Automatic and Rapid Identification of Infarct Slices and Hemisphere in DWI Scans

Abstract

Accurate, free of observer's bias, and fast identification of acute infarct is critical in visual and automatic processing of stroke images. An automatic and rapid algorithm has been developed to identify the infarct slices and the hemisphere in diffusion-weighted imaging (DWI) scans. Thirty-six DWI scans were acquired from five centers with the slice thickness of 4-14 mm. We also derive images from the original scans to assess the accuracy of the algorithm by using a wide range of infarct size and number of artifacts per unit area. Based on the difference in percentile characteristics of intensity normalized (infarct/noninfarct) images, two parameters are defined: R(s) for infarct slice identification and R(h) for infarct hemisphere identification. Using the identified infarct slices the infarct hemisphere is subsequently determined. The average sensitivity and specificity for slice and hemisphere identification were 98.1%, 51.4% and 91.7%, 91.7%, respectively. The processing time is approximately 3-5 seconds on Matlab platform and on VC++ it is predicted approximately 10 milliseconds. Based on simulation study, we can infer that the algorithm produces accurate results in most of the situations although the sensitivity goes down by approximately 15% when the infarct size is small (<2-3% of image area) and the artifacts per unit area are large. The proposed algorithm applied as a preprocessor can be useful to: 1) estimate location (hemisphere) and extent of infarct (number and location of slices), 2) reduce time and labor of infarct volume study, 3) cross-check visual interpretation, 4) form a part of an infarct segmentation module, and 5) improve localization of the midsagittal plane.