Posterior circulation stroke: machine learning-based detection of early ischemic changes in acute non-contrast CT scans.

Objectives: Triage of patients with basilar artery occlusion for additional imaging diagnostics, therapy planning, and initial outcome prediction requires assessment of early ischemic changes in early hyperacute non-contrast computed tomography (NCCT) scans. However, accuracy of visual evaluation is impaired by inter- and intra-reader variability, artifacts in the posterior fossa and limited sensitivity for subtle density shifts. We propose a machine learning approach for detecting early ischemic changes in pc-ASPECTS regions (Posterior circulation Alberta Stroke Program Early CT Score) based on admission NCCTs.

Methods: The retrospective study includes 552 pc-ASPECTS regions (144 with infarctions in follow-up NCCTs) extracted from pre-therapeutic early hyperacute scans of 69 patients with basilar artery occlusion that later underwent successful recanalization. We evaluated 1218 quantitative image features utilizing random forest algorithms with fivefold cross-validation for the ability to detect early ischemic changes in hyperacute images that lead to definitive infarctions in follow-up imaging. Classifier performance was compared to conventional readings of two neuroradiologists.

Results: Receiver operating characteristic area under the curves for detection of early ischemic changes were 0.70 (95% CI [0.64; 0.75]) for cerebellum to 0.82 (95% CI [0.77; 0.86]) for thalamus. Predictive performance of the classifier was significantly higher compared to visual reading for thalamus, midbrain, and pons (P value < 0.05).

Conclusions: Quantitative features of early hyperacute NCCTs can be used to detect early ischemic changes in pc-ASPECTS regions. The classifier performance was higher or equal to results of human raters. The proposed approach could facilitate reproducible analysis in research and may allow standardized assessments for outcome prediction and therapy planning in clinical routine.