Computer Aided Detection of Early Ischemic Changes in Non-Contrast CT Images of Acute Ischemic Stroke
Abstract
An ischemic stroke occurs from a brain artery blockage. Rapid treatment improves outcomes. However, there are two treatment prerequisites. First, early ischemic changes (EIC) in brain tissue must be established, often by interpreting a computed tomography (CT) scan. However, EIC identification from CT is difficult. Second, treatments are only effective within the first few hours. Consequently, the maximum delay between stroke onset and brain imaging must be estimated. This is difficult if the stroke onset was not observed. The goal of my PhD research was to create computer-aided detection (CADe) tools to: a) improve assessment of EIC in CT images; and, b) estimate the delay between stroke onset and brain imaging using CT scan features.
My algorithm evaluated CT image features associated with EIC. Three stages of assessment determined software performance. First, I determined physician reliability when outlining EIC on CT images. Next, I determined how closely my algorithm matched physician driven outlines. Finally, I determined physician reliability when assisted by my algorithm. My CADe methods were used to calculate a new stroke severity index (SSI). The relationship between SSI and the delay between stroke onset and imaging time was then examined.
Results showed that my CADe method identified EIC and helped physicians. Unassisted, the inter-rater reliability of identifying EIC was moderate. My software identified areas of EIC, but not as well as the unassisted physicians. With my software, physicians achieved a higher measure of agreement (Dice coefficient) than stroke fellows. However, fellows still achieved a Dice coefficient close to that of the physicians without software assistance. Inter- and intra-rater reliabilities were similar between the physicians and the fellows. The SSI did not correlate time from stroke onset, but did fit with known understanding of ischemic stroke pathology.
This thesis showed that stroke CADe may have clinical utility. It quickly identified EIC in early CT images and increased the reliability of EIC identification by physicians. My ischemic stroke CADe tool brought quantitative methods to a subtle, subjective field.
Description
Keywords
Neuroscience, Computer Science, Engineering--Biomedical
Citation
Thomson, Q. P. (2014). Computer Aided Detection of Early Ischemic Changes in Non-Contrast CT Images of Acute Ischemic Stroke (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/25399