Results of Applying Image Segmentation Expert Combination

The aim of our experiments was to (i) perform a comparison between the four proposed models of image segmentation. The baseline comparison with a simple GMM based image segmentation and an MRF model in [18] shows that our proposed models easily outperform the baseline models. (ii) To compare the performance of the AWM combination strategy against the ensemble combination rules. Section 11.4.5.1 compares the four models on the two databases, and section 11.4.5.2 compares the AWM approach with ensemble combination rules approach on the two databases.

Our segmentation performance evaluation is performed on 400 mammograms selected from the DDSM. The first 200 mammograms contain lesions and the remaining 200 mammograms are normal (used only for training purposes). Each of these mammograms has been categorized into one of the four groups representing different breast density, such that each category has 100 mammo-grams. The partitioning of the mammograms has been performed manually on the basis of the target breast density according to DDSM ground truth. The results will be reported in terms of the Az value that represents the area under the ROC curve as well as sensitivity (the segmentation evaluation for testing is based on ground-truth information as given in DDSM).

The grouping of mammograms by breast density is applicable only to the supervised approaches. Supervised approaches segmenting a mammogram with a specific breast density type use a trained observed intensity model constructed with only training samples from that breast type. Thus, each trained observed intensity model will be specialized in the segmentation of a mammo-gram with a specific breast type. We adopt a fivefold cross-validation strategy. Using this procedure, a total of five training and testing trials are conducted, and each time the data appearing in training does not appear as testing. For each of the fivefolds, equal numbers of normal and suspicious pixels are used to represent training examples from their respective classes. These sample pixels are randomly sampled from the training images. In the unsupervised

Table 11.10: Mean AZ for each breast type and segmentation strategy.

Table 11.10: Mean AZ for each breast type and segmentation strategy.

Breast type

WGMMS

wgmmMrf

WGMMU

wgmmMrf

1

0.68

0.70

0.66

0.59

2

0.66

0.66

0.66

0.60

3

0.72

0.80

0.75

0.75

4

0.66

0.76

0.68

0.74

Mean

0.68

0.73

0.68

0.67

Winning strategies are given in bold.

Winning strategies are given in bold.

case, there is no concept of training and testing and each image is treated individually.

0 0

Post a comment