Results from False Positive Reduction

Table 11.15 summarizes the results from applying the false-positive reduction strategy to 200 abnormal segmented DDSM mammograms. Three sets of results are shown for each stage in the false-positive approach described for each breast type grouping.

The first column shows the sensitivity and average number of false positives per image following mammogram segmentation. The segmentation was obtained by combining 10 segmentation expert outcomes using the AWM described earlier. Each expert was constructed using the WGMM constrained with a MRF utilizing a supervised learning approach WGMMiSrf.

The second column shows the sensitivity and average number of false-positives regions per image obtained after applying the region prefiltering. These results demonstrate the utility of the region prefiltering stage. The average number of false-positive regions per image has dropped from approximately 147 to just 9 when testing on the complete dataset of 200 abnormal mammograms. This result has been obtained at a reduction in the sensitivity to the detection of breast lesions, from 0.63 to 0.60, for all breast types.

The final column shows the results obtained after classifying each region passing the prefiltering using an optimized ANN based on the 37-dimensional PCA feature vector for each sample. Using ROC analysis, the threshold for the detection of positive cases is set using the operating point of each ANN [40]. From these results it can be seen that the sensitivity is reduced still further to just 0.54 for all 200 abnormal mammograms, with a reduced average number of false-positive regions per image of 3.84. The results indicate that the biggest drop in sensitivity is obtained for the fatty breasts, breast types 1 and 2. This may be attributed to the increased variability of breast lesions in these breast types compared with that of the denser breasts.

0 0

Post a comment