Mammographic Cluster Classification TwoView Application

In this application of the algorithm, the two views of the same cluster were combined for the selection of the classification features [59]. A total of 101 paired clusters were available for this test. The 14 features of Table 13.2 were first determined on the 101 CC and 101 MLO views of the cluster and then averaged. The set of average feature values were then used as input to the classification stage of the algorithm (Fig. 13.3). The computer generated ROC curves of the classification performance of the algorithm obtained with the symmlet wavelet and the donut filter are shown in Fig. 13.17. Similar to the previous experiment, the classifier with the donut filter outperformed the classifier with the

fc Ph H

Figure 13.17: Computer ROC plots of the TPF and FPF pairs obtained from the classification of 101 clusters using two-view information for feature estimation. The dashed curve corresponds to the results obtained with the symmlet wavelet filter and the solid curve corresponds to the results obtained with the donut filter. The estimated area indices AZ and corresponding SE values are included in the insert.

Figure 13.17: Computer ROC plots of the TPF and FPF pairs obtained from the classification of 101 clusters using two-view information for feature estimation. The dashed curve corresponds to the results obtained with the symmlet wavelet filter and the solid curve corresponds to the results obtained with the donut filter. The estimated area indices AZ and corresponding SE values are included in the insert.

symmlet wavelet. And both outperformed their respective performances on the single-view application. The results suggest that the combination of views for feature estimation seems beneficial to the classification process.

Two views could lead to the definition of more robust features improving classification performance independent of the segmentation method used in the process. But, is averaging the best approach to feature selection from the two mammographic views? Our results seem to indicate that averaging is promising. However, they are somewhat counterintuitive since averaging carries the risk of introducing a fuzziness to an otherwise good descriptor, i.e., a feature that was a good descriptor in one view but poor in the other may lose its robustness once averaged. So, should we average or possibly combine features from the two views for the generation of a larger feature set? The answer to this question is not clear and more work is needed to determine the best feature combination from two mammographic views.

0 0

Post a comment