Risk stratification of probably benign mammographic lesions (BI-RADS 3) with Bayesian networks

• Applicant: Privatdozent Dr. Matthias Benndorf
• Subject Area: Nuclear Medicine, Radiotherapy, Radiobiology
• Term: from 2013 to 2014
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 251959626

Lesions detected in mammographic examinations which are assigned BI-RADS (Breast Imaging: Reporting and Data System) category 3 have a probability of malignancy less than two percent, per definition. There are several descriptors (e.g. round, clustered microcalcifications) that allow for the assignment of BI-RADS category 3. The aim of the project is to use Bayesian networks to examine whether stratification of these lesions into groups with higher, and groups with lower probability of malignancy is possible. Bayesian networks allow for the calculation of the positive predictive value, by taking the probability distributions of multiple diagnostic variables into account. Variables are represented as nodes, whereas the statistical dependencies among them are represented as edges. There are several algorithms available to induce a Bayesian network, among them being the naive case (conditional independence is assumed for all diagnostic variables) and a tree-augmented naive Bayes, that allows one additional dependency per variable.To derive reliable estimates of the probability distributions (for example, how many cancers are hyperdense, given that they have a spiculated margin), a huge amount of empirical data is necessary. The University of Wisconsin, Madison, has access to a dataset of 132.319 mammographies rated according to the BI-RADS lexicon. The examinations were performed between 1999 and 2011. The whole dataset was matched with the US national cancer registry. This means that there is a reliable disease status (malignant/benign) for observed changes in the examinations. Additionally, for every patient there is a record about family history of breast cancer, personal history of breast cancer, hormone replacement therapy status and breast density.With this data, classification algorithms for the BI-RADS 3 lesions are induced. A preliminary discussion has resulted in the conviction that the inclusion of an expert node into empirical Bayesian networks is a promising approach. This expert node will not get its probability distribution from the empirical data. It will solely be based on the scientific literature. For example the node distinguishes between typical and atypical BI-RADS 3 lesions. In this way it is possibly able to distinguish groups that were assigned BI-RADS 3 without proper justification, and have a much lower or much higher risk for malignancy. The clinical use of the whole approach is to lower the rate of equivocal findings in mammography, and therefore reduce the number of additionally required imaging.

• DFG Programme: Research Fellowships
• International Connection: USA
• Participating Institution: University of Wisconsin; School of Medicine and Public Health; Department of Radioloy
• Host: Professorin Elizabeth Burnside