Study scans

The NCCT scans in this study were obtained from the Toshiba Aquilion One from two INternational Stroke Perfusion Imaging REgistry (INSPIRE) sites (Canon, Tokyo, Japan). The NCCT lesion was segmented automatically with the AI-based HDT algorithm on MIStar software as illustrated in Figure 1 (Apollo Medical Imaging, Melbourne, Australia). It consisted of the following steps: (1) assessing the symmetry of the density histograms of the left and right hemisphere (after registration to a template), (2) definition of potential seeds, and (3) iterative optimization of level-set thresholds.

The clinicians were instructed to manually segment on axial views, a slice of each of the five NCCT scans. Clinicians were provided the whole brain NCCT for review only. The segmentation was completed using the paintbrush mode in the ITK-SNAP software application.¹

MRI-DWI images were manually segmented by trained personnel using ITK-SNAP to extract infarct lesions as reference. MRI-DWI lesions were registered to the NCCT images using Advanced Normalization Tools (ANTS) (16).