It’s still relatively rare for artificial intelligence to deliver a crushing victory over human physicians in a head-to-head test of medical expertise. But a deep neural network approach managed to beat 42 dermatology experts in diagnosing a common nail fungus that affects about 35 million Americans each year.
The latest successful demonstration of AI’s capabilities in the medical field relied heavily upon a team of South Korean researchers putting together a huge dataset of almost 50,000 images of toenails and fingernails. That large amount of data used to train the deep neural networks on recognizing cases of onychomycosis—a common fungal infection that can make nails discolored and brittle—provided the crucial edge that enabled deep learning to outperform medical experts.
“This study was the first to show that AI has overwhelmed the specialists,” says Seung Seog Han, a dermatologist and clinician at I Dermatology in Seoul, South Korea. “Until now, in many studies, AI was similar to that of a specialist in diagnosis of diabetic retinopathy, diagnosis of skin cancer, and chest X-ray readings.”
Past testing involving AI versus doctors in dermatology has typically shown AI to be roughly on par with human expertise. But in this study, just one dermatologist out of 42 experts did slightly better than the deep neural networks in one out of three trials during a specific test scenario. The research appears in the 19 January 2018 online issue of the journal PLOS One.
Notably, the deep neural networks performed much better than the dermatology experts on especially difficult cases as opposed to easier cases, Han says. Besides Han, primary researchers in the group included Gyeong Hun Park, a professor of dermatology at Hallym University, and Sung Eun Chang, a professor of dermatology at Ulsan University.
Han’s day job as a physician involves treating a wide variety of skin conditions. But he has also learned a number of computer programming languages such as C++ and Python over the years. And he took an interest in deep learning after seeing news about AlphaGo, DeepMind’s self-taught Go-playing AI, defeating the world’s best human players such as South Korea’s Lee Sedol.
Deep learning algorithms often prove a good fit for tackling specialized tasks that rely upon detecting patterns humans might miss in large amounts of data. In this case, the South Korean researchers saw an opportunity to use deep learning algorithms developed by Microsoft Research to help physicians identify possible cases of nail fungus infections from digital photos.[shortcode ieee-pullquote quote=""This study was the first to show that AI has overwhelmed the specialists."" float="left" expand=1]
But any deep learning model needs a lot of data to train the AI in recognizing related patterns. Collecting usable images related to cases of nail fungus infection represented a huge challenge, because there is usually no standard formatting for the images. Many pictures are taken from a wide variety of angles and show a mix of both healthy nails and nails with onychomycosis. Furthermore, all images needed to be resized to 224 x 224 pixels because of technical limitations in the deep learning algorithms being used, which could make many pictures unrecognizable.
Han and his colleagues trained an object detection algorithm called Faster R-CNN to identify and crop images so that they only included affected toenails or fingernails, and then enlarged the resulting images to provide a dataset that could train the deep neural networks. Most of the images came from MedicalPhoto, a clinical photo management program for dermatology that Han developed back in 2007.
Still, Han had to go through a collection of 100,000 photos generated by that R-CNN cropping process to manually read and tag each photo twice to make sure the training data was accurate: incorrect or inadequate nail images were tossed out. That cost him about 550 hours of work over 70 days, even when he was processing one image every 10 seconds for hours each day.
That dataset helped train additional convolutional neural networks—Microsoft’s ResNet-152 and Oxford University’s VGG-19 model—to perform the actual job of identifying possible cases of nail fungus infections. That combined deep learning approach outperformed the group of 42 dermatology experts, which included 16 professors, 18 clinicians, and eight medical residents.
The researchers also showed that deep learning generally outperformed the five best dermatologists in additional testing. And for good measure, they found that the AI’s diagnostic assessments also proved superior to those of general physicians, medical students, nurses, and non-medical personnel.
The research team has released an early demo version of their deep learning approach that anyone can try out through a website or by downloading the app for Android smartphones. By collecting data through the website and app, the researchers hope to identify possible problems that could still arise if the AI was used in real medical practice.
Han and his colleagues have also been testing deep learning in tackling other skin diseases such as skin cancer. A related paper by the group was published in the 8 February 2018 online issue of the Journal of Investigative Dermatology.
Such research shows that AI could prove especially helpful in telemedicine cases that rely more heavily upon clinical photography to diagnose certain conditions such as nail fungus infections. Still, human dermatologists will be needed to confirm such diagnoses using the patient’s general medical history and a broader set of factors such as foot odor; few would necessarily feel comfortable making a diganosis based on pictures alone.
Han and his colleagues think their research could prove especially useful for general practitioners who may see patients complaining of related nail and skin conditions. “The diagnosis of AI is much more accurate than the general clinical diagnosis, and I think it will be helpful for the general practitioner to determine the direction of onychomycosis treatment,” Han says.
Jeremy Hsu has been working as a science and technology journalist in New York City since 2008. He has written on subjects as diverse as supercomputing and wearable electronics for IEEE Spectrum. When he’s not trying to wrap his head around the latest quantum computing news for Spectrum, he also contributes to a variety of publications such as Scientific American, Discover, Popular Science, and others. He is a graduate of New York University’s Science, Health & Environmental Reporting Program.