this might suggest that there could be many factors responsible for impeding the reproducibility of the results, ranging from the expertise, amount of stress and tiredness of the pathologists to the level of diligence when looking at the specimens Furthermore, various researchers implicitly acknowledge this problem when they recommend, for complicated cases, the triple approach: the agreement among the surgeon, the radiologist and the pathologist with their respective findings (clinical,mammographic and cytological). Due to the mentioned situations, a natural and sensible question arises: is it possible to construct computational support tools that help reduce this subjectivity? In this paper, we explore the possibility of extracting some “objective” features, from a database, relevant for determining the presence/absence of breast cancer using a graphical-modeling approach called Bayesian networks (BNs). Such an approach allows both the visual representation of the probabilistic interactions among variables of interest and the quantitative measure of the impact of those interactions. These two important properties permit to perform some inferential processes, such as prognosis and diagnosis. Thus, the main contributions of this paper are twofold: (a) we assess the performance of seven BN classifiers in order to determine their effectiveness for accurately diagnosing breast cancer using two real-world breast cancer datasets, and (b) by using such an approach, we measure the magnitude of the interobserver variability implicitly contained in these data.