Background: A reliable method of infarct volume measurement is needed if infarct volume is to be used as an outcome measure in clinical stroke trials. We investigated the reproducibility of a semi-automated method of computed tomography (CT) infarct volume measurement amongst three stroke research fellows with no formal neuroradiology training and two consultant neuroradiologists. Methods: CT brain scans for volumetric analysis were performed at 5 to 7 days in 34 patients with acute ischaemic stroke, of which 28 scans showed visible recent infarction. Five observers independently traced the infarct boundary on digitised images with a cursor. Volumetric analysis incorporated pixel thresholding with preset Hounsfield thresholds. One of the observers repeated the analyses on 21 of the scans in order to assess intraobserver variation. Results: Median infarct volume was 35.7 cm³ (range 0.2-318 cm ³). The closest limits of observed agreement (mean ± 1.96 SD) between pairs of observers were between a research fellow and neuroradiologist (-29 to 21 cm³). The widest limits of agreement were between a different research fellow and the same neuroradiologist (-39.1 to 41.4 cm ³). The limits of agreement between infarct volumes measured on two separate occasions by one of the research fellows were -7 to 8 cm³. Conclusions: Intraobserver reliability of CT infarct volume measurements performed by a stroke research fellow was superior to interobserver reliability between any pair of observers. The wide limits of agreement between different observers using manual tracing may not be acceptable in multicentre trials of acute ischaemic stroke treatment, but volume measurement by a single observer appears to be more reliable.