Original title: GazeRefine: Expert Gaze as a Test-Time Prompt for Training-Free Medical Image Segmentation
Mohammed Oussama Benyahia, Marouane Tliba, Mohamed Amine Kerkouri, Taifour Yousra +9 · September 2026
Teaching computers to outline diseases in medical scans is hard because doctors must draw tedious, detailed masks by hand. Researchers built GazeRefine to skip this step entirely. Instead of hand-drawn outlines, it tracks where a doctor looks. The tool takes these brief, weighted eye fixations and marks them as targets or background. Next, it maps these points into a frozen visual model. It then grows these points into full outlines by comparing nearby visual patterns and sticking close to the doctor's original gaze. This allows it to trace whole areas the doctor only glanced at, without drifting off target. The tool needs zero extra training or tuning. It showed strong results on colonoscopy images and competitive results on prostate scans.
Imagine you're sorting laundry by color without any labels on the clothes. Normally, you'd need to study each item or have a guide point out what counts as 'red' or 'blue'. Here, a system uses where your eyes fixate while glancing at the pile to instantly group all similar-colored items together, even ones you didn't directly look at. In this paper, the expert's gaze fixations act as the initial focus, and the algorithm extends that to segment the entire image by spreading information based on visual similarity.
Doctors spend hours manually annotating images for diagnosis and research. This approach lets them use where they look to automatically mark areas, cutting down the need for pre-labeled data and making analysis faster in clinics.
The abstract does not specify limitations, but the method is evaluated only on polyp and prostate MRI segmentation. It may not generalize to other medical imaging tasks without further testing.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
GazeRefine: Expert Gaze as a Test-Time Prompt for Training-Free Medical Image Segmentation — Mohammed Oussama Benyahia, Marouane Tliba, Mohamed Amine Kerkouri 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.