Double Network Mapping Overview

The double network mapping (DNM) method is used to predict the target contrast enhancement using two ANNs for each enhancement method. The aim is to learn a mapping based on a set of gray-scale features FROI from a given mammogram, with a quantitative measure of segmentation performance, S. The segmentation performance is quantified following contrast enhancement, for each enhancement method m, where 1 < m < M from a set of M enhancement methods. The two submappings are detailed below:

1. ANNmNMenh: Foramammogram I, enhancedusing enhancement methodm, this ANN learns the mapping between the set of F gray-scale input features FROI = (f1, f2,..., fF) extracted from a suspicious ROI, and a set of P quantitative measures Q = (q1, q2,...,qP) of enhancement performance as described previously in section 11.3.1.

2. ANNmNMseg: For a mammogram I, enhanced using enhancement method m, the ANN learns the mapping between the set of quantitative measure Q = (q1, q2, ...,qP) of enhancement performance and a set of R measures quantifying the performance of lesion segmentation S = (Si, S2,..., SP) identified in Table 11.2.

A diagrammatic overview of the mappings learnt is given in Fig. 11.6 and the training and testing phases are described in more detail below. To evaluate the strategy, a firefold cross-validation approach is used to reduce bias and ensure that a test result is produced for each mammogram image.

11.3.4.1.1 Training the DNM Approach. Using this strategy, ANNmNUfmh and ANNmNUseg are trained independently for each enhancement method, Em where m e {1,..., M}. For a training image, a border comprising normal pixels of the same areas as the target ROI is constructed around it. The set of grayscale input features FROI are extracted from the target ROI and background regions as described in section 11.3.3.1. Each training mammogram is contrast enhanced with each method and a set of quantitative measures of enhancement

Figure 11.6: Diagrammatic overview of the DNM strategy.

Q are calculated from the target ROI and border. Thus ANNmNMenh learns the mappings:

Similarly for enhancement method Em to train ANNmNMseg, the set of quantitative measures of enhancement Q are used as input features in learning the mapping with the set of quantitative measures of segmentation, S. Thus ANNmNMseg learns the mappings:

11.3.4.1.2 Testing the DNM Approach. The first step in determining the optimal contrast enhancement method for a test mammogram I is to locate a suspicious ROI. To do this, the HMRFu segmentation algorithm is used to segment the test image and it results in several candidate regions. Regions with a Euler number > 1 (i.e, enclose a smaller region totally) are removed and from the remaining regions, the most likely suspicious regions are selected on the basis of area and morphological tests using a previously trained ANN. For the single suspicious ROI identified, a surrounding border is constructed of equal area, and the set of input gray-scale features FROI are extracted. These are used as inputs to ANNmNMenh for each enhancement method Em where m e{ 1,..., M}. The output of these networks is then supplied as input to ANNmNMseg for each enhancement

Table 11.6: Results from using optimized strategy DNM showing the mean percentage improvement in segmentation performance compared with the segmentation of the unenhanced original mammogram

A

Total

1

-0.16

1.75

0.10

2

0.28

-0.09

0.14

3

0.20

1.00

0.39

4

0.00

2.00

0.31

Mean

0.08

1.16

0.24

method m e{1,..., M}. The mth ANNmNMseg maximising the value for the sum of TP and SUBTP outcomes is predicted as the optimal enhancement method for a test mammogram I, thus assign Em ^ m if ANNmNMseg = argmax^^ ANNmNMseg for Vm ={1,..., M}

11.3.4.1.3 Model Order Selection. Optimization of the model order for ANNDmNMseg and ANNmNMseg for each enhancement method Em, where m e {1,..., M}, is performed independently by varying the number of hidden nodes between 2 and 30. The mean squared error (MSE) resulting for each model on test using fivefold cross validation is minimized. The optimal number of hidden nodes selected is the one which minimizes the MSE over all configurations of hidden nodes.

11.3.4.1.4 DNM Framework Results. Table 11.6 shows the mean percentage improvement in segmentation performance compared with the segmentation of the unenhanced original image, using the predicted expert enhancement for each breast type. The DNM strategy results are significantly poorer than those obtained using the target expert contrast enhancement methods reported in Table 11.4. They are also inferior to the use of the FUZZY method on all breasts as shown in Table 11.5 (part c).

0 0

Post a comment