EPRA U-Net: An Efficient Pyramid Residual Attention Framework for Infarct Segmentation in Diffusion-Weighted MRI

This paper introduces EPRA U-Net, a task-specific architecture for acute ischemic infarct segmentation in diffusion-weighted MRI. It combines an EfficientNet encoder, an R2 recurrent residual block, atrous spatial pyramid pooling, dual attention, and Tversky loss designed to emphasize lesion sensitivity. On an in-house dataset of 167 patients and 4,895 DWI slices, the model reported pixel-aggregated Dice of 0.8984, per-sample Dice of 0.9469, IoU of 0.8155, recall of 0.8887, lesion F1 of 0.9378, and HD95 of 11.62 pixels. The abstract reports 16% to 29% fewer missed lesions than UNet++, DeepLabV3+, and TransUNet.