Early detection can mean a world of difference for patients diagnosed with skin cancer. In fact, those lucky enough to detect skin cancer early achieve a five-year survival rate of 99 percent. Resource or time constraints often keep people from heading to the doctor early, however, which is why two developers decided to take the world of cancer screening into their own hands and create a free, artificial intelligence-powered screening program completely online.
Peter Ma, an independent developer and part of the Intel Software Innovator Program, along with co-founder Mike Borozdin, has developed an AI solution that has the power to determine and classify skin cancer types with the same level of intelligence as a dermatologist. The technology, known as Doctor Hazel, uses deep-learning neural networks to screen and classify skin cancer with 80 percent accuracy.
“Doctor Hazel uses real-time images from the endoscope camera, while simultaneously pulling from 8,000 variables to determine four different outcomes for the mole in question: a normal mole, melanoma, another form of cancer or nothing,” says Ma.
Moreover, it can be used by anyone, anywhere, at any time — even by those without an internet connection — thanks to the use of Intel’s Movidius Neural Compute Stick, an edge device that's able to classify cancer images in real time and offline, making image classification much more interactive, says Ma.
“Anyone can go to the website, upload a photo of a concerning mole, and get results within seconds. If the AI feature detects the mole is cancerous, an individual will then receive a recommendation to see a doctor for further testing,” Ma continues. “Because of the edge device running offline, we are able to use the device in places where internet access is scarce.”
Edge, Cloud Make Healthcare AI Possible
Doctor Hazel, which is still in beta, was inspired by the death of Ma’s close friend, who passed away in his early 30s from cancer, prompting Ma and Borozdin to dig into how they could use their skills to improve cancer prevention and screening.
“At the time of prototype, I was learning and deploying AI on the edge. The AI is very good at classifications, and skin cancer was a perfect use case for it,” says Ma.
The project officially began in 2017 as an idea to be presented at TechCrunch Disrupt’s hackathon, cobbled together from several different technologies. Ma and Borozdin then trained Doctor Hazel on over 10,000 images on the Intel AI DevCloud, sourcing them from the International Skin Imaging Collaboration project, University of Iowa and many other places, says Ma.
“We got to around 80 percent accuracy. To increase that we would need more data,” says Ma, who has a goal of reaching 90 percent accuracy.
Next, Ma and Borozdin are seeking to commercialize Doctor Hazel “so that initial screening can be done at either a primary care's office or at the pharmacy,” says Ma. However, the pair are facing several roadblocks in doing so, from acquiring a massive amount of data to getting Food and Drug Administration approval.
Despite these challenges, Ma sees a bright future for the technology and AI in healthcare, particularly as technology evolves that can power AI from the cloud edge and enable classification in near real time, as with Doctor Hazel.
“I believe the AI classification will power many other healthcare capabilities in the future, both reducing time for the physicians and improving quality of care for the patients,” says Ma.