DEVELOPMENT OF AI MODELS FOR ISCHEMIC STROKE DETECTION: A COMPARATIVE STUDY OF CNN AND TRANSFORMER ARCHITECTURES

Abstract

Ischemic stroke is the second leading cause of death worldwide, responsible for approximately 6 million disability-adjusted life years annually. Because each hour of untreated ischemia destroys roughly 1.9 billion neurons, fast and accurate lesion segmentation has direct clinical consequences. This paper compares four deep learning architectures for pixel-wise acute ischemic stroke (AIS) segmentation on the TEKNOFEST-2021 dataset (877 CT and 230 MRI studies from 819 patients): a CNN baseline (U-Net) and three Transformer-based hybrids (TransUNet, UNETR, Swin-UNet). All models were trained under identical conditions using the Adam optimizer, a composite Dice + cross-entropy loss, and a 60/20/20 patient-level split over 50 epochs. All four exceed the clinically accepted DSC threshold of 80%. Swin-UNet achieves the highest overall DSC (88.6%) and pixel accuracy (98.7%); UNETR achieves the best boundary precision (HD95 = 8.9 mm). The largest single gain is a 10.1-point DSC improvement on lacunar stroke detection. Ablation experiments show ImageNet pretraining contributes 5-6 DSC points and elastic deformation augmentation 3.9 points. All Transformer models outperform the CNN baseline with large effect sizes (d > 0.82, p < 0.001).