Atypical adenomatous hyperplasia (adenosis) of the prostate: development of a Bayesian belief network for its distinction from well-differentiated adenocarcinoma.

The diagnosis of atypical adenomatous hyperplasia (AAH) of the prostate and its distinction from well-differentiated prostatic adenocarcinoma with small acinar pattern (PACsmac; Gleason primary grades 1 or 2) are affected by uncertainties that arise from the fact that the knowledge of AAH histopathology is expressed in descriptive linguistic terms, words, and concepts. A Bayesian belief network (BBN) was used to reduce the problem of uncertainty in diagnostic clue assessment, while still considering the dependencies between elements in the reasoning sequence. A shallow network was designed and developed with an open-tree topology, consisting of a root node containing two diagnostic alternatives (eg, AAH v PACsmac) and 12 first-level descendant nodes for the diagnostic features. Eight of these nodes were based on cell features, three on the type of gland lumen contents and one on the gland shape. The results obtained with prototypes of relative likelihood ratios showed that belief for the diagnostic alternatives is high and that the network can differentiate AAH from PACsmac with certainty. The features that best contributed to the highest belief were those concerning the nucleolar size, frequency, and location. In particular, after the analysis of five nucleolar features (prominent nucleoli, inconspicuous nucleoli, nucleoli with diameter greater than 2.5 micron, nucleolar margination, and nuclei with multiple nucleoli), the belief for AAH was 1.0, being already close to 1.0 when three were evaluated (the value range is 0.0 to 1.0; the closer to 1.0, the greater the belief). The contribution of the three features concerning the gland lumen contents (mucinous material, corpora amylacea, and crystalloids) was such that the final belief did not exceed 0.8. Results with the group of remaining features (eg, basal cell recognition, gland shape variation, cytoplasm appearance, and nuclear size variation) were slightly better. These features allowed a substantial accumulation of belief that was already greater than 0.9 when three were polled. However, the maximum belief value was never obtained. In conclusion, a BBN for AAH diagnosis offers a descriptive classifier that is readily implemented, and allows the use of linguistic, fuzzy variables, and the accumulation of evidence presented by diagnostic clues.