Overlap Analysis of Segmentation Results

The results in the previous section show that the performance obtained following ROC analysis of the knowledge-based segmentation component is greater than that obtained from the best performing segmentation expert. By thresholding each probability image using a ROC operating point following optimal expert combination, region boundaries can be identified. In general, the ROC operating point [40] can be selected for each individual mammogram by associating a cost for a false positive, CFp, and a false negative, CFN. In this chapter, the operating point cannot be determined using this method. This is because the ground truth knowledge cannot be used during testing.

To determine an estimate of the operating point, the mean operating point is calculated from all mammograms contained within a training fold. Only mam-mograms that following segmentation, give lesion detection with an operating point greater than 0.95 are considered. The mean operating point is calculated from each training fold, for each breast type. To compute each operating point, the relative cost of a false positive is chosen as CFP = 1 and for a false negative CFN = 20. In addition, the probability of apositive outcome, P(D+) = 0.03, computed as the mean percentage of abnormal pixels in all training mammograms.

Table 11.16: Sensitivity and average number of false positives per image after segmentation of 200 abnormal images using adaptive knowledge-based model

Type

Sensitivity

FP/i

1

0.79 (0.57)

175.03 (163.31)

2

0.80 (0.58)

172.42 (132.09)

3

0.96 (0.72)

136.48 (132.55)

4

0.79 (0.66)

121.47(158.79)

Mean

0.84 (0.63)

151.35 (146.73)

This is compared to the results (in brackets) from WGMMf experts combined using AWM model.

This is compared to the results (in brackets) from WGMMf experts combined using AWM model.

Following identification of region boundaries, overlap analysis is performed using the outcomes described in Table 11.2. The sensitivity results and average number of false-positive regions per image are shown in Table 11.16. The average number of false-positive regions per image decreases as the breast type increases. This can be attributed to the stricter ROC threshold used for thresholding these groups of probability images. Note that these results improve significantly on those without using adaptive model. This is because of the difficulty in selecting values for the costs CFP and CFN when setting the operating point. Different image segmentations result in different distributions of a posteriori estimates of positive (abnormal) and negative (normal) pixels. Ideally CFP and CFN need to be optimized on a per image basis, but this optimization is outside of the scope of this study.

0 0

Post a comment