Teaching a Machine to Learn

Leveraging artificial intelligence requires humans to teach computers how to recognize patterns in data. Duke faculty and staff feed in data from ophthalmic images, and program and refine complex algorithms (series of mathematical operations) that can help identify a variety of eye conditions.The training process begins slowly. Input an image into the computer’s algorithm. Let it guess whether a disease is present in that image or not. Confirm or correct and reweight the algorithm’s parameters as needed. Then another image. Then another. As the images build, the computer begins to see commonalities and differences, makes its own adjustments, and gets better at detecting which patterns indicate a disease and which do not. Eventually, nourished by data from hundreds and thousands of images, the computer becomes even better at recognizing complex patterns of disease presence –and faster, and less expensive. Now it can be deployed to interpret new images it has never seen.

Felipe Medeiros, MD PhD used an AI algorithm called neutral transfer to transfer the painting style of famous artists DaVinci, Van Gogh, Picasso and Kandinsky to visualize how they would paint an optic nerve with glaucoma. 

Share