In brachytherapy, thin catheters are inserted into tumors. On CT scans, they look like tangled glowing threads. Previously, doctors would manually outline each one, spending up to an hour.
The new algorithm learns from synthetic scans — drawing metal strips in an empty volume, like threads without a body. The neural network masters photometry (measuring brightness) and fine-tunes on just ten real catheters. Then a physical tracker steps in: it untangles the threads, tracking their direction and preventing them from sticking together. This principle of separating overlapping signals is familiar from spectroscopy, where a prism splits a mixture of colors into individual beams. The tracker reduces entropy — the disorder in data.
The entire analysis takes 12 seconds — faster than manually opening the software. This allows doctors to adjust the radiation plan right during the procedure. An unexpected fact: the tracker relies on the Hough transform, a mathematical trick from 1962 devised to analyze bubble chamber images. Today, it saves hours in oncology.
Brachytherapy catheters are made of flexible metal with gold markers — on X-rays they sparkle like reflectors on a jacket.