3D Ischemic Stroke Lesion Segmentation in MRI with Attention U-Net and Voxel-Level Uncertainty Quantification

Abstract Ischemic stroke is a major cause of death and disability around the world, and accurate segmentation of stroke lesions from MRI diagnosis are critical for treatment planning and clinical decision making. Compared to traditional methods, while deep learning has advanced automated stroke segmentation, but existing approaches predict single binary masks with no estimate of prediction confidence, a crucial limitation since clinical usage would require knowledge on where the model can be trusted just as much as the prediction itself. We propose a reliability-aware 3D framework combining an Attention U-Net with Monte Carlo Dropout inference and an optimized decision threshold, tested on the ISLES 2022 benchmark (240 multi-center DWI + ADC cases; 168/36/36 split). At inference, T = 20 stochastic forward passes generate both a segmentation mask and a voxel-level uncertainty map. This work achieved a Dice of 0.6816 (IoU = 0.5526, Precision = 0.8764 and Recall = 0.6595) for single-model volumetric methods on this benchmark dataset, meanwhile boiling Monte Carlo Dropout increased Recall by + 2.65% at the zero-training cost. Uncertainty maps were semantically aligned but spatially localized to the edges of lesions, highly correlated with segmentation errors, and inversely associated with lesion size automatically marking small and difficult lesions as cases of high uncertainty.