Comparison of Combination Strategies Ensemble Combination Rules vs AWM

In order to develop a number of experts that can be combined, we extract different gray-scale and texture data per pixel in the images. The gray-scale values of the pixels are intensity values, and texture features are extracted from pixel neighborhood. The following table shows the different feature experts used in our analysis based on different features. Each expert can be implemented with one of the four segmentation models described earlier.

Expert

Description of pixel feature space

Dimensionality

gray

Original gray scale

1

enh

Contrast enhanced gray scales

1

dwt1

Wavelet coefficients from {DLh, D1hh, D1hl, S[l}

4

dwt2

Wavelet coefficients from {DLh, D2hh, D2hl, SLl}

4

dwt3

Wavelet coefficients from {D3LH, D3HH, D3HL, SLL}

4

laws1

Laws coefficients from E5 impulse response matrix

5

laws2

Laws coefficients from L 5 impulse response matrix

5

laws3

Laws coefficients from R5 impulse response matrix

5

laws4

Laws coefficients from W5 impulse response matrix

5

laws5

Laws coefficients from S5 impulse response matrix

5

We now present the results on 200 test mammograms that contain lesions. The details of training and testing scheme are the same as detailed in section 11.4.2. As we mentioned earlier, each breast is classified as one of the four types (1, predominantly dense; 2, fat with fibroglandular tissue; 3, hetero-geneously dense; and 4, extremely dense) and the results are presented for data from each type. Table 11.11 shows the test results on sensitivity of the

Table 11.11: Mean sensitivity for each testing strategy for DDSM image database

Breast type 1

Breast type 2

Breast type 3

Breast type 4

wgmmS

laws1

laws4

laws4

laws4

0.740

0.545

0.675

0.510

wgmmMRF

laws1

laws1

enh

laws1

0.690

0.650

0.650

0.640

wgmmu

enh

laws2

enh

laws1

0.525

0.575

0.660

0.550

wgmmMrf

laws1

laws1

laws4

laws1

0.690

0.640

0.690

0.540

Results are shown for all breast types. Winning segmentation expert are shown in bold per breast type.

Results are shown for all breast types. Winning segmentation expert are shown in bold per breast type.

different segmentation models with different features without expert combination. The following key conclusion can be drawn from these results: (a) A single feature is not always the winning feature. In general, features enh, laws\, and laws4 do quite well. (b) It is easier to segment fatty breasts as opposed to dense breasts which is to be expected. (c) Models using MRF work better than those that do not use them. (d) There is no clear cut winner between supervised and unsupervised strategy; depending on which features they use, they can outperform the other. (e) For three of the breast types 1, 2, and 4, the model WGMMmrf is a clear winner, whereas for breast type 3, WGMMmrf performs the best.

Table 11.12: Mean sensitivity for each combination strategy for DDSM database

Breast type 1

Breast type 2

Breast type 3

Breast type 4

WGMMS

Mv

AWM

AWM

Min

0.510

0.520

0.701

0.505

wgmmMrf

AWM

Sum

AWM

AWM

0.575

0.630

0.727

0.680

WGMMU

AWM

Mv

Mv

Mv

0.320

0.532

0.515

0.525

wgmmMrf

AWM

AWM

AWM

AWM

0.550

0.667

0.705

0.625

Results are shown for all breast types. Winning combination method shown in bold per breast type.

Results are shown for all breast types. Winning combination method shown in bold per breast type.

Table 11.13: The results from best performing (a) expert strategy and (b) AWM combination strategy.

T

Seg

Expert

Sens

wgmmMrf wgmmMrf

0.650 0.690 0.640

.15 .23 .31 .28

T

Seg

Cmb

Sens

4

wgmmMrf wgmmMrf wgmmMrf wgmmMrf

AWM AWM AWM AWM

0.575 0.667 0.727 0.680

.25 .26 .38 .37

Winning strategy shown in bold. T = breast type; Seg = segmentation strategy; Cmb = combination strategy; Sens = sensitivity; % mass = mean percentage of target lesion detected as true positive.

Winning strategy shown in bold. T = breast type; Seg = segmentation strategy; Cmb = combination strategy; Sens = sensitivity; % mass = mean percentage of target lesion detected as true positive.

We next compare the ensemble combination rules with the AWM expert combination strategy on the four breast type data testing. The results are shown in Table 11.12. The key results can be summarized as follows: (a) The AWM method result always turns out to be the overall best result compared to all ensemble combination rules on all breast types. (b) The AWM results are best with the WGMM'Srf segmentation method on breast types 1, 3, and 4, and best with WGMM^F on breast type 2. (c) The combination methods Max and Prod never win. (d) Segmentation models using MRF are better than those that do not use them.

In Table 11.13 we compare single best experts with the best combination of experts for the four breast types. The results show that only on breast type 1, using the single best expert WGMMS with laws\, features will outperform all other experts and combination of experts (sensitivity of 0.74). For the remaining three breast types, the AWM expert combination method is the best. For breast types 3 and 4 (dense breasts), the supervised learning based models with MRF are better, whereas for fatty breast of type 2, unsupervised learning model with MRF is the best.

SEGMENTED IMAGES

SEGMENTED IMAGES

Figure 11.7: Schematic overview of false-positive reduction strategy within the adaptive knowledge-based model.
0 0

Post a comment