Earlier this year, Sebastian Thrun, an ARTIFICIAL intelligence (AI) scientist, and his colleagues at Stanford University demonstrated that a “deep learning” algorithm could diagnose potentially cancerous skin changes, according to Technologyreview, Accuracy comparable to that of a board-certified dermatologist.
Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include information robot, editing robot, writing robot and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.
Nature reports that the cancer findings were an important part of a series of reports this year that provided an early glimpse into a new era of “software diagnostics.” In this new era, AI can not only help doctors diagnose diseases, but even compete with human doctors. Medical images such as photos, X-rays and MRIs can be matched almost perfectly to the strengths of deep learning software, experts say. In the last few years, deep learning software has made breakthroughs in recognizing faces and objects in pictures.
Many companies have woken up to the opportunity. In December 2016, Verily, the life sciences subsidiary of Google’s parent company Alphabet, teamed up with Nikon to develop an algorithm that can discover the cause of blindness in diabetics. The field of radiology, meanwhile, is also known as the “Silicon Valley of medicine” because it produces so many detailed images.
Black box medical
While Thrun’s team’s predictions were accurate, no one was sure which features the deep-learning algorithm used to classify moles as cancerous or benign. As a result, this has become the medical version of deep learning’s “black box” problem. Unlike traditional visual software, where programmers can define rules in deep learning, algorithms can find the rules themselves, but often leave no clues to explain their decisions.
“In black-box medicine, doctors don’t know what’s going on because nobody knows, and it’s inherently opaque,” said Nicholson Price, a scholar at the University of Michigan who focuses on health law. Price added that it probably won’t do serious harm to the medical field. He likens deep learning to a drug, except that people don’t know where its effects come from.
Take lithium, for example. The exact biochemistry of how it affects mood is unclear, but the drug is approved to treat bipolar disorder. Aspirin is also widely used, but the pharmacology behind it has remained unsolved for more than 70 years. Like them, the black-box issue doesn’t affect the FOOD and Drug Administration, which in addition to approving new drugs also regulates software if it’s developed with the goal of treating or preventing disease, Price said.
In a statement, the FOOD and Drug Administration said it has approved many image analysis applications over the past two decades that rely on a variety of pattern recognition, machine learning and computer vision techniques. The agency also confirmed that we’ll see more deep-learning-enabled medical software, and believes that companies can keep the details of their algorithms secret.
In addition, the FOOD and Drug Administration has given the green light to at least one deep learning algorithm. In January, the agency approved Arterys, a San Francisco medical imaging company, to legally sell its software. The software’s DeepVentricle algorithm analyzes MAGNETIC resonance imaging (MRI) images of the interior contour of the ventricle and calculates how much blood the patient’s heart can hold and pump. The calculation can be done in 30 seconds, compared with an hour in the conventional way.
The FOOD and Drug Administration ordered Arterys to conduct more extensive testing to make sure the results of its algorithm matched doctors’ diagnoses. “Statistically, you need to show that your algorithm is safe and effective for its intended use,” says John Axerio-Cilies, the company’s chief technology officer.
Huge demand
To train its software, the team, led by Thrun, a former vice president of Google’s driverless car team, provided 129,405 expert-evaluated skin images. The images covered 2,032 diseases, including 1,942 images from patients diagnosed with skin cancer. In the end, the software beat 21 human dermatologists in determining which moles were potentially cancerous.
“When dermatologists see the potential of this technology, I think most will choose to support it,” said study author Robert Novoa, a dermatologist at Stanford University. Mr. Novoa and other team members declined to say whether they planned to commercialize the software.
Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.
Allan Halpern, a Memorial Sloan Kettering dermatologist and president of the International Association for Digital Skin Imaging, said: “Any fears that doctors are about to lose their jobs because of AI are overblown. Instead of being a threat, I think algorithms will drive demand for dermatology services.” That’s because a positive screening test also requires a biopsy. Deep learning software has a place in primary care, Dr. Halpern said, but if extensive screening is to be done, or through consumer apps, there may not be enough dermatologists to follow through.
Mr. Selis also said companies could be tempted to offer deep learning tools directly to consumers. For example, people might scan their moles to see if they need to see a doctor. Some non-AI mobile apps, such as Mole Mapper, already allow people to track suspicious moles and record their changes over time.
But Mr. Halpern said he did not think consumers were ready for diagnostic systems that might tell them that a mole has a 5 percent chance or a 50 percent chance of being cancerous. “We’re not good at using these probabilities,” he says.