Modeling Medical Advancements With AI
Keri Olson describes how Power Systems technology is making medical advancements possible.
By Keri Olson06/01/2020
In 2006, I was diagnosed with melanoma, the most aggressive and deadly form of skin cancer. While this was a stressful and frightening time, I was fortunate that it was found early. Through diagnostic testing, it was determined the cancer was locally contained to a spot on my left shoulder and had only progressed to Stage II. The spot was removed with clean margins, and I am happy to report that I have been cancer free for nearly 15 years.
Melanoma was responsible for 7,230 deaths in the United States in 2019. With fair skin and blue eyes, I was at higher risk of developing skin cancer, but I also had a greater chance of detecting and treating it at early stages. African-Americans have a significantly lower risk of melanoma than Caucasians, but the mortality rate for African-Americans in the United States is higher, primarily due to misdiagnosis. Studies also show lower survival rates for melanoma among other ethnic groups, including Hispanics, Asians and Pacific Islanders.
Melanoma is unique in that it typically arises as a visible mark on the skin’s surface, unlike many other cancers that develop hidden from patients’ view. This suggests that computer vision, which has demonstrated human equivalency in visual recognition tasks, would be well suited to aid in its early detection.
Real Data for AI Models
The biggest roadblock to date for AI vision-based skin cancer diagnosis has been the lack of large, well-designed, public data sets of skin images with requisite metadata to train systems for detection. The greater the number of images used to train visual detection models, the greater the accuracy of the model.
The International Skin Imaging Collaboration (ISIC) is addressing this need though the creation of a large, open-source archive of high quality, annotated skin images. At present, the ISIC Archive contains over 13,000 images, including more than 1,000 images of melanomas, with a long-term goal of housing millions of images for many uses including the development of computer vision algorithms for skin cancer detection.
Using the ISIC data set, IBM Research teamed with Memorial Sloan Kettering Cancer Center, Emory University and Kitware Inc. in 2016 to launch the first international melanoma image detection challenge. This challenge was designed to help participants develop image analysis tools to enable the automated diagnosis of melanoma. In this and subsequent challenges, the average performance of dermatologists has been surpassed by machine learning fusion algorithms.
AI for Good
IBM’s legacy of providing ground-breaking technology for the betterment of humankind continues in other ways as well. For example, IBM is aiding scientists to fight against the COVID-19 pandemic. Researchers at the Department of Energy’s Oak Ridge National Laboratory used the Summit supercomputer, powered by POWER9™ servers, to simulate 8,000 compounds in a matter of days. The outcome is a model that could impact the COVID-19 infection process by binding to the virus’s spike. They have identified 77 small-molecule compounds that have shown the potential to impair COVID-19’s ability to dock with and infect host cells.
IBM researchers took that data and found that using IBM’s Bayesian optimizer, which invokes state-of-the-art AI optimization algorithms, they could identify the most promising drugs after screening only 500, instead of hundreds of thousands. This kind of acceleration will translate to saving lives because treatments can be identified sooner.
I am truly grateful to IBM, our clients and our partners for their contribution to scientific discovery to improve health outcomes worldwide.
Keri Olson: Director, HPC and Analytics Offerings, IBM