Deep Learning Enabled Acute Ischemic Stroke Lesion Segmentation for Smart Healthcare Systems

Intelligent Decentralized Edge Computing for Assisted Smart Healthcare Systems is developing rapidly. The diagnosis and prediction of global diseases are key challenges nowadays. The primary factor and a common challenge in efficient stroke management is in early diagnostics. The decisive method and the first task is neuroimaging, based mainly on CT examination. Thus, immediate detection of any abnormal hypodensities in the brain tissue is a key factor in efficiently treating the patient. This study uses an image to mask translation approach with Conditional Generative Adversarial Networks (CGANs) to detect abnormal hypodensity in the brain tissue. The translation of GANs conditioned on the symmetrical nature of the brain is applied to compare healthy brain tissue with incoming CT scan, thus identifying the onset of pathology. A dataset of CT scans used in this study was collected from 4 different hospitals in Kazakhstan, comprising 920 real cases, with 21068 images. Experimental results show high efficiency in detecting locations containing abnormal hypodensities associated with a notably higher completion error than in zones without abnormalities. The proposed solution will allow early detection of the brain infarction zone and determine the extent of the lesion.