An Advanced Probabilistic Framework for Assisting Screening Mammogram Interpretation - Computational Intelligence in Healthcare: Advanced Methodologies - page 377

based on LDA and neural networks fail to account for uncertainty and an ex-
plicit representation of view dependences, as done in our probabilistic approach.
In addition, the former approaches aimed mostly in improving the detection of
the breast cancer at a regional level, whereas we focused on modelling the whole
image and breast context information for better distinction between cancerous
and normal patients-the ultimate goal of a breast cancer screening program.
Although we demonstrated that the proposed framework has the potential
to assist screening radiologists in the evaluation of breast cancer cases, a num-
ber of issues can be addressed for extending the proposed multi-view approach.
Currently, we consider only cases where both MLO and CC views are present.
However, in practice, for half of the patients taking subsequent mammographic
exams, only MLO views are available. The question then is to modify the system
such that cases with missing CC views can also be analysed. Another interesting
direction for extension is with respect to the region linking. Instead of linking
all regions from MLO view to all regions from CC view of the same breast, as
currently done, it would be natural to consider only the potential true links and
avoid a large number of false positive links. To do so, an important piece of do-
main knowledge used by radiologists in mammogram analysis is that a cancerous
lesion projected on MLO and CC views is located on approximately the same
distance from the nipple. Thus, links to regions that clearly violate this rule can
be ignored and the system can become more precise.
In conclusion, we emphasise that in the coming years, the development and
practical application of computer-aided detection systems is likely to expand
considerably. This trend is expected to be most observed in the screening pro-
grams where the mammograms are to be digitised, the breast cancer incidence
rates are very low, the workload is tremendous and the detection of breast cancer
is dicult due to the small and subtle changes observed. By providing a “sec-
ond opinion” in the image analysis, CAD can help radiologists increase cancer
detection accuracy at an earlier stage. This can create both health and social
benefits by saving women's lives, increasing families' well-being, reducing unnec-
essary check-up costs and compensating the shortage of radiologists. To create
reliable and widely applied CAD, however, it is crucial to integrate the knowl-
edge and the working principles of radiologists in the CAD development. The
multi-view CAD system described in this chapter is a promising step forward in
this direction.
Acknowledgement
This research is funded by the Netherlands Organisation for Scientific Research
under BRICKS/FOCUS grant number 642.066.605.
Appendix
Figure 9 depicts an example of a causal-independence model for breast cancer
prediction. We have two binary cause variables mass (MASS) and microcalci-
fications (MCAL), which are the two main mammographic indicators for the