Objective: To examine the potential of different constructs of Bayesian belief networks (BBN) to manage uncertainty in breast cancer grading.
Study design: We developed four networks, two based on Bloom-Richardson's and two on Helpap's grading systems. The function of the networks was based either on an expert's experience or frequency counts derived from subjective grading of a large number of samples. The four BBNs were tested on 20 specimens, and the resulting final beliefs were compared with the subjective gradings.
Results: The BBNs showed agreement with the subjective gradings in 60-85% of cases. Different constructs of BBNs, however, differed in their performance. The mean beliefs in frequency-based networks were slightly higher than in experience-based networks. In addition, as compared with the Bloom-Richardson-based networks, the Helpap-based BBNs resulted in higher maximum beliefs but produced a larger fraction of discrepancies with the subjectively graded cases. Depending on the type of network, 65-90% of the BBN grades were associated with high beliefs.
Conclusion: The results suggest that for reliable results, grading systems with more than three or four variables may be necessary. When based on relevant information, BBNs seem to have the potential to become a valuable method of assisting the pathologist in breast cancer grading.