Four different classification experts were implemented: radial basis function (RBF) neural network, multilayer perceptron (MLP), K -nearest neighbour ( K -NN) and Gaussian classifiers using the feature sets FS1 to FS4 respectively. The Receiver Operating Characteristic (ROC) curve was then used to identify the a priori probabilities which will guarantee a 100% true positive detection.
Two different false positive figures of merit, denoted Error-1 and Error-2, were adopted as a basis for fusion strategy assessment. Error-1 measures the number of falsely detected clusters of microcalcifications (a group of at least three single MCs within a prespecified radius) per image. Error-2 reflects the percentage of normal images misclassified as abnormal. Clearly the two errors are not independent. They represent two significant characteristics of the microcalcification detection system. The former measure would be relevant if the system was intended as a source of second diagnostic opinion. On the other hand the second measure would be decisive if the system was to be used for pre-reading the mammograms in an extensive screening program.
The performance of the individual classifiers using the two different figures of merit is presented in Table 1 . The MLP classifier achieves a minimum error for both test sets while the K -NN classifier yields the worst performance in terms of Error-1. The RBF classifier gives the best performance on both test sets in terms of Error-2 while the Gaussian classifier produces the worst result on both sets.
Table 1:
Error rates produced by each expert on independent test sets A and B.
Table 2:
Error rates produced on independent test sets by various expert fusion
strategies.
S Ali Hojjatoleslami
