Lesion Stroke Segmentation: A Deep Learning Approach

Stroke remains one of the leading causes of death and long-term disability worldwide. Accurate and rapid detection of stroke lesions in brain MRI scans is essential for timely intervention and effective treatment planning. Manual segmentation is tedious and prone to variability. This research proposes an automated method for lesion stroke segmentation using a 2D U-Net deep learning architecture enhanced by the Focal Tversky loss function to address class imbalance. Training on the ATLAS v2.0 dataset ensures robustness, while rigorous evaluation using Dice Similarity Coefficient, Intersection over Union (IoU), and Sensitivity demonstrate the model’s ability to segment small, irregular lesions. This automated approach aims to support radiologists, minimize manual workload, and contribute to improved patient outcomes.