Evaluation on DDSM Mammograms

This section presents the results obtained from the segmentation of 200 mam-mograms from the DDSM. The aim of the experiment is to identify the optimal contrast enhancement expert for each of the 200 abnormal mammograms. Each mammogram image has been grouped according to its target breast type. There are 50 images per breast type grouping and results will be presented on a per breast type basis. Each mammogram is contrast enhanced using each enhancement method identified in section 11.3.2.1.1 and segmented using the unsupervised HMRFU segmentation method. The sensitivity in the detection of breast lesions following segmentation of the enhanced images is quantified using the outcomes given Table 11.2 and the ground truth definition.

From a set of M enhancement methods (E\,..., EM) for a given mammogram, the target contrast enhancement, Em where m e{1,..., M} is the enhancement method giving the largest value of (TPr + SUBTPr) following segmentation using HMRFU. The target contrast enhancement expert Em is identified as assign Em ^ m if (TPT + SUBTP' )m = arg max (TP' + SUBTP' )m

Table 11.3: Segmentation results from using the original mammogram

Type

TPr

SUBTPr

Total

1

0.25

0.04

0.29

2

0.18

0.11

0.29

3

0.20

0.04

0.23

4

0.15

0.02

0.16

Mean

0.20

0.05

0.24

Values given are mean percentage of mass detected with TP and SUBTP outcome together with their sum.

Values given are mean percentage of mass detected with TP and SUBTP outcome together with their sum.

The target contrast enhancement Em is found for every mammogram from all M enhancement methods, m e{1,..., M), keeping the segmentation method and associated initialization parameters constant. Having identified each of the target enhancement experts, the following important observations can be made (see sections 11.3.3.2.1-11.3.2.2.6).

11.3.2.2.1 Original vs. Contrast-Enhanced Mammograms. Each original unenhanced image is segmented and out of the 200 mammograms used, 80 (40%) images give no sensitivity in the detection of target regions, that is, for these images the value of (TPr + SUBTPr)< 0. Of these 80 unenhanced images, 34 still give no sensitivity in detection after application of each evaluated contrast enhancement method. Only the remaining 166 abnormal mammograms are considered in the evaluation of the optimal strategies described in the following sections. Table 11.3 presents the results from segmenting the original unenhanced mammograms grouped by breast type. The table shows the mean percentage of target region detected with each outcome, grouped by breast type. We observe that the segmentation performance decreases as the breast density increases.

11.3.2.2.2 Improved Segmentation in Contrast-Enhanced Mammograms. Of the 200 mammograms used in this evaluation, 150 (75%) give a greater sensitivity after the target, contrast enhancement method, compared with the sensitivity obtained from the original images.

25 20

15 10

Figure 11.5: Frequency of each contrast enhancement method being selected as the target method for each mammogram grouped by breast type.

11.3.2.2.3 Reduced Sensitivity in Enhanced Images. Of the 166 enhanced mammograms that give a positive sensitivity result, 12 (7%) reported an inferior (reduced) sensitivity on application of the target contrast enhancement method compared with the original mammogram. No attempt is made to learn that "no contrast enhancement" is best suited for these images because of the small sample number.

11.3.2.2.4 Frequency of Optimal Contrast Enhancement Methods.

The target contrast enhancement for a given mammogram is defined above using Eq. (11.10). Figure 11.5 presents the frequency that each ofthe six enhancement method are identified as the target contrast enhancement for a mammogram, grouped according to its breast type. From this figure, it can be seen that each enhancement methods is a target optimum for at least one mammogram. The variability in choice of target method is greater in the fatty breast (breast types 1 and 2) cases compared to the dense types (types 3 and 4). Note that the proposed novel extension to the ACE combining fractal dimension, method ACELFD, outperforms the classic ACE method more frequently for the fatty breast types. Additionally, it should be noted that the HISTOEQ method outperforms all other

25 20

15 10

Figure 11.5: Frequency of each contrast enhancement method being selected as the target method for each mammogram grouped by breast type.

Table 11.4: Mean percentage improvement in segmentation performance by using the target-enhancement method compared to the segmentation of the unenhanced image for each mammogram grouped by breast type

Type

TPT

SUBTPT

Total

1

0.76

1.25

0.83

2

1.28

0.64

1.03

3

0.80

4.00

1.43

4

1.47

6.50

2.25

Mean

1.07

3.09

1.38

methods for fatty breasts but is noticeably less effective in the segmentation of dense breasts (types 3 and 4).

11.3.2.2.5 Mean Percentage of Target Mass Detected as TPt or SUBTPT. Using the target contrast enhancement expert, Table 11.4 tabulates the mean percentage improvement in segmentation performance compared with the segmentation of the unenhanced image for lesion ground-truth detected with outcomes TPt and SUBTPt for all mammograms, grouped by breast type. The greatest improvement can be seen in the segmentation of the densest breast type using the target contrast enhancement method, compared to the segmentation obtained from the unenhanced original (e.g., on average, for breast type 1, using target enhancement method result in 83% improvement in segmentation compared to the unenhanced original).

11.3.2.2.6 A Classical Solution in Choosing a Single Optimal Contrast Enhancement Method for CAD. In analyzing the 166 selected mammo-grams, a common approach to identify a single optimal contrast enhancement for use in CAD is to evaluate a selection of enhancement methods on a range of different training mammograms from different breast types. By determining the mean value of TPt + SUBTPt for all enhancement methods across all mammograms, the CAD researcher can choose to select the single method maximizing the value of TPT + SUBTPT from the training set. Table 11.5 lists the

Table 11.5: Mean percentage improvement in segmentation performance for each contrast enhancement method compared with the segmentation of the enhanced image, grouped by breast type

Type TPr SUBTPr Total

Mean

Mean

Mean

Mean

Mean

Mean

percentage improvement in segmentation compared with the original segmentation, in the detection of ground truth defined lesions detected with outcomes TPt and SUBTPt obtained using each individual enhancement method. From the data obtained, in identifying the single best enhancement method, the FUZZY method (Table 11.5, part c) is chosen for each breast type. Notice that the segmentation performance obtained using the FUZZY method shown in Table 11.5 (part c) for all mammograms, is suboptimal compared with that of the target contrast enhancement for each mammogram shown in Table 11.4.

0 0

Post a comment