The algorithm was able to achieve 97% accuracy when training on one CT scan (with a visible stroke region) and testing on data from that same scan:
When tested on entire CT scans, with visible stroke regions, the algorithm performed quite well, and the probability image produced was very clear. However, the algorithm was not able to generate a usable probability image on CT scans with non-visible stroke.
At left is the probability image produced for a CT scan with a visible stroke region. The neural network output clearly identifies the stroke region. At right is the probability image produced for a CT scan with no visible stroke region (but does contain ischemic stroke – as shown by MRI). This image does not display any clearly identified stroke region.