Scientists describe a potential screening method for COVID-19 based on eye images analyzed by artificial intelligence. Scanning a set of images from several hundred individuals with and without COVID-19, the tool accurately diagnosed coronavirus infections more than 90 percent of the time, the developers reported in a preprint posted to medRxiv September 10.
“Our model is quite fast,” Yanwei Fu, a computer scientist at Fudan University in Shanghai, China, who led the study, tells The Scientist. “In less than a second it can check results.”
Currently, screening for coronavirus infection involves CT imaging of the lungs or analyzing samples from the nose or throat, both of which take time and require professional effort. A system based on a few images of the eyes that could triage or even diagnose people would save...
Volunteers at Shanghai Public Health Clinical Centre in Fudan each had five photos of their eyes taken using common CCD or CMOS cameras. Of 303 patients, 104 had COVID-19, 131 had other pulmonary conditions, and 68 had eye diseases. A neural network tool extracted and quantified the features from different regions of the eye and an algorithm recognized the ocular characteristics of each disease. A neural network is a series of algorithms for solving AI problems, learning as it goes along in a way that mimics the human brain. The researchers then carried out a validation experiment on a small dataset from healthy people, COVID-19 patients, pulmonary patients, and ocular patients.
Of 24 people with confirmed coronavirus infections, the tool correctly diagnosed 23, Fu tells The Scientist. And the algorithm accurately identified 30 out of 30 uninfected individuals.
Coronavirus infections, not just those caused by SARS-CoV-2, have long had associations with the eye, causing inflammation of the transparent membrane that covers the inside of the eyelid and whites of the eyeball, a condition called conjunctivitis, or pink eye. The eyes also offer a route to infection for respiratory viruses, including coronaviruses.
Human coronavirus NL63, which causes common cold symptoms, was first identified in 2004 in a baby with bronchiolitis and conjunctivitis. Subsequent studies showed that a minority of children infected with this coronavirus suffer from this eye condition.
Although conjunctivitis remains a potential symptom of coronavirus infections, less than 5 percent of COVID-19 patients actually present with eye symptoms, notes Daniel Ting, ophthalmologist at the Singapore National Eye Centre, who has published on this topic and deep learning in ophthalmology. “If you look to develop an AI system to detect COVID-19 based on [limited numbers of] eye images, I think the performance is not going to be great,” especially given the low prevalence of eye symptoms. He doubts the performance of the algorithm also because “a lot of eye manifestations could be due to reasons other than COVID-19.”
Ting cautions that the sample size of 303 patients and 136 healthy individuals in the Shanghai study is too small to draw strong conclusions. “To develop a good deep learning system to automatically detect some unique features from any medical imaging requires more patients,” he says. “In order to increase the reliability of this study, the same size would need to be multiplied by at least ten times, so, thousands of patients.”
Fu has started down this road, increasing the number of participants and broadening the types of subjects. “We are now doing more double-blind tests in the hospitals, with patients, some with eye diseases,” he says. The group also plans to introduce an online screening platform that uses the algorithm to screen for COVID-19.
“As an ophthalmologist it would be very surprising if there is a distinct COVID viral conjunctivitis pattern as opposed to other similar forms of viral conjunctivitis,” ophthalmologist Alastair Denniston, the director of the Health Data Research Hub for Eye Health in Birmingham, UK, writes in an email to The Scientist. “This is unlike building an algorithm for conditions which are biologically more distinct like macular degeneration,” he writes.
He notes that if there were a unique pattern evident in COVID-19 cases, “then the comparison for training and testing should be against cases that look similar,” such as non–COVID-19 viral conjunctivitis or other causes of a red eye associated with colds caused by adenovirus or rhinovirus. He also faults the paper in not providing “the necessary description to really critique the science in terms of how they built and (tried to) validate the model.”
Denniston recently reviewed more than 20,000 AI studies on detecting disease from medical imaging, but found that less than 1 percent were sufficiently robust in their design and reporting that independent reviewers had high confidence in their claims. This led him to convene a group of experts to define the international standards for the design and reporting of clinical trials of AI systems. These standards were published this month in Nature Medicine, The BMJ, and Lancet Digital Health and are supported by leading medical journals.
The Shanghai study has some potentially controversial applications, even if the AI works. Their algorithm could be used in public places, Fu says, though this would raise data privacy concerns in many countries. “In China, for example, we have a lot of high-resolution cameras everywhere,” he notes. “In airports or at train stations, we could use these surveillance cameras to check people’s eyes.” The program would be most accurate if people looked directly at the camera, but Fu says “as long as our camera can clearly watch the eye region it would be good enough.”
Screening the public without expressed consent using this algorithm would be ruled out of bounds in some parts of the world. “In Europe, this would be highly problematic and most likely illegal, in violation of the EU Charter of Fundamental Rights and general data protection legislation,” says computer scientist Barry O’Sullivan of University College Cork in Ireland who is an expert in AI. The gathering of health data and biometric data in Europe requires consent.
O’Sullivan echoes the concern that the paper falls short on detail regarding its methodology. “It is an interesting hypothesis,” he says. But, as currently written, it isn’t ready for publication in a machine learning journal, he concludes.