IBM is developing a platform that uses deep learning to diagnose melanoma from images of skin lesions in collaboration with dermatologists from Memorial Sloan Kettering Cancer Center.
Over the past few years, we’ve seen IBM dive into healthcare including the recent launch of Watson Health. From helping guide cancer therapy decisions to simply collecting & collating unbelievable amounts of health data, IBM has undertaken a number of exciting endeavors.
In this work, IBM researchers describe development of a platform that uses deep learning technology to analyze images of skin lesions and identify those that could be melanoma. The technical details of how they accomplished this are described in detail in a paper recently published online by IBM in detail that goes way beyond my understanding of this space.
To develop the system, they used a database of about 1300 dermascope images of skin lesions. Among those images, roughly 20% were melanoma. Most of the images (80%) were used to train the system while 20% were set aside for validation. Testing a number of different approaches, they described different combinations of sensitivity, specificity, accuracy, and precision. Highlighting one of the key challenges with this kind of work, it doesn’t seem like there was necessarily a “best” approach – an improvement in one area often was at the detriment of another area.
They also did a comparison with the assessments of 8 dermatologists looking at a set of 100 images (half melanoma) and reported that their system was more accurate than the dermatologists. At the sensitivity achieved by the group of dermatologists (82%), the IBM system had slightly higher specificity (62% vs. 59%).
There’s clearly a lot more work to be done before this kind of technology is ready for prime time, including more rigorous validation on a more representative dataset (not “enriched” for melanoma and more reflective of real world practice). That said, its certainly got some exciting potential help more effectively screen large groups of patients and potentially identify lesions earlier, before they get a lot harder to treat.