Abstract
Objectives Posterior circulation infarction (POCI) is a common but frequently missed diagnosis. Hyper acute POCI is poorly visualised using traditional CT and MRI sequences. Deep learning is a branch of artificial intelligence, which facilitates automated detection of imaging features not readily identified by clinicians. We aimed to develop a novel convolutional neural network (CNN) to classify POCI using CT perfusion (CTP).
Methods Data were analysed from the International-stroke-perfusion-registry (INSPIRE). Patients with baseline multimodal-CT and follow up diffusion-weighted MRI at 24–48 hours were included. Patients with POCI on follow up MRI were identified. A reference group of randomly selected patients with non-POCI diagnosis were collated to form a dataset in a 1:4 POCI to reference-ratio. A 3D-DenseNet was trained to classify participants into POCI or non-POCI using CTP deconvolved maps.
Results Eighty-eight patients with POCI were included (median age: 69 with interquartile range [60- 78]; NIHSS at baseline: 8 [5–14]; DWI lesion volume: 3 [0.6–16] ml). Three-hundred-two patients were included in the reference group (median age: 72.5 [61–80.8]; NIHSS at baseline: 12 [6–17]; DWI lesion volume: 15.1 [3–50] ml). Optimal model performed was achieved using Delay Time (DT) maps with an accuracy of 0.89 (sensitivity: 0.77; specificity: 1). Mean Transit Time and Cerebral Blood Flow yielded lower but acceptable accuracies of 0.83 (sensitivity/specificity: 0.61/0.97) and 0.80 (sensitivity/specificity: 0.51/0.97), respectively.
Conclusions Classification of POCI using a CNN is highly accurate. Optimal model performance was achieved using DT maps. A CNN classification model may aid in the rapid and accurate diagnosis of POCI.
