Researchers at Stanford University have developed an algorithm that can classify skin cancers as benign or malignant as well as a dermatologist can, according to paper published Wednesday in Nature.
Melanoma prevalence and diagnosis
According to the researchers, one in five U.S. residents will be diagnosed with skin cancer during their lifetimes. The five-year survival rate for melanoma, which accounts for about 75 percent of all skin cancer-related deaths, is 97 percent if the disease is detected in its earliest stages. The five-year survival rate for melanoma detected in its latest stages is 14 percent.
A visual examination is the first step in diagnosing skin cancer. If the examination's findings are inconclusive or indicate that a lesion could be cancerous, the next step is a biopsy.
About the algorithm
To develop their algorithm, Stanford researchers tweaked a Google-created algorithm that was trained to differentiate cats from dogs. Brett Kuprel, a graduate student at Stanford and co-lead author of the paper, said the researchers made a dataset for skin cancer by finding images of the cancers on the internet and working with the medical school to create a taxonomy.
The researchers tested the algorithm against the performance of 21 board-certified dermatologists. To gauge the dermatologists' skills, the researchers evaluated the dermatologists' ability to correctly diagnose cancerous and non-cancerous lesions in more than 370 images. The researchers measured the algorithm's performance on a sensitivity-specificity curve, looking at three diagnostic tasks:
- Keratinocyte carcinoma classification;
- Melanoma classification; and
- Melanoma classification when viewed using dermoscopy.
Overall, the researchers found that the algorithm's identifications were "on par" with the dermatologists' identifications, Medscape reports.
Andre Esteva, a graduate student at Stanford and co-author of the paper, and colleagues hope the algorithm will allow doctors and patients to track skin lesions proactively and detect cancer early.
The team would like to see the algorithm, which currently exists on a computer, made compatible with smartphones. Esteva said, "Everyone will have a supercomputer in their pockets with a number of sensors in it, including a camera. What if we could use it to visually screen for skin cancer? Or other ailments?"
However, the researchers said the algorithm should be further tested in clinical trials.
Susan Swetter—a co-author on the paper, professor of dermatology, and director of the Pigmented Lesion and Melanoma Program at the Stanford Cancer Institute—said, "Advances in computer-aided classification of benign versus malignant skin lesions could greatly assist dermatologists in improved diagnosis for challenging lesions and provide better management options for patients. However, rigorous prospective validation of the algorithm is necessary before it can be implemented in clinical practice, by practitioners and patients alike."
Separately, Jana Witt of Cancer Research UK said, "It's unlikely that [artificial intelligence (AI)] will replace all of the other information your clinician would consider when making a diagnosis, but AI could help guide ... referrals to specialists in the future" (Scutti, CNN, 1/26; Lowry, Medscape, 1/26; Stanford release, 1/25; Gallagher, BBC, 1/26).
Should radiologists embrace or fear machine learning?
Machine learning and artificial intelligence (AI) technology is gaining ground in medical imaging. For many health IT leaders, machine learning is a welcome tool to help manage the growing volume of digital images, reduce diagnostic errors, and enhance patient care.
Despite its benefits, some radiologists are concerned that this technology will diminish their role, as algorithms start to take a more active part in the image interpretation process while ingesting volumes of data far beyond what any human can do. This briefing discusses how machine learning works, provides examples of how it has helped organizations make strides in imaging, and examines the benefits and challenges of adopting it into an imaging strategy.