A research team has published a new AI model designed to find acute stroke damage in MRI scans more reliably than existing tools.
The model, called EPRA U-Net, combines several established deep learning components — an EfficientNet encoder, recurrent processing blocks, and a dual attention mechanism — into a single architecture tuned specifically for diffusion-weighted MRI, the scan type clinicians use to spot fresh stroke damage. Researchers trained and tested it on scans from 167 patients, totaling 4,895 image slices. Against three commonly benchmarked models — UNet++, DeepLabV3+, and TransUNet — EPRA U-Net posted a per-sample Dice score of 0.9469 and missed fewer lesions across the board: 16% fewer than UNet++, 25% fewer than DeepLabV3+, and 29% fewer than TransUNet.
Missed lesions in stroke imaging are not a minor accuracy footnote — they can directly affect treatment decisions, including whether a patient qualifies for clot-removal therapy within a tight time window. A model that is measurably more sensitive to small or subtle infarcts could, if validated in clinical settings, reduce the rate at which early-stage strokes go undetected on initial read. The team leaned into this by choosing a Tversky loss function that explicitly trades some specificity for higher recall.
The dataset of 167 patients is narrow, and the evaluation is in-house — independent, multi-site validation would need to follow before anyone should read these benchmark numbers as a clinical performance guarantee.