Abstract Background Virtual bronchoscopy (VB) could obviate flexible bronchoscopy (FB) if no endobronchial lesion is detected in patients presenting with a suspicion of malignancy. Our objectives were to evaluate the accuracy (in terms of sensitivity and specificity) of VB in detecting endobronchial lesions, and to determine the anatomical limit of detection of endobronchial lesions by VB. Methods This study involved, in a blind comparison of VB and FB, consecutive patients presenting with symptoms or plain chest radiography abnormalities raising the suspicion of pulmonary neoplasm. After the standard chest computed tomography (CT), additional helical CT data were acquired from the aortic arch to the origin of the segmental bronchi of the inferior lobes in one 20-second breath hold using an helicoidal CT scan (3.0-mm collimation with a pitch of 1.5 and 1.5-mm reconstruction intervals). Results One hundred ninety patients were enrolled; 136 patients (including 63 with an endobronchial lesion at FB) contributed to the primary analysis. The sensitivity and specificity of VB to detect endobronchial lesions were 68% (95% confidence interval [CI]: 55% to 79%) and 90% (95% CI: 81% to 96%), respectively. Overall, the agreement between VB and FB regarding the location on endobronchial lesions was substantial (weighted kappa: 0.66). However, VB detected only 26 of the 34 lobar lesions (sensitivity: 76%; CI: 59% to 89%) and 11 of the 23 segmental lesions (sensitivity: 48%; CI: 27% to 69%). Conclusions Beyond the mainstem bronchi, VB is not accurate enough to detect endobronchial lesions and to obviate FB in patients presenting with a suspicion of malignancy.