An Advanced Probabilistic Framework for Assisting Screening Mammogram Interpretation - Computational Intelligence in Healthcare: Advanced Methodologies - page 370

A i / B j = ( x 1 , x 2 , …, x n )

A 1 B 1

A 2 B 2

Stage 1
R EG N ET

L 11 L 12

21 L 22

C A1 C B1 C B2 C A2

C A1 C A2 C B1 C B2

Stage 2
V IEW N ET

I A1 I A2 I B1

I B2

OR OR

MLO

CC

MLO CC NScM NScC

Stage 3
B REAST N ET

I 1

I 2

OR

Breast

LBr NScLBr RBr NScRBr

Stage 4
C ASE N ET

I 1

I 2

MAX

Case

Fig. 6. Bayesian network framework for representing the dependencies between multi-
ple views of an object

model a link L ij is to use the corresponding regions A i and B j as causes for the

link class, i.e., creating the so-called v-structure A i −→ B j , defined in

Section 4.1. Since the link variable is discrete and the regions are represented
by a vector of real-valued features ( x 1 ,x 2 ,...,x n ) extracted from an automatic
detection system, we apply logistic regression to compute P ( L ij = true

L ij ←−

| A i ,B j ):

exp β ij
0

x k

+ β ij
1

+ β ij
k

x 1 + ···

1+exp β ij
0

x k

P ( L ij = true | A i ,B j )=

+ β ij
1

+ β ij
k

x 1 + ···

where β 's are the model parameters we optimise and k is the total number
of features of regions A i and B j . Logistic regression ensures that the outputs

P ( L ij = true | A i ,B j ) lie in the range [0 , 1] and they sum up to one. Next we