Original title: BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net
Marven Sherif, Amgad Elmasry, Youssef Ghazal, Ayman Elghotni · September 2026
Spotting diseased tissue or lesions in medical body scans is difficult for computers. Different radioactive imaging chemicals, called tracers, make tissues look very different. To solve this, researchers built a tool that lets human doctors guide the computer by drawing rough marks, or scribbles, on the scans. These scribbles tell the system what is healthy tissue and what is diseased. The computer network takes in four data streams: two normal body scans plus separate maps for the healthy and diseased marks. Because it was already trained on regular scans, scientists carefully added these extra mark layers without wiping out its prior knowledge. Testing showed that without any human marks, the tool achieved an accuracy score of 0.554. Adding just five rounds of scribble corrections raised that score to 0.751. About 85% of that entire improvement happened right after the very first scribble.
Imagine sorting a mixed pile of laundry where some items are clean and some are dirty, but they look similar. It's hard to tell them apart just by looking, so you might make mistakes and mix clean and dirty clothes. In this paper, the system is like an assistant that helps sort laundry, but it needs hints from you—like pointing out one dirty sock—to figure out the rest more accurately.
Doctors use PET/CT scans to find and track cancer, but automated tools often miss lesions or make mistakes. Giving users a way to correct these errors with simple drawings makes the process faster and more accurate, directly improving patient care.
The abstract does not specify limitations, but the work is presented in the context of the autoPET/CT V challenge, so its performance may be specific to that dataset and not generalize to all PET/CT scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net — Marven Sherif, Amgad Elmasry, Youssef Ghazal et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.