FastMRI-Guided Mamba UNet for Real-Time Emergency Brain Lesion Detection

Abstract

 Emergency imaging often relies on fast MRI protocols, which introduce motion artifacts, low resolution, and reduced contrast, posing substantial challenges for lesion detection. TGMamba-UNet enhances robustness under these conditions by integrating temporal guidance vectors learned from high-quality MRI with a Mamba-UNet backbone capable of modeling longrange structural dependencies. The model effectively suppresses artifact-induced distortion and stabilizes predictions in noisy environments. Experiments conducted on 2,280 MRI scans (1,460 with motion artifacts and 820 high-quality references) show that TG-Mamba-UNet improves Dice from 0.812 to 0.894 (+10.1%) and reduces artifact-induced segmentation error by 17.8%.