To develop and assess a deep learning algorithm (“CT Infarct”) for detection and volume quantification of the ischemic core on non-contrast CT head imaging.

Stroke care has been revolutionized since the demonstration of benefit from endovascular thrombectomy. An important selection criterion involves estimating the region of permanently damaged ischemic core, so that the possible benefit through restoring vascular supply to the ischemic penumbra can be determined. Clinicians often rely on advanced imaging techniques such as CT perfusion and MRI that are not as widely accessible as non-contrast CT.

We used a training dataset of paired CT head and MRI brain studies (within 3 hours for 1,896 positive pairs; 5 days for 1,670 negative pairs). We segmented the regions of reduced diffusivity on the MRI brain and registered them to the CT head. We then trained the algorithm using only the CT head images and the segmentations. A set of 150 pairs (90 unilateral MCA infarcts; 60 negative) was used to compare the algorithm with three experienced neuroradiologists in interpreting the CT head.

The CT Infarct algorithm performed with sensitivity 96% and specificity 72% for ischemic core >0mL detection, and sensitivity 78% and specificity 98% for ischemic core ≥5mL detection (the latter is the planned volume threshold for future clinical implementation to reduce false positives). The 3 neuroradiologists had mean sensitivity 64% and specificity 91%. The algorithm predicted an ischemic core volume in the correct range for 83% of cases in the range 20-50mL and 97% in the range >50mL. The neuroradiologists predicted the correct volume for 23% and 38% in each range respectively.

The CT Infarct algorithm can detect and quantify the volume of ischemic core better than three experienced neuroradiologists. It has the potential to help with stroke care, especially in hospitals with access to only routine CT imaging.