Agreement of Cercare artificial intelligence with neuroradiologist interpretation and Vitrea CT perfusion outputs in acute ischaemic stroke: A dual-centre retrospective study.
Authors
Affiliations (4)
Affiliations (4)
- School of Medicine & Dentistry, University of Lancashire, Preston, United Kingdom; The University of Chester, Chester, United Kingdom; Countess of Chester Hospital, Chester, United Kingdom; Aintree Hospital, Liverpool, United Kingdom.
- The University of Chester, Chester, United Kingdom; Countess of Chester Hospital, Chester, United Kingdom; North Bristol Hospital, Bristol, United Kingdom. Electronic address: [email protected].
- Countess of Chester Hospital, Chester, United Kingdom.
- St Helens and Knowsley Teaching Hospitals NHS Trust, St Helens and Knowsley, United Kingdom.
Abstract
To assess agreement between Cercare artificial intelligence (AI) and expert neuroradiologist interpretation for CT angiography (CTA) large vessel occlusion (LVO) detection, and to compare Cercare and Vitrea CT perfusion (CTP) outputs in acute ischaemic stroke. This dual-centre retrospective study included 214 adults undergoing multimodal CT for suspected acute ischaemic stroke. LVO agreement was assessed in 162 CTA datasets using consultant neuroradiologist interpretation as the comparator; no angiographic or consensus reference standard was used. The quality-controlled primary CTP analysis included 196 paired datasets after exclusion of five studies with documented technical compromise; an all-output sensitivity analysis included 201 datasets. Treatment classification was assessed in 65 quality-controlled neuroradiologist-confirmed LVO cases with paired perfusion outputs, with an all-output sensitivity analysis in 67 cases. Cercare-derived cerebral metabolic rate of oxygen (CMRO<sub>2</sub>) volumes were examined in exploratory, hypothesis-generating analyses against perfusion core volumes and discharge modified Rankin Scale (mRS). Cercare showed positive percent agreement of 82.3% (95% CI 71.0-89.8) and negative percent agreement of 83.0% (74.5-89.1) with neuroradiologist interpretation, with substantial agreement (kappa = 0.641). In the quality-controlled CTP cohort, core limits of agreement were - 34.7 to 30.1 mL (ICC 0.930) and penumbra limits were - 94.6 to 104.7 mL (ICC 0.812). Inclusion of five technically compromised studies widened the corresponding all-output limits to - 90.5 to 86.4 mL and - 191.4 to 219.5 mL. In 65 quality-controlled LVO patients, agreement was 93.8% for EXTEND-IA-style imaging mismatch and 87.7% for DAWN-style clinical-core mismatch classification. In exploratory, hypothesis-generating analyses, CMRO<sub>2</sub> correlated with Vitrea core (ρ = 0.758-0.787) and discharge mRS (ρ = 0.390-0.401). Cercare showed substantial agreement with expert CTA interpretation. Quality control materially influenced CTP agreement estimates, and residual individual-patient variability, particularly for penumbra and treatment-threshold classification, indicates that outputs should not be regarded as interchangeable. The DAWN-derived analysis represents a threshold classification rather than full trial eligibility. CMRO<sub>2</sub> findings are exploratory and hypothesis-generating and require prospective validation.