IBM's computer vision research zeroes in on identifying skin cancer

IBM's methods for identifying skin cancer in its tested data sets have performed as well as medical specialists.

After years of working with New York's Memorial Sloan-Kettering Cancer Center, IBM is using its cognitive computing power to zero in on a specific, all-too-common form of cancer: melanoma.

Playing House: How IBM's Watson is helping doctors diagnose the most rare and elusive illnesses

Marburg Hospital's Centre for the Rare and Undiagnosed Diseases is using the cognitive computing system to solve some of the most complex medical cases.

Read More

A team of IBM research scientists has been testing computer vision techniques that could one day give doctors better tools to diagnose skin cancer. In a recently submitted report, the team is announcing that their system is now as good at recognizing melanoma as a medical specialist.

Given how deadly late-stage melanoma is, giving patients and doctors access to better diagnostic tools could have a major impact on public health and health care costs, Noel Codella, a cognitive computing and computer vision researcher at IBM, told ZDNet.

"Doctors can catch melanoma early, but it's costly to see your dermatologist, and dermatologists are not always accessible in all regions," he said. "Even if you do see a dermatologist, they are prone to errors at times."

Dermatologists use an imaging technique called Dermoscopy to detect the disease in its early stages, but they tend to be overly cautious: up to nine lesions are surgically biopsied for every one melanoma discovered. Yet if the disease is caught early, patients who receive proper treatment have a five-year survival rate higher than 98 percent. If the disease progresses to the lymphatic system or beyond, the survival rate drops as low as 16 percent.

IBM's vision is to use computer vision to analyze dermoscopic images for medical professionals: A doctor or nurse would use a Dermoscope attached to a smartphone or other camera to take a picture of the patients' lesion, send it to a cloud-based analytics service and get back a detailed report on it. The report could include information like a confidence indicator to help the doctor determine whether the patient has melanoma. It could also point out things like visual patterns on the lesion that could help determine a patient's risk.

Codella likened the tool to a blood test. In other words, it would be "a tool that helps a doctor provide a diagnosis, not a tool that provides the diagnosis in itself."

The IBM research team is still improving their algorithms and working to determine at what point their system could be clinically useful. There's currently "no solid timeline" for developing a product, Codella said. First, the team needs more data.

As Codella pointed out it a blog post, the IBM system -- even though it already performs as well as a clinical specialist -- has seen fewer than 3,000 lesions. That gives it the experience level of a medical student.

The IBM research team has trained and tested its system with a data set provided by the International Skin Imaging Collaboration (ISIC) -- the largest publicly available benchmark data set of dermoscopic images.

While it's easy to envision a tool that patients could one day use to assess skin lesions themselves, Codella said "there are a lot of concerns" that need to be addressed before you put diagnostic equipment in the hands of consumers.

"You don't want to put a tool out there that unnecessarily calms them," he said. That being said, he added, there are already apps that help patients keep track of skin lesions so they can relay the information to their doctors.

"There potentially can be a lot of benefit from consumer applications," Codella said. "At the end of the day, I think the field needs to try to figure out what is the best type of service that need to be provided that can help complete the workflow of clinical care while at the same time minimizing harm that could be done."

Doctors want to use IBM's supercomputer to diagnose health conditions