- Artificial intelligence performs as well as dermatologists at screening for skin cancer
- Recent advances in machine learning contribute to multiple healthcare breakthroughs
- Mobile app may lead to earlier cancer detection and improve survival rates
Imagine that you’re taking a shower, and you notice an odd-looking mole on your arm. It might be nothing. Or it might be a malignant melanoma, one of the deadliest forms of skin cancer.
You should probably go to a doctor and have it checked out. But what if you live in a remote area and the nearest clinic is three hours away. Would you make the journey? Or would you just ignore it and hope for the best?
Someday soon there may be a better choice. Andre Esteva, a PhD candidate in the Stanford Artificial Intelligence Laboratory has trained a deep convolutional neural network (CNN) to screen skin lesions (such as moles) and classify them as cancerous or benign.
In Esteva’s research, the CNN performs as accurately as board-certified dermatologists and can identify lesions from photographs, which may eventually lead to a mobile app that could make an initial classification without the need for a doctor’s visit.
“I've always believed that AI could have a profound impact in healthcare, where massive data is constantly being generated,” says Esteva. “Dermatology has been a great use case, thanks to the visual nature of screenings, the maturity of computer vision algorithms, and the rising ubiquity of mobile phones.”
Although melanomas represent fewer than five percent of all skin cancers in the US, they account for approximately 75 percent of all skin-cancer deaths, totaling over 10,000 deaths a year.
Early diagnosis is critical, with the survival rate dropping dramatically if detection is delayed. But if scientists can put the tools for early identification in the hands of patients — especially those with limited access to healthcare, or even those with an aversion to doctor’s offices — survival rates can increase.
Esteva and his team began with a GoogleNet Inception v3 CNN architecture that was pretrained for image recognition. They then fed the network an additional 127,463 dermatologist-labeled images sourced from several clinician-curated open-access repositories, and trained it using transfer learning.
When the network was tested on its ability to identify both melanomas (the deadliest form of skin cancer) and keratinocyte carcinomas (the most common), it performed as well as 21-board certified dermatologists, with an accuracy rate of over 70 percent.
Esteva acknowledges that a dermatologist’s classification of a lesion and subsequent recommendation for treatment involves much more than a visual screening. “We have built an algorithm, in collaboration with dermatologists, to extend their reach of care regarding visual screens.”
Deep learning renaissance
Melanomas account for approximately 75 percent of all skin-cancer deaths, totaling over 10,000 deaths a year.
Building on these advances, Esteva anticipates that his work could also prove useful in radiology, pathology, and other areas that involve medical imaging.
“There are a myriad of different applications in healthcare that I see becoming feasible as the field matures,” Esteva says. “I'm particularly excited by unsupervised learning techniques, and improved algorithms for working with unstructured data, of which there is vastly more than structured data.”
So, until that mobile app becomes available, when you spot a funny-looking mole, go to the doctor. Because you just never know.