Ischemic Stroke Lesion Segmentation

Overview

An academic Taif University project exploring deep learning for medical image segmentation using the ISLES 2022 dataset.

The work trained a U-Net model to identify stroke lesions in brain MRI scans and compared predictions against the provided ground-truth masks.

Problem

Stroke lesion segmentation is challenging because lesions vary substantially in size, location, and appearance.

Model performance also depends heavily on the quality of dataset annotations used during training and evaluation.

Solution

I implemented and trained a U-Net model using preprocessed and normalized MRI images resized to 224 × 224, with Dice Loss optimization and segmentation visualization.

The model initially reached an approximately 0.83 Dice score on the provided dataset.

Post-training error analysis identified inconsistent and incorrect ground-truth labels in some samples, showing that part of the measured error came from annotation quality rather than the model alone.