Information Technology Reference
In-Depth Information
A lesion projected in MLO and CC views is displayed as a circle; thus, the
whole breast and the patient is cancerous. An automatic single-view system
detects regions in both views and a number of real-valued features are extracted
to describe every region. In the figure regions
A
1
and
B
1
are correct detection
of the lesion, i.e., these are true positive (TP) regions whereas
A
2
and
B
2
are
false positive (FP) regions. Since we deal with projections of the same physical
object we introduce links (
L
ij
) between the detected regions in both views,
A
i
and
B
j
. Every link has a class (label)
L
ij
=
ij
defined as follows
ij
=
true
if
A
i
OR
B
j
are
TP
,
(2)
false
otherwise
.
This definition allows us to maintain information about the presence of cancer
even if there is no cancer detection in one of the views. In contrast to the clinical
practice, in the screening setting the detected lesions are usually small and due
to the breast compression they are sometimes dicult to observe in both views.
A binary class
C
,withvaluesof
true
(presence of cancer) and
false
, for region,
view, breast and patient is assumed to be provided by pathology or a human
expert.
In any case, multiple views corresponding to the same cancerous part contain
correlated characteristics whereas views corresponding to normal parts tend to
be less correlated. For example, in mammography an artefactual density might
appear in one view due to the superposition of normal tissue whereas it dis-
appears in the other view. To account for the interaction between the breast
projections, in the next section we present a Bayesian network framework for
multi-view mammographic analysis.
5.3 Model Description
The architecture of our model for two-view mammographic analysis is inspired
by the way radiologists analyse images, where several levels of interpretation are
distinguished. At the lowest image level, radiologists look for regions suspicious
for cancer. If suspicious regions are observed on both views of the same breast,
then the individual suspiciousness of these regions increases implying that a
lesion is likely to be present. As a result, the whole breast as well as the exam
and patient is considered suspicious for cancer; otherwise, the breast, exam and
patient is considered normal.
From a CAD point of view the first step-identifying suspicious regions-have
already been tackled by the single-view CAD systems developed by our group.
Furthermore, in [31,32] we proposed a two-step Bayesian network framework for
multi-view detection using the regions from the single-view system. Here, we
extend this system to create an unified framework to model the following stages
in the mammographic analysis as described above. Figure 6 represents the new
advanced framework.
At first we compute the probability that a region in one view is classified as
true
given its links to the regions in the other view. A straightforward way to
Search WWH ::
Custom Search