Abstract
INTRODUCTION/PURPOSE: Deep learning (DL) algorithms have shown promising results in detecting anterior circulation large vessel occlusions (LVO) in non‐contrast (NCCT), but its utility in posterior circulation LVO detection remains uncertain. We aimed to evaluate the diagnostic performance of the Methinks POST‐LVO DL algorithm for identifying posterior circulation LVO using NCCT. MATERIALS/METHODS: This is a retrospective, multicenter, observational cohort study that included patients with posterior circulation LVO, specifically those with basilar artery (BA) or proximal posterior cerebral artery (PCA) occlusions who underwent both NCCT, and computed tomography angiography (CTA). Ground truth labels were established through consensus readings by experts neuroradiologists on the CTA, and the performance of the DL algorithm in posterior circulation LVO detection was assessed on this dataset, analyzing sensitivity, specificity, and area under the curve (AUC). For comparative analysis, the AUC of the DL algorithm was also evaluated against the performance of neuroradiologists interpreting NCCT blinded to CTA using a Delong test. Subgroup analyses were performed according to clot location and National Institutes of Health Stroke Scale (NIHSS) score. RESULTS: A total of 77 patients with posterior circulation LVO were included, of which 43 had BA occlusions and 34 had proximal PCA occlusions. The overall sensitivity and specificity of Methinks algorithm were 55.4% and 80.9%, respectively, with an AUC of 0.723 whereas the neuroradiologist achieved a sensitivity of 27.4% and specificity of 91.8% with an AUC of 0.59. In Figure 1, the comparison between the DL algorithm and neuroradiologists on NCCT showed a significantly higher AUC for the algorithm, with a difference of 13.3% (95% CI: 5.0%‐21.6%, p = 0.0017). Among patients with proximal PCA occlusion, sensitivity was 55.9%, while for BA occlusion, it was 53.5%. When stratified by NIHSS > 10, sensitivity in proximal PCA was 56.2% and for BA with NIHSS > 10, it increased to 61.5%. CONCLUSION: Our initial experience with a DL algorithm for the detection of posterior LVO showed positive and promising results. However, further improvements are required to overcome the current limitations and before the algorithm can be implemented in clinical practice. [Image: see text]