Abstract TP141: Can the Human Eye Match a Computer Algorithm in Identifying Hypoperfusion in Asymptomatic Carotid Artery Stenosis?

Abstract

Background: Perfusion weighted imaging on MRI (MRP) and computerized tomography perfusion (CTP) are increasingly required to manage large vessel disease. Computerized algorithms can quantify perfusion data, but the programs are expensive and not widely used outside acute stroke evaluation. We aimed to determine how well human observers can identify asymmetries in cerebral perfusion images compared with an automated computer algorithm. Methods: Ten clinicians experienced in treating carotid artery disease (4 vascular neurologists, 3 neuroradiologists, 1 vascular surgeon, 1 neurosurgeon, 1 interventional radiologist) were given 28 post-processed, color-coded, axial-slice MRP scans from patients in the Carotid Revascularization Endovascular versus Stenting Trial - Hemodynamics (CREST-H) study. All patients had >70%, unilateral, asymptomatic carotid artery stenosis and had varying degrees of time-to-peak (TTP) delay on the side of stenosis, ranging from 0 to 2 secs, quantified by a semi-automated system that computes quantitative perfusion maps, using deconvolution of tissue and arterial signals (Olea, Cambridge, MA). A minimum volume of 10cc was required for a given TTP delay. Clinicians were asked to determine asymmetry (y/n) and side of occlusion for each case. Number of correct responses that matched the computer output were tallied. Results: We averaged correct responses by the 10 clinicians for cases at each increment of TTP delay; (Figure). At TTP delays ≥1.5 seconds, accuracy was ≥80%. At 1.25 sec accuracy fell to 60%, and at ≤ 1 sec, accuracy was ≤50%. For TTP=0 (no asymmetry), accuracy was 71%. Conclusions: Visual impression of hemodynamic asymmetry among experienced clinicians was reasonably accurate for TTP delays ≥1.5 seconds, but declined with more subtle asymmetries. Depending on the clinical impact of TTP delays (for CREST-H: correlation with cognitive decline), experienced clinicians may perform as well as an automated algorithm.