Iii

Number cJ in 11ยป Turing H jtKf:

FIGURE 4 Improvement in ANN performance with size increase of the training database.

and that it may require a very large, carefully selected sample set of cases to ensure that the entire variable domain is adequately covered. In reality, the training data samples that are selected, segmented, and used to optimize a classifier in diagnosis of medical images is very sparse in relationship to the feature space. Under this situation, one is more likely to "stress" the system during testing by using cases that were "never seen," because at least some of these may cover areas in the feature space that had not been (or at best had been sparsely) represented in the training set. Therefore, a large training database is often required to efficiently model the complex relationships among the medical findings, and another independent database also is required to test the performance and generalization of the network.

0 0

Post a comment