An Advanced Probabilistic Framework for Assisting Screening Mammogram Interpretation - Computational Intelligence in Healthcare: Advanced Methodologies - page 375

where N is the number of records for a unit, Cl i E i is the class value

and the feature vector of the i -th observation, respectively. Thus, the value of
ALL ( Cl ) indicates how close the posterior probability distribution is to reality:
when P ( Cl i |E i )=1thenln P ( Cl i |E i ) = 0 (no extra information); otherwise
−

and

ln P ( Cl i |E i ) > 0.
The log-likelihood results are given in Table 3. The lowest ALL ( C )isachieved
for the links meaning that the estimated probabilities best fit the link probability
distribution. A possible explanation is that in our Bayesian network framework
the links are directly dependent on the original region features and thus they are
better fitted. On the other hand, the rest of the units are based on combining
estimated probabilities from previous levels where noise could play a role. The
MV-CAD-Causal and MV-CAD-LR obtain comparable results for all units whereas

1

0.8

MV-CAD-Causal
AUC=0.847

0.6

MV-CAD-LR
AUC=0.835

MV-CAD-NB
AUC=0.833

0.4

SV-CAD
AUC=0.797

0.2

0 0.2 0.4 0.6 0.8

False positive rate

Fig. 8. ROC analysis per case

Table 3. Average log-likelihood of the class based on the multi-view scheme for differ-
ent units

Method
MV-CAD-Causal MV-CAD-LR MV-CAD-NB

Unit

Link 0.19

Region 0.38

MLO/CC 0.34/0.31

Breast 0.300 0.235 0.553

Case- max 0.482 0.483 0.790

Case- train 0.476 0.477 0.543