations are a big step in that direction.