The reliability of cranial ultrasound diagnosis in the premature neonate was examined using data from an ongoing multicentre study of the epidemiology and long-term consequences of neonatal brain haemorrhage. First week ultrasound films (obtained at 4 hours, 24 hours and 7 days) from 60 study subjects were randomly selected for independent review by two groups of experienced interpreters, and results were recorded separately for observations (i.e. presence or absence of an abnormal echodense area on a film) and interpretations (i.e. presence or absence of haemorrhage or ventricular dilatation) in each hemisphere. Because of deaths in the first week of life, the total number of films examined was 138. Concordance on the presence or absence of an abnormal echodensity was examined for each individual film for three areas of interest: the germinal matrix, the ventricles and the parenchyma. Concordance on the presence or absence of haemorrhage or ventricular dilatation was examined only for the seventh-day film, or the final film prior to death. Finally, concordance was analysed with the diagnostic interpretations grouped into categories thought to differ prognostically for long-term outcome. In general, concordance was poorest for germinal matrix lesions and best for parenchymal lesions. Concordance was lower for observations made on each individual film than it was for interpretation of the final film in each case. Fifty-five of 60 cases (92%) were assigned to the same major prognostic category by both readers. Ultrasound review conferences were held periodically and there was evidence that concordance in ultrasound reading and interpretation improved during the course of the study.