CADMIUM II is a system for the interpretation of mammograms. A novel aspect of the system is that it combines symbolic reasoning with image processing, in contrast with most other approaches, which use only image processing and rely on artificial neural networks (ANNs) to classify mammograms. A problem of ANNs is that the advice they give cannot be traced back to communicable diagnostic inferences. Our approach is to provide advice based on explicit knowledge about the diagnostic process. To this end, we have conducted a knowledge elicitation study which looked at the descriptors used by expert radiologists when making diagnostic decisions about mammograms. The analysis of the radiologists' reports yielded a set of salient diagnostic features. These were used to inform the advice provided by the symbolic decision making component of CADMIUM II.