A research team has published RUFNet, a neural network framework designed to segment brain tumors in MRI scans while training on very few labeled images.
Most medical imaging models need large, carefully annotated datasets to perform well — a bottleneck in clinical settings where labeled scans are scarce. RUFNet takes a few-shot approach, learning from a small set of "support" images and then generalizing to new "query" scans. The framework combines three components: a Hybrid Mamba backbone that tracks relationships between support and query images at manageable computational cost; a mask refinement module that uses query features to clean up noisy annotations in the support set; and an uncertainty module that flags low-confidence pixel predictions and blends them with a more conservative prior. On the BraTS 2020 benchmark — the standard Brain Tumor Segmentation Challenge dataset — RUFNet scored Dice coefficients of 84.3% in the one-shot setting and 86.1% when given five support examples.
Those numbers matter because few-shot segmentation models typically trail fully supervised ones by a wide margin; closing that gap even partially could reduce how much expert annotation time a clinical deployment requires. The uncertainty module is the less-obvious contribution: rather than producing a single confident-sounding prediction on ambiguous pixels, it quantifies doubt and defers — a more honest output for a high-stakes domain.
Mamba-based architectures have been gaining ground in medical imaging as a lighter alternative to full attention transformers, and RUFNet fits that pattern. Whether benchmark performance on BraTS 2020 — a relatively well-studied dataset — translates to messier real-world scans remains the open question, as it always does.