Automated Segmentation of Infarct Core in Non-Contrast CT Scans of Ischemic Stroke Patients

Ischemic stroke is a crippling disorder that affects many people globally, and timely detection and prognosis are vital for efficient treatment. Despite the increasing use of Computed Tomography (CT) examination in stroke evaluation, automated assessment of CT images for detecting early symptoms of ischemia in stroke patients has been challenging. This research aims to automatically detect and segment the infarct core areas in Non-Contrast CT (NCCT) scans of ischemic stroke patients using deep transfer learning. Leveraging the advantages of transfer learning, we evaluate multiple pre-trained ImageNet models of the U-Net architecture for segmentation of the infarct regions in NCCT scans. We observe that the U-Net model based on the VGG-16 backbone successfully identifies and outlines infarct regions in NCCT scans, achieving an impressive Intersection over Union (IoU) score of 0.76 and a Dice coefficient of 0.79.