Precise mediastinal lymph node (LN) staging is imperative in otherwise operable non-small cell lung cancer (NSCLC), as it determines subsequent treatment and possible inclusion in a neoadjuvant trial. The roles of mediastinoscopy and computed tomography (CT) remain controversial. To determine the accuracy of current CT scanners, a prospective study was performed. From April 1993 until September 1995, 100 consecutive patients with NSCLC without distant metastases underwent staging by CT and cervical mediastinoscopy. Naruke's map was used for classification, and LNs larger than 1 cm were considered CT positive. There were 91 males and 9 females, with a mean age of 64 (range 45-82) yrs. Fifty nine tumours were central and 41 peripheral, 64 right-sided and 36 left-sided. Thoracotomy with mediastinal LN sampling was performed in 74 patients, nonoperated patients having multilevel stage IIIA or stage IIIB disease. Twenty five (25%) mediastinoscopies were positive and three were false-negative (3%). There were 29 false-positive CT scans and 12 false-negative. Overall sensitivity and specificity of CT were 63 and 57%, respectively, and of mediastinoscopy 89 and 100%, respectively. Positive and negative predictive values of CT were 41 and 77%, respectively, and of mediastinoscopy 100 and 96%, respectively. Accuracy of CT was 59% and of mediastinoscopy 97%. Accuracy of CT was lowest for left-sided and centrally located tumours, and for LN station 7. Even with current computed tomography scanners, sensitivity and specificity remain low. Although overall cost may increase, routine cervical mediastinoscopy is necessary for precise staging of non-small cell lung cancer, and subcarinal lymph nodes should be routinely sampled.