Information Technology Reference
In-Depth Information
mixture of kernels to detect maximum extremum points of microcalcifications.
Experimental results with benchmark mammographic data of 22 mammograms
have demonstrated computational eciency (detection per image in less than 1
second) and the potential reliability of the classification outcome. However, these
results required a better validation with a larger dataset in order to provide an
insight for the clinical application of the method.
In [14], a logistic generalized additive model with bivariate continuous inter-
actions is developed and applied for automatic mammographic mass detection.
The main goal is to determine the joint effect of the minimum and maximum
gray level value of the pixels belonging to each detected region on the probability
for malignant mass, depending on the type of breast tissue (fatty or dense). The
results on a large dataset of detected mammographic regions showed that the
breast tissue type plays a role in the analysis of detected regions. Advantage of
this method is its insightful nature, allowing interpretation and understanding of
the results obtained from the CAD system. However, this approach focuses only
on the lesion localisation on a mammogram rather than on the case classification
using multiple images.
Despite the potential of these studies some general drawbacks are observed.
On the one hand, the methods using descriptors provided by the human expert
allow evaluation on a patient level-a desired outcome in screening. However,
they are highly dependent on the expertise level and subjective evaluation of
the reader, which might lead to high variation in the classification results. On
the other hand, the approaches developed to automatically extract features from
the mammograms allow more consistent and ecient computer-aided detection.
But they tend to analyse independently every image belonging to a patient's
examination and prompt localised suspicious regions rather than to use the in-
formation from all the images to provide an overall probability for the patient
having cancer. To overcome some of the limitations of these single-view CAD sys-
tems, a number of approaches that deal with multi-view breast cancer detection
have been suggested, as discussed in the next section.
3.2 Multi-View Mammographic Analysis
In [15] Gupta et al. use LDA to build a model that also makes use of BI-RADS
descriptors from one or two views. The results of two classifiers, one for MLO
and one for CC, are combined in order to improve the classification of the CAD
system. However, an overall interpretation of a case-based on its related images is
not taken into account. In [16], Good et al. have proposed a probabilistic method
for true matching of lesions detected in both views, based on Bayesian network
and multi-view features. The results from experiments demonstrate that their
method can significantly distinguish between true and false positive linking of
regions. Van Engeland et al. develop another linking method in [17] based on
Linear Discriminant Analysis (LDA) classifier and a set of view-link features
to compute a correspondence score for every possible region combination. For
every region in the original view the region in the other view with the highest
Search WWH ::
Custom Search