A Bayesian Network Model of Mammography Findings
We have built a bayesian network to serve as decision support tool for radiologists interpreting mammograms. The radiographic findings reported on mammography have been standardized using the BI-RADS terminology. The probability of different breast diseases depends on the different BI-RADS findings (descriptors) observed. Our bayesian network relates probabilistic relationships among these BI-RADS descriptors with breast diseases. The BI-RADS terminology is a taxonomy of terms, with each term having different possible distinctions (possible values for that descriptor). This is shown in the figure to the right.
Our model is shown in this figure at the left. BI-RADS terms are shown as circles (nodes) surrounding the central "disease" node, that represents the possible breast diseases. Each node has several discrete states, corresponding to the possible values for the corresponding BI-rads descriptor.
This table shows the different diseases (benign and malignant) that our model assesses.
In addition to using the Bayesian network to assess the probabilities of different diseases of the breast, this network can be used to help the radiologist interpret the concordance of histopathology of breast biopsy with mammography findings. Because the observance of particular combinations of BI-RADS findings on mammography implies that some diseases are much more likely than others, certain histopathologies will be discordant with the mammography findings.
The disease probabilities provided by our model can be combined with the histopathology to produce a probability of sampling error. We recently studied 92 patients who underwent breast biopsy and compared the ability of the Bayesian network to predict concordant and non-concordant cases compared with experts. The network performed extremely well, detecting all non-concordant cases (100% sensitivity) with few false positives (calling a concordant case "non-concordant"; 91% specificity). The histogram above shows that the network produced a high probability for sampling error for the non concordant cases. Among the concordant cases, most had a low probability of sampling error.