By Clinical Informatics News Staff 

December 18, 2014 | IBM and Memorial Sloan Kettering announced yesterday a project to apply cognitive computing to the problems of melanoma diagnosis.

IBM’s cognitive visual computing can be trained to identify specific patterns in images by gaining experience and knowledge through analysis of large collections of educational research data. The technology can be applied to any images—satellite images, medical images, natural images, even videos.

By applying the technology to dermatology, IBM and MSK researchers hope to more accurately diagnose skin cancer, said Noel Codella, a research scientist in the multimedia analytics group at IBM.

The IBM system uses a variety of approaches—different machine learning approaches and different algorithms—to determine how well each approach works and what combination might work best. “The system is going to try many algorithms and evaluate how well each one of them work, and how well different combinations of them work,” said Codella.  

To train and test the system, IBM needed an educational dataset. Memorial Sloan Kettering provided a training dataset containing over 3,000 cases of melanoma, atypical lesions, and benign lesions.

The images in the dataset are all dermoscopy images, a special imaging technique in dermatology designed to minimize color, angle, and light distortion. The visual computing system scans the dermoscopy images for variations in color, shape, density, texture, and pattern and predicts which samples are likely diseased and which are likely benign.

So far, the system as proven very adept. “We’re seeing some very encouraging results,” Codella said. “Our system is getting 97% sensitivity and 95% specificity.”

But there’s definitely room to improve, Codella said. In the U.S. each year there are nearly 5 million people treated for skin cancer, and melanoma causes 9,000 deaths a year. The next step is to grow the dataset, improve the recognition performance of the system, and incorporate other types of data analysis.

“For example,” he said, “the way the system works right now is that within the educational database of images, doctors have isolated images that represent melanoma, and images of lesions that are not melanoma. Our cognitive system basically looks over these images and tries to determine what type of automated approaches are going to best discriminate between these two classes… Moving forward, there’s a lot of additional information that could be brought into the analytics framework.  For example, how does a lesion change over time? That can be a very good indicator of whether something is melanoma or not based on how it grows. How does the lesion compare to other lesions on the same patient?”

Codella said that the IBM team’s work with Memorial Sloan Kettering goes beyond just collecting test sets. Different doctors have different approaches to diagnoses. “This is really an effort to standardize the field, both in terms of the educational content, but also what are the current practices, the terminology to describe the disease and its features.”  

The research isn’t yet ready for the clinic. “It’s our hope that the research we do is going to lead to technologies that improve clinical practice,” Codella said, “but right now it’s still in the early stages of research. It’s going to be a little while before you see what we’re working in your dermatologist’s office.”