Abstract Two consecutive series of 107 invasive carcinomas and 108 carcinomas in situ of the breast were considered retrospectively. The main symptomatic, clinical and cytological features were retrieved from clinical records, whereas the features at mammography and/or ultrasonography were reclassified at a blinded review by an expert radiologist. The power of predicting invasion on the basis of these features was then evaluated. The accuracy of such a prediction was compared to that of a mathematical model using the same diagnostic features by multivariate logistic regression analysis. Neither approach reached a satisfactory accuracy, but the radiologist's judgment (sensitivity 97.1%, specificity 81.9%, positive predictive value 98.4%, negative predictive value 71.6%, overall accuracy 95.8%) was slightly superior to the mathematical model (sensitivity 93.2%, specificity 87.9% positive predictive value 98.8%, negative predictive value 53.7%, overall accuracy 92.8%). Although some presentations are significantly associated with invasive or in situ status, diagnostic features alone are not sufficiently accurate to bypass histological confirmation, although they might give useful clinical indications.