Background Accurate carotid imaging is important for effective secondary stroke prevention. Non-invasive imaging, now widely available, is replacing intra-arterial angiography for carotid stenosis, but the accuracy remains uncertain despite an extensive literature. We systematically reviewed the accuracy of non-invasive imaging compared with intra-arterial angiography for diagnosing carotid stenosis in patients with carotid territory ischaemic symptoms. Methods We searched for articles published between 1980 and April 2004; included studies comparing non-invasive imaging with intra-arterial angiography that met Standards for Reporting of Diagnostic Accuracy (STARD) criteria; extracted data to calculate sensitivity and specificity of non-invasive imaging, to test for heterogeneity and to perform sensitivity analyses; and categorised percent stenosis by the North American Symptomatic Carotid Endarterectomy Trial (NASCET) method. Results In 41 included studies (2541 patients, 4876 arteries), contrast-enhanced MR angiography was more sensitive (0.94, 95% Cl 0.88-0.97) and specific (0.93, 95% CI 0.89-0.96) for 70-99% stenosis than Doppler ultrasound, MR angiography, and CT angiography (sensitivities 0.89, 0.88, 0.76; specificities 0.84, 0.84, 0.94, respectively). Data for 50-69% stenoses and combinations of non-invasive tests were sparse and unreliable. There was heterogeneity between studies and evidence of publication bias. Interpretation Non-invasive tests, used cautiously, could replace intra-arterial carotid angiography for 70-99% stenosis. However, more data are required to determine their accuracy, especially at 50-69% stenoses where the balance of risk and benefit for carotid endarterectomy is particularly narrow, and to explore and overcome heterogeneity. Methodology for evaluating imaging tests should be improved; blinded, prospective studies in clinically relevant patients are essential basic characteristics.