Dataset and Adaptive Knowledge Based Model Configuration

This section presents the results from evaluating the optimal contrast enhancement and segmentation knowledge-based components of the adaptive knowledge-based model on a dataset of 200 DDSM mammograms containing abnormalities. The 200 mammograms comprise 50 images from each of four different breast types. To obtain a testing result for each mammogram, knowledge-based components utilize separate training and testing folds such that no image from a test fold exists in a corresponding training fold. Training data for the abnormal mammograms is based on redefined DDSM ground truth boundaries.

Figure 11.8 shows the configuration of the adaptive knowledge-based model for contrast enhancement and mammogram segmentation used for performance evaluation. Enhancement and segmentation experts are identified in the black boxes. Knowledge-based components, providing optimal enhancement and optimal segmentation, are identified in dotted boxes. Associated with each expert and knowledge-based component in Fig. 11.8 is a table with four rows, one for each breast type. The right-hand column of the table identifies the performance of the associated expert or knowledge-based component for all mammograms of the predicted breast type. This performance measure is computed differently for contrast enhancement and segmentation components as follows:

(a) Enhancement component: Performance is measured by the mean percentage improvement in the segmentation of the contrast-enhanced image compared to that of the unenhanced original for all mammograms of a given breast type.

(b) Segmentation component: Performance is measured by the mean area (AZ) under the ROC curve, for all mammograms of a given breast type. Use of this measure in evaluating the adaptive knowledge-based model reflects the underlying sensitivity and false-positive count across all ROC thresholds and has been used in other studies [41] to compare classification tasks.

Ttlr

OPTIMAL ENHANCE

ENHANCED I

OPTIMAL SEGMENT

Mean 0.72

Figure 11.8: Evaluation of a given configuration of the adaptive knowledge-based model. Performance shown for each breast type for each component is interpreted as a percentage.

The following paragraphs briefly review the contrast enhancement and segmentation of digitized mammograms described in previous sections.

Contrast enhancement: The trained contrast enhancement knowledge-based component selects the optimal contrast enhancement method for a test mammogram, as one from a subset of six selected enhancement methods. Each of the enhancement methods has been described in section 11.3.2.1.1. The BPM strategy is used to implement the knowledge-based contrast enhancement component, and following training predicts the optimal enhancement method for a testing mammogram on the basis of an extracted feature vector. A different feature vector is used depending on the predicted breast type. A feature vector comprising a selected number of principal components FBP26 is used for mam-mograms of breast types 1-3. For breast type 4, the complete feature vector FBP316 is used.

Segmentation: To segment a mammogram, the semisupervised WGMM constrained with a WGMMMMRF strategy is used. Ten different segmentation experts are trained and each one gives a segmentation decision for the test mammogram. The 10 experts have been trained to operate on specific groupings of input feature spaces. The experts for this configuration of the adaptive knowledge-based model are described in section 11.4.4. The decision of each expert is combined using a knowledge-based segmentation component implemented using the AWM described earlier. The AWM will predict the optimal blend of expert decisions to maximize the segmentation performance.

11.6.1.2 Results

This section presents the results from contrast enhancement and segmentation using 200 abnormal images, such that the image processing pipeline is constructed on the basis of the predicted breast type.

Knowledge-based contrast enhancement: From the results presented in Fig. 11.8, it can be seen the best performing expert is the FUZZY contrast enhancement method over all breast types. The average improvement in segmentation performance is 54% for all 200 abnormal images. Using the predicted optimal enhancement method from the knowledge-based contrast enhancement component, the average improvement in segmentation performance increases to 60%. The knowledge-based contrast enhancement component is determining the optimal enhancement based on component knowledge learnt during supervised training. By utilizing this hidden knowledge, the resultant performance is improved compared with that obtained by simply using the single best contrast enhancement method, the FUZZY contrast enhancement method.

Knowledge-based mammogram segmentation: From the results presented in Fig. 11.8, it can be seen that the single best performing expert is the grayscale contrast enhancement method for all predicted breast types. The mean AZ value is 0.65 for all 200 abnormal images. Using predicted optimal enhancement method from the knowledge-based segmentation component, the mean AZ value rises to 0.72. Clearly in combining the decisions of each expert, the knowledge-based segmentation component is performing better than that obtained by selecting the single best performing expert. The outcome of each segmentation expert is considered when forming the optimal segmentation. The statistically motivated AWM component, determines optimal weights for each segmentation expert that are most likely to have given rise to a resultant combined single segmentation. By doing this, the resulting performance is improved over all other segmentation experts.

0 0

Post a comment