An Advanced Probabilistic Framework for Assisting Screening Mammogram Interpretation - Computational Intelligence in Healthcare: Advanced Methodologies - page 367

n

P ( e | C 1 ,...,C n )= P ( I k | C k ) .

f ( I 1 ,...,I n )= e k =1

Finally, it is assumed that P ( i k

|

c k ) = 0 (absent causes do not contribute to

the effect); otherwise, P ( I k C k ) > 0.

An important subclass of causal-independence models is obtained if the de-
terministic function f is defined in terms of separate binary functions g k ;itis
then called a decomposable causal-independence model [28]. Usually, all functions
g k ( I k ,I k +1 ) are identical for each k . Typical examples of decomposable causal-
independence models are the noisy-OR [30] models, where the function f repre-
sents a logical OR. A simple example for the application of the logical OR in the
domain of breast cancer is given in the Appendix, motivating the use of noisy-OR
models in the general theoretical framework presented in the next section.

|

5 Bayesian Multi-View Detection

The inputs for the proposed multi-view detection scheme are the regions detected
by a single-view CAD system presented in [17] and for completeness it is briefly
described here.

5.1 Single-View CAD System

As its name implies, the single-view CAD system analyses independently each
mammographic view of each breast. Figure 4 depicts a schematic representation
of the single-view system used in this study. It consists of the following four main
steps:

1. Mammogram segmentation into background area, breast, and for MLO, the
pectoral muscles.
2. Local maxima detection based on pixel-based locations. For each location
in the breast area a number of features are computed that are related to

Original
mammogram

Mammogram
segmentation

Local maxima
detection

Region
segmentation

Region
classification

0.91 0.62

Fig. 4. Stages in the single-view CAD system