Results

This section presents the results from applying the false-positive reduction methodology on the suspicious regions resulting from the knowledge-based

Table 11.19: Sensitivity and average number of false-positives per image for 200 abnormal and 200 normal images

After segmentation After falsepositive After segmentation prefiltering reduction (OP)

Table 11.19: Sensitivity and average number of false-positives per image for 200 abnormal and 200 normal images

After segmentation After falsepositive After segmentation prefiltering reduction (OP)

Breast type

Sensitivity

FP/i

Sensitivity

FP/i

Sensitivity

FP/i

1

0.79

207.26

0.77

8.98

0.64

3.41

2

0.80

162.48

0.80

10.95

0.70

7.05

3

0.96

161.86

0.89

7.96

0.89

7.76

4

0.80

136.45

0.76

6.71

0.76

7.63

Mean

0.84

167.01

0.81

8.65

0.75

6.46

Values after segmentation, after region prefiltering, and after false-positive reduction using optimized classifier at ROC operating point, each by breast type (FP/ i = average number of false-positive per image, OP = operating point).

Values after segmentation, after region prefiltering, and after false-positive reduction using optimized classifier at ROC operating point, each by breast type (FP/ i = average number of false-positive per image, OP = operating point).

contrast enhancement and segmentation components. Table 11.19 shows the sensitivity in the detection of breast lesions and average number of false-positive regions over all mammograms of each predicted breast type. The first column of shows performance after segmentation, the second column shows the results after region prefiltering, and the final column shows the results following classification using the optimized ANN.

The aim of false-positive reduction is to reduce the average number of false-positive regions per image while maintaining sensitivity levels. The prefiltering stage can be seen from the results as being very effective in reducing the false-positive count. After region prefiltering sensitivity has dropped by just over 3.5% and the average number of false-positive regions to 8.65 for all 400 images. The mean sensitivity drops more sharply when reducing false-positive regions with the trained ANNs. The performance of each ANN is reported at the operating point on the ROC curve where the cost of a false positive and false negative are set equal (Cfp = CpN = 1) and the a priori probability of a positive case is set, P(D+) = 0.5. These costs and priors can be adjusted by the expert radiologist to reflect the required level of sensitivity. Optimization of their values is outside the scope of this study. The largest reduction in false-positive regions using the ANN is seen for the fatty, type 1 breasts. For the denser breasts, types 3 and 4, the operating point selected maintains the level of sensitivity but does not significantly reduce the false-positive count. In fact for the densest breasts, the false-positive count rises indicating the nontrivial nature of this classification problem.

0 0

Post a comment