An artificial-intelligence (AI) algorithm performs just as well as individual ophthalmologists in diagnosing congenital cataracts, reports a paper published online this week in Nature Biomedical Engineering. The algorithm, which has been implemented as a cloud-based platform for multihospital collaboration, may also help to improve the management of other rare diseases.
Convolutional neural networks - artificial networks with connectivity patterns inspired by the organization of the visual cortex - can accurately identify specific motifs in images when trained with massive amounts of curated data. However, such AI schemes have not been tested in the clinic when data is scarce, as is the case for rare diseases.
Yizhi Liu and colleagues report the implementation of convolutional neural networks for the diagnosis of congenital cataracts - a rare disease that causes clouding of the eye lens and is responsible for about 10% of all vision loss in children worldwide. They used a set of 50 cases involving various challenging clinical situations, a needle-in-a-haystack test with a realistic ratio of disease-to-healthy cases, and 57 cases from a phase-I multihospital clinical trial involving infants. The authors show that the AI algorithm diagnosed the disease, identified its severity, and suggested treatment with overall accuracies exceeding 90%.
Zoology: Mineral armour discovered in insectsNature Communications
Neuroscience: Social isolation evokes craving responses in the human brainNature Neuroscience
Ecology: Migration associated with faster pace of lifeNature Communications
Gene therapy: Concerns for the long-term safety of AAV gene therapyNature Biotechnology