OBJECTIVE To describe the endobronchial ultrasonographic characteristics and the cut-off value for diagnosis of peripheral lung cancer, and therefore to evaluate its diagnostic value. METHODS During June 1st, 2005 and June 30th, 2006, 78 patients with peripheral pulmonary lesions were enrolled. The lesions were all detectable by endobronchial ultrasonography (EBUS) and a final diagnosis was made. The endobronchial ultrasonographic structure of peripheral pulmonary lesions were analyzed, differentiated and classified into malignant or benign groups. RESULTS According to the result of binary multivariable logistic regression analysis on the 9 variables and by calculating the area under ROC curve, 5 variables were found to be useful in predicting the presence of malignancy: (1) clear borderline; (2) internal hypoechoic echo; (3) heterogeneous pattern; (4) without internal hyperechoic dots and linear arcs; (5) adjacent blood vessels representing shift, narrow or break-off. The equation of malignancy probability for any patient was: P = 1/[1 + e(-) (6.321-3.097X(2)-1.537X(1) + 1.898X(5) + 2.390X(3) + 3.003X(4))], X(1) for borderline; X(2) for internal hyperechoic dots and linear arcs; X(3) for adjacent blood vessels; X(4) for internal echo intensity; X(5) for internal echo distribution. The areas of ROC curve illustrated that multivariable logistic regression model discriminated benign and malignant lesions better than univariable logistic regression. The optimal cut-off value of the malignancy probability, which was greater or equal to 0.52 according to the ROC curve. This model gave a sensitivity and specificity of 87.2% and 80.6%, and the accuracy was 85.9%. CONCLUSIONS Endobronchial ultrasonographic characteristics of peripheral lung cancer included clear borderline, internal hypoechoic echo, heterogeneous pattern, without hyperechoic dots and linear arcs, and adjacent blood vessel shift, narrow or break-off. Multivariable logistic regression model discriminated benign and malignant lesions better than univariable logistic regression. Combination of multiple variables increases the sensitivity, specificity and accuracy for prediction of malignancy.