Diagnosing skin cancer usually starts with a visual exam of the lesion, followed by dermoscopic analysis, then a biopsy if the clinician thinks the appearance of the lesion warrants one, and then histologic confirmation.
Now, researchers have developed a new algorithm that can classify skin cancers as benign or malignant from photographs of the lesions, and they show that the algorithm performs as well as dermatologists.
The research is described in a letter in Nature.
"In the United States, there are only 10,000 dermatologists. But there are 360 million Americans. There are 5.4 million new cases of skin cancer in the US every year. One in 5 Americans will be diagnosed with skin cancer in their lifetime," the researchers, led by PhD candidate Andre Esteva, Stanford University, California, write.
"Although melanomas represent fewer than 5% of all skin cancers in the United States, they account for some 75% of all skin-cancer-related deaths, and responsible for over 10,000 deaths annually in the United States alone. Early detection is critical, as the estimated 5-year survival rate for melanoma drops from over 99% if detected in its earliest stages to about 14% if detected in its latest stages," the authors note.
Esteva and his group developed a computational method that they hope will allow doctors and patients to proactively track skin lesions and detect cancer early.
"This is a type of algorithm, it is called a deep-learning algorithm, that is inspired by a rudimentary understanding of neurons in the brain," Esteva explained to Medscape Medical News.
"Basically, it's composed of a number of layers that are rather simple in the processing that they do, but which, when stacked together, combine massive amounts of data and is really effective at certain key tasks. In this deep-learning algorithm we used 129,450 clinical images, including 3374 dermoscopy images of skin diseases," he said.
In addition to object classification, deep-learning algorithms have been used for natural language processing and speech recognition.
Esteva and his group tested the performance of their algorithm and 21 board-certified dermatologists from across the United States to see how they compared at doing three tasks: recognizing keratinocyte carcinoma vs benign seborrheic keratosis, recognizing malignant melanoma vs benign nevus, and recognizing melanoma under dermoscopy.
"We would present to the dermatologist an image, and we would ask if they would biopsy or treat the lesion, or reassure the patient. We did the same with the algorithm," he explained. Only biopsy-proven images were used.
The results showed that for all three tasks, the algorithm performed on par with the dermatologists.
Universal Access to Healthcare the Goal of This Research
"This project has been driven by two fundamental tenets," Esteva said.
"The first is that we are trying to work towards universal access to healthcare. As a simple proof of concept, given the maturity of computer vision algorithms, we decided to start with skin care," he said.
The second tenet is to extend the reach of healthcare providers outside of the clinic.
"Experts estimate that in 5 to 6 years, there will be over 6 billion smartphone subscriptions worldwide. If you could use your smartphone for healthcare, this would extend the reach of providers outside of the clinic," Esteva said.
He added that he welcomes feedback from the medical community on how best to use this new technology.
"I'm not a doctor, I'm an electrical engineer. But I would like to hear from doctors as to how they would use this algorithm that lets them classify benign vs malignant skin lesions. We hope we have made a bridge between artificial intelligence and medical practice," he said.
Mr Esteva has disclosed no relevant financial relationships.
Nature. Published online January 25, 2017. Abstract
Follow Medscape Oncology on Twitter: @MedscapeOnc
Medscape Medical News © 2017 WebMD, LLC
Send comments and news tips to news@medscape.net.
Cite this: AI as Good as Docs for Diagnosing Skin Cancers From Photo - Medscape - Jan 26, 2017.
Comments