Information Technology Reference
In-Depth Information
MV-CAD-NB
leads to considerably larger average log-likelihoods, especially at a
case level using the MAX combining scheme. The latter can be explained by the
nature of the naive Bayes classifier, which assumes independence of the views and
leads to less accurate breast probabilities as indicated in Tables 1 and 2. With
the TRAIN combining scheme, the case probabilities computed by
MV-CAD-NB
are better fitted to the truths.
7 Discussion and Conclusions
In this work, we proposed a unified Bayesian network framework for automated
multi-view analysis of mammograms, which is an extension of the work presented
in [31,32]. The new framework is based on a multi-stage scheme, which models
the way radiologists interpret mammograms. More specifically, using causal in-
dependence models, we combined all available image information for the patient
obtained from a single-view CAD system at four different levels of interpretation:
region, view, breast and case. In comparison to our previous work, where the
breast and case classification were done using logistic regression, the new scheme
allowed a better representation of the multi-view dependencies and handling un-
certainty in the estimation of the breast/case probabilities for suspiciousness.
Furthermore, based on experimental results with actual screening data, we
demonstrated that the proposed Bayesian network framework improved the de-
tection rate of breast cancer at lower false positive rates in comparison to a
single-view CAD system. This improvement is achieved at view, breast and case
level and it is due to a number of factors. First, we built upon a single-view CAD
system that already demonstrates relatively good detection performance. By ap-
plying a probabilistic causal model we linked the original features extracted by
the single-view CAD system for all the regions in MLO and CC views and we
combined all the links for one breast to obtain a single measure for suspicious-
ness of a view, breast and case. Another factor for the improved classification is
that our approach incorporated domain knowledge. Following radiologists' prac-
tice, we applied a straightforward scheme to account for multi-view dependencies
such that (i) correlations between the regions in MLO and CC views are consid-
ered per breast as whole and (ii) the classification of breast/case as “suspicious”
is employed through the logical OR. Thus the proposed methodology can be
applied to any domain (e.g., fault detection in manufacturing processes) where
similar definitions and objectives hold.
We also note important advantages of our probabilistic system in compari-
son to previous approaches for automated mammographic analysis. On the one
hand, most Bayesian network-based systems for breast cancer detection require
that radiologists provide categorical description of the findings a priori to clas-
sification. In contrast, our causal approach is completely automated and it is
entirely based on the image processing results from the single-view CAD sys-
tem. This implies that in screening setting, where time is a crucial factor, the
proposed multi-view approach has more realistic application and it can be of
great help to the human expert. On the other hand, fully automated approaches
Search WWH ::
Custom Search