Information Technology Reference
In-Depth Information
correspondence score is selected as the corresponding candidate region. The
proposed approach demonstrates an ability to discriminate between true and
false links.
In [18], Van Engeland and Karssemeijer extend this matching approach by
building a cascaded multiple-classifier system for reclassifying the level of suspi-
ciousness of an initially detected region based on the linked candidate region in
the other view. Experiments show that the lesion-based detection performance of
the two-view detection system is significantly better than that of the single-view
detection method. Paquerault et al. also consider established correspondence be-
tween suspected regions in both views to improve lesion detection ([19]). LDA
is used to classify each object pair as a true or false mass pair. By combining
the resulting correspondence score with its one-view detection score the lesion
detection improves and the number of false positives reduces. Other studies on
identifying corresponding regions in different mammographic views have also
used LDA and Bayesian artificial neural networks, in order to develop classifiers
for lesions [20,21]. The aim is, once again, identifying a lesion rather than giving
an overall interpretation of the case based on its related images.
In [22], the authors proposed a neural network approach to compute the likeli-
hood of paired regions on MLO and CC views, which are within similar distance
from the nipple (a common landmark, used by radiologists for finding correspon-
dence between view regions). Their system keeps the same case-based sensitivity
(true detection rate) while reducing significantly the case-based false positives.
In another study [23], some of the previous authors examined further the ef-
fect of three methods for matching regions on both views on the performance
of the CAD system. The matching was based on different search criteria for
corresponding region on the other view. Results showed that the straight strip
method required a smaller search area and achieved the highest level of CAD
performance.
In recent study [24], Wei et al. extended a previous dual system for mass de-
tection trained with average and subtle masses, with a two-view analysis to im-
prove mass detection performance. Potential links between the mass candidates
detected independently on both views are established using regional registration
technique and similarity measure is designed using paired morphological and
texture features to distinguish between true and false links. Using a simple ag-
gregation strategy of weighting the similarity measure with the cross-correlation
of the object pair, the authors obtained two-view score for malignancy, which
is further combined using Linear Discriminant Analysis (LDA) with the single-
view score to obtain the final likelihood for malignant mass. Experimental results
showed that the two-view dual system achieves significant better performance
than the dual and the single-view system for average masses and improves upon
the single-view system only for subtle masses, which are by default more dicult
to classify.
All these studies dedicated their attention on establishing whether a region
in one view corresponds to a region in the other view and demonstrate improve-
ment in the localised detection of breast cancer, mostly for prompting purposes.
Search WWH ::
Custom Search