Key Observations

This section has presented a configuration of the adaptive knowledge-based model. The performance of the model has been evaluated on a dataset of 200 abnormal mammograms from four different breast types. The aim of the evaluation has been to demonstrate the utility of the model compared with that of individual experts. Following this, the performance of a specific configuration of the adaptive knowledge-based model has been evaluated on a dataset of 400 mammograms, comprising 200 abnormal and 200 normal images. From these evaluations, the following key observations can be made:

1. Utility of knowledge-based contrast enhancement component: Using a dataset of 200 abnormal mammograms, the utility of the knowledge-based contrast enhancement expert has been demonstrated to be greater than that of the best performing expert contrast enhancement method. Using the predicted optimal contrast enhancement method in image segmentation results in a 60% improvement in the detection of abnormal regions over the original segmentation. This is compared to a 54% improvement from the single best performing expert, the FUZZY contrast enhancement method.

2. Utility of knowledge-based segmentation component: By optimally combining the segmentation outcomes of 10 different segmentation experts, each operating on a unique feature space partition, the knowledge based segmentation component resulted in a mean ROC AZ value of 0.72 for 200 mammograms from four breast types. This is compared to the best performing gray-scale segmentation expert reporting a mean AZ value of 0.65.

3. Utility of adaptive knowledge-based model in presence of normal mammograms: Evaluation of the performance of this configuration of the adaptive knowledge-based model on a dataset of 400 mammograms comprising 200 abnormal and 200 normal images results in a segmentation sensitivity of 0.84 for the detection of breast lesion with 167.01 false-positive regions per image. This demonstrates a high level of sensitivity in the presence of a complete spectrum of mammogram types.

4. False-positive reduction: The results following region prefiltering in the false-positive reduction methodology demonstrate the utility of the region size thresholding strategy. Following classification by each trained optimized ANN results in a sensitivity of 0.75 with 6.46 false-positive regions per image.

0 0

Post a comment