Attention-Gated 3D U-Net for Automated Ischemic Stroke Lesion Segmentation from Multi-Parametric Brain MRI

Abstract

Manual delineation of ischemic stroke lesions on brain MRI is time-consuming and subject to disagreement between readers, which motivates reproducible automated methods. This work develops an Attention-Gated 3D U-Net for lesion delineation from three MRI sequences, namely diffusion-weighted imaging (DWI), the apparent diffusion coefficient (ADC) map, and fluid-attenuated inversion recovery (FLAIR). The study uses 250 subjects, each accompanied by an expert-drawn binary lesion mask. A four-level encoder-decoder with attention gates on its skip connections is trained under a joint Dice and binary crossentropy objective. Prior to training, every volume is resampled to 1.0 mm isotropic spacing, z-score normalized over in-brain voxels, and cropped to the brain region. Augmentation is applied during training and combines spatial and intensity perturbations. Optimization proceeds for 116 epochs using AdamW with a cosineannealed learning rate, attaining a best validation Dice of 0.6821. Evaluation employs both five-fold cross-validation and a held-out partition, while mixed-precision training, gradient accumulation, and a channels-last memory layout keep memory consumption modest. The attention gates concentrate the network response on lesion tissue and reduce spurious activations elsewhere.