ANN and BBN use different machine learning concepts for pattern recognition and have very different characteristics. Table 1 summarizes and compares the basic characteristics of these two networks.
In medical image processing, usually a large number of features are extracted. In a complex and multidimensional
TABLE 1 Characteristics of an ANN and a BBN
Characteristic ANN BBN
Optimization method Simple heuristic approach Probabilistic approach
Explainable learning process No Yes
Flexibility in topology design No Yes
Weights inside the network Trained Computed or assigned
Possible data overfitting Yes No
Limitation on input data type No Yes (only discrete data)
Implementation Easy Difficult space, it is often difficult to use conventional mathematical analytic methods, such as linear discriminant function or polynomial regression, to define the separation boundary between different classes. The ANN has been used as the classifier in a large number of computer-assisted diagnosis systems in medicine . The biggest advantage of the ANN is that it is relatively easy to implement, since the ANN does not need much human intervention and knowledge of the problem. ANNs can learn and automatically define boundaries between different pattern clusters in a multidimensional space based on the optimization approach of "hill-climbing," even when the training data contain noise and incomplete information. However, as a heuristic optimization method, there is no guarantee that the ANN will find the most accurate model overall. To help the ANN avoid being trapped inside a locally optimal solution, an appropriate momentum (a) and learning rate should be applied. The major concern about the ANN is that it acts like a "black box" and cannot meaningfully explain its reasoning .
Because of the "black box" approach and possible data overfitting in training an ANN, some researchers found that it might be difficult to convince the physicians to accept and act on advice from an ANN-based computer-assisted system . In contrast, a BBN can explain the reasoning process and offer an efficient and principled approach for avoiding data over-fitting . Thus, some investigators suggested that a BBN should be more reliable than an ANN used as a computerassisted tool for diagnosis of medical images . However, there was no experimental evidence to support that a BBN is better than an ANN in medical image processing. In fact, an important principle in pattern classification indicates that feature extraction is the most crucial aspect of classification. If features are not capable of discriminating classes of interest at all, the resulting recognition performance will be poor regardless of what decision rules and classifiers are employed. On the other hand, if features exhibit simple statistical structures and separate classes well, various decision rules will make little appreciable difference in the overall classification performance . Recently several studies have been reported to compare the performance of ANNs and BBNs in detecting masses from digitized mammograms  and diagnosing breast cancers from a set of mixed image and nonimage medical features . These studies demonstrated that after proper training, both an ANN and a BBN yielded very similar performance under the same testing conditions. Thus, the conclusion is that the kind of classifier may not be as important as the quality of training database and input features. Future efforts aimed at improving the performance and generalization of an ANN or a BBN in medical image processing will benefit considerably from the refinement of training databases and feature selection.
Was this article helpful?