Breast Profile Mapping Overview

The second strategy used for learning the expert contrast enhancement for a mammogram is the breast profile mapping (BPM) strategy. For a mammogram

I, enhanced using enhancement method Em, where m e{1,..., M}, the BPM strategy learns the mapping between the set of N gray-scale input features Fbpn detailed in section 11.3.3.1, and a l = {1., ...,L} indicates the target contrast enhancement for a training mammogram. Both feature sets {FBP316, {FBP26) are evaluated separately in their utility for learning the expert contrast enhancement. The expert l is based on a set of R measures quantifying the performance of lesion segmentation S = {s1; s2,..., sR} described in Table 11.2. The expert l is identified as the one maximising the sum of TPT and SUBTPT outcomes for each enhancement method Em where m e{1,..., M} as defined previously in Eq. (11.10).

Unlike the DNM strategy, this method utilizes a single classifier to predict the target contrast enhancement method. The k-nearest neighbor (k-NN) classifier has been show to be effective at learning nonparametric mappings with a small sample size [27] and for this reason it is employed in the knowledge-based contrast enhancement expert. To evaluate the strategy a fivefold cross validation is used to reduce bias and provide a test decision for each mammogram.

II.3.4.2.1 Training the BPM Approach. To train the BPM strategy, the set of gray-scale input features Fbpn = (f1; f2,..., fN), where N identifies the original and PCA feature sets (N = {316, 26}), are extracted from the segmented breast profile. Each training mammogram is contrast enhanced with each enhancement method. The quantitative measures of segmentation are calculated for the target ROI. For each enhancement method, the winning predicted enhancement method identified by the label l is used to learn the mapping between F and l with the k-NN classifier.

11.3.4.2.2 Testing the BPM Approach. To determine the predicted target enhancement method El for a test mammogram I, the set of gray-scale input features Fbpn are extracted from the segmented breast profile. Using the trained k-NN classifier, the predicted actual expert contrast enhancement is determined.

Table 11.7: Percentage improvement in segmenting an unenhanced mammogram compared to that obtained when segmenting the image enhanced using the predicted enhancement method from the optimized BPM strategy based on feature set FBP316

Type

ß

SUBTPß

Total

1

-0.16

1.50

0.07

2

0.11

0.91

0.41

3

0.10

4.75

0.96

4

0.47

7.50

1.44

Mean

0.13

3.66

0.72

11.3.4.2.3 Model Order Selection. In order that the BPM strategy is to perform optimally, the number of nearest neighbors k must be correctly set. For each input feature set FBP316 and FBP26 for different values of k the validation set error is plotted and the value of k corresponding to the least error is chosen.

11.3.4.2.4 BPM Framework Results

1. Feature set FBP316: Using an optimized value of k = 23, Table 11.7 shows the percentage improvement in segmentation performance when using the predicated actual enhancement method, compared with that obtained with the unenhanced original from the FBP316 set. These results show that the segmentation improvement obtained over the unenhanced image, when segmenting an image enhanced using a enhancement method predicted by the BPM strategy, is greater than that obtained using the DNM strategy predicted enhancement method. However, segmenting the BPM strategy's predicted enhanced image results in inferior performance to that using the target enhancement method identified in Table 11.4. The result for breast type 4, the densest breast type, shows a small improvement over using the FUZZY method, shown in Table 5 (part c), for all mammograms of that type.

2. Feature set FBP26: Using an optimized value of k = 19, Table 11.8 shows the percentage improvement in segmenting the unenhanced image compared

Table 11.8: Percentage improvement in segmenting an unenhanced mammogram compared to that obtained when segmenting the image enhanced using the predicted enhancement method from the optimized BPM strategy based on feature set FBP26

Type

ß

SUBTPß

Total

1

0.04

2.00

0.31

2

0.33

1.00

0.59

3

0.30

5.00

1.17

4

0.13

7.00

1.06

Mean

0.20

3.75

0.78

to that when segmenting the image enhanced using the predicted enhancement method by the optimized BPM strategy with the FBP26 feature set. These results indicate better performance than the DNM strategy but are still inferior to the segmentation using the target expert enhancement method shown in Table 11.4. The result for breast type 1-3 show an improvement over using the FUZZY method in Table 5 (part c) for all mammograms of that type. Comparing the results from the evaluation of the two feature sets, FBP26 and FBP316 from the BPM strategy, the results indicate that the feature set FBP26 is better suited to processing mammograms with breast types 1-3, whereas the feature set FBP316 gives better performance on the densest breast type, i.e., type 4. Interestingly, for both feature sets, the performance improvement is worse over the fattiest breast types, type 1, compared with the densest, type 4. This is because of the variability of optimal enhancement method for the fatty breast types, whereas the denser breasts tend to be optimal enhanced by the FUZZY method more often.

0 0

Post a comment