A CBR-based CAD system reasons based on stored
knowledge prior cases with associated ground truth and its
final diagnosis suggestion is based on the ground truth of the
stored cases that are most similar to the query case. Hence,
its reasoning process is much easier to comprehend for the
physician than an ANN system that acts like a black box. In
this paper we propose a novel approach, based on decisiontree
learning DT, for building a CAD system that predicts
breast cancer biopsy outcomes based on BI-RADS compliant
lesion descriptions. Similar to the cited CAD systems based
on case-based reasoning, a CAD system based on decision
trees features a very transparent reasoning process. However,
in contrast to a CBR system, a decision-tree learner abstracts
a global model of the decision process from the prior cases
with associated ground truth instead of directly using these
cases in the decision process. This global model is even
easier to understand and predictions based on it are even
easier to comprehend for the physician than those of a CBR
system. In a second approach, we propose an extension of
the state of the art CBR approaches that features an entropic
distance measure. It provides a solid mathematical basis for
measuring the similarity of cases that have attributes of different
types e.g., nominal and numeric. It furthermore provides
a clean mathematical foundation for handling missing
attributes, in contrast to previous approaches.15 We have
evaluated the proposed CAD approaches on two large, publicly
available mammography databases to compare both
their performance and features.