Information Technology Reference
In-Depth Information
tumour characteristics such as presence of spicules (star-like structure) and
focal mass. Based on these features, a neural network (NN) classifier is then
employed to compute the region likelihood for cancer. The locations with a
likelihood above certain threshold are selected as locations of interest.
3. Region segmentation with dynamic programming using the detected loca-
tions as seed points. For each region a number of continuous features are
computed based on breast and local area information, e.g., contrast, size,
location.
4. Region classification as “normal” and “abnormal” based on the region fea-
tures. A likelihood for cancer is computed based on supervised learning with
a NN and converted into normality score (NSc): the average number of nor-
mal regions in a view (image) with the same or higher cancer likelihood.
Hence, the lower the normality score the higher the likelihood for cancer.
Based on the single-view CAD system the likelihood for cancer for a view,
breast or case is computed by taking the likelihood of the most suspicious region
within the view, breast or case, respectively. Although this system demonstrates
good detection rate at a region level, its performance deteriorates at a case level.
One of the main reasons is that the single-view processing fails to account for
view interactions in computing the region likelihood for cancer. To overcome
this limitation, we consider the problem of multi-view breast cancer detection,
presented in the next section, and subsequently we propose a Bayesian network
framework to model this problem.
5.2 Multi-View Problem Description
The objective of mammographic multi-view detection is to determine whether
or not the breast and respectively the patient exhibits characteristics of abnor-
mality by establishing correspondences between two-dimensional image features
in multiple breast projections. Figure 5 depicts a schematic representation of
multi-view detection.
MLO
CC
B 1
L 11
L 12
L 21
L 22
A 1
B 2
A 2
Fig. 5. Schematic representation of mammographic multi-view analysis with automat-
ically detected regions
Search WWH ::




Custom Search