In a dark room, on a table, sit a laser pointer, a piece of glass with tissue, and an ordinary phone camera. Light passes through the sample, and the image shows not cells but a pattern of bright and dark stripes, like the shadow of leaves on a wall. From this pattern, the computer learns to guess how much HER2 protein is on the surface of breast cancer cells.
The HER2 protein is like a door handle that a drug can grab onto. If there are few handles, the drug won't work; if there are many, doctors prescribe special treatments. Usually, handles are counted under a heavy microscope with expensive lenses. The new system does away with lenses: a laser with three colors illuminates the tissue, and the camera captures how the light scatters. A neural network analyzes these patterns and assigns a score from 0 to 3+.
To avoid mistakes, the program itself assesses its confidence. If it's uncertain, it flags the answer as unreliable, and then a human checks it. On 412 samples, this approach yielded 84.9% accurate answers for the four stages and 94.8% for the key question: is there a drug target. Every third error was caught and corrected.
Such a device weighs less than a kilogram and costs tens of times less than a typical scanner. In a small hospital without a pathologist, it could review dozens of samples in an hour and suggest who needs expensive drugs and who doesn't.
The laser pointer in the experiment shines with three colors at once (red, green, and blue, like in old TVs) and separately illuminates different layers of tissue.