Initial Experience With a Deep Learning Algorithm for Detecting Posterior Circulation Large Vessel Occlusion on Noncontrast Computed Tomography.
Authors
Affiliations (10)
Affiliations (10)
- Department of Neurology (A.B., L.C.-C., N.A., E.K., W.D.H., J. Cespedes, S.O.-G.), University of Iowa Health Care, Iowa City.
- Methinks Software SL, Barcelona, Spain (J. Cendrero, L.T., V.S., A.U., C.M.).
- Department of Radiology, Stanford University, CA (P.C.).
- Stroke Unit, Neurology Department, Hospital Universitario Vall d'Hebron, Barcelona, Spain (P.C., M.R.).
- Cooper Neurological Institute, Cooper University Health Care, Camden, NJ (T.G.J.).
- Department of Neurology, Cooper Medical School of Rowan University, Camden, NJ (T.G.J.).
- Department of Neurosurgery, Advocate Lutheran General Hospital, Park Ridge, IL (D.L.).
- Department of Neuroradiology, Rockefeller Neuroscience Institute, West Virginia University, Morgantown (A.R.).
- Department of Neurosurgery (S.O.-G.), University of Iowa Health Care, Iowa City.
- Department of Radiology (S.O.-G.), University of Iowa Health Care, Iowa City.
Abstract
Detection of large vessel occlusions using a deep learning (DL) algorithm for the anterior circulation has shown promising results. However, the role of DL algorithms in detecting posterior circulation large vessel occlusion (PC-LVO) remains uncertain. We aimed to evaluate the diagnostic performance of a DL algorithm (Methinks PC-LVO) for detecting PC-LVO using noncontrast computed tomography. This is a retrospective, multicenter, observational cohort study that included patients with PC-LVO who underwent both noncontrast computed tomography and computed tomography angiography. The diagnostic performance of the DL algorithm was assessed by analyzing sensitivity, specificity, and area under the curve for PC-LVO detection of consecutive PC-LVO strokes. For comparative analysis, the area under the curve of the DL algorithm was also evaluated against the performance of a neuroradiologist interpreting noncontrast computed tomography. Ground truth labels were established through consensus readings by expert neuroradiologists. Subgroup analyses were performed according to clot location (proximal posterior cerebral artery and basilar artery) and National Institutes of Health Stroke Scale (NIHSS) score (NIHSS score ≥6, NIHSS score ≥8, and NIHSS score ≥10). A total of 196 patients were included, of whom 74 patients had PC-LVO. Among these, 43 had basilar artery occlusions, and 34 had proximal posterior cerebral artery occlusions. The overall sensitivity and specificity of the software were 55.4% and 80.9%, respectively, with an area under the curve of 0.72. The neuroradiologist achieved a sensitivity of 27.4% and specificity of 91.8%. Among patients with proximal posterior cerebral artery occlusion, sensitivity was 55.9%, whereas for basilar artery occlusion, it was 53.5%. When stratified by NIHSS score ≥10, sensitivity in the proximal posterior cerebral artery was 56.2%, and for the basilar artery with NIHSS score ≥10, it increased to 61.5%. Our initial experience with a DL algorithm for the detection of PC-LVO showed promising results. However, further improvements are required before the algorithm can be implemented in clinical practice.