An Advanced Probabilistic Framework for Assisting Screening Mammogram Interpretation - Computational Intelligence in Healthcare: Advanced Methodologies - page 358

In reading mammograms, radiologists judge for the presence of a lesion by
comparing views. The general rule is that a lesion is to be observed in both
views (whenever available). CAD systems, on the other hand, have been mostly
developed to analyse each view independently. Hence, the correlations in the le-
sion characteristics are ignored and the breast cancer detection can be obscured
due to the lack of consistency in lesion marking. This limits the usability and
the trust in the performance of such systems. That is the reason why we develop
a Bayesian network model that combines the information from all the regions
detected by a single-view CAD system in MLO and CC to obtain a single proba-
bilistic measure for suspiciousness of a case. We explore multi-view dependencies
in order to improve the breast cancer detection rate, not in terms of individual
regions of interest but, at a case level.
In detail, the power of our model relies on
- Incorporating background knowledge with respect to mammogram reading
done by experts: multiple views are not analysed separately, and the cor-
relation among corresponding regions in different views are preserved as a
consequence;
- Analysing beyond the features of given regions: the interpretation is extended
up to case level. In the end, an interpretation to whether a given case can,
or cannot, be considered as cancerous is our ultimate goal;
- The design itself as representing a step forward on a better understanding of
the domain, establishing clearly the qualitative and quantitative information
of the relevant concepts in such domain.
Our design of a Bayesian network model in the context of image analysis has
important implications with respect to the structure and content of that network,
i.e., the understanding of the variables being used, and how those relate to one
another. This is seen as one of the major contributions of the present work.
The remainder of the chapter is organised as follows. In the next section we
give some introduction on terms of the domain, which will be used through-
out this chapter. In Section 3 we briefly review previous research in multi-view
breast cancer detection. In Section 4 we introduce basic definitions related to
Bayesian networks and in Section 5 we then describe a general Bayesian network
framework for multi-view detection. The proposed approach is evaluated on an
application of breast cancer detection using actual screening data. The evalua-
tion procedure and the results are presented in Section 6. Finally, conclusions
and directions for extension of our model are given in Section 7.
2 Terminology
To get the reader acquainted with the terminology used in the domain of breast
cancer and throughout this chapter, we introduce next a number of concepts.
By lesion we refer to a physical cancerous object detected in a patient. We call
a contoured area on a mammogram a region . A region can be true positive (for
example, a lesion marked manually by a radiologist or detected automatically