Reduction of False Positives

False positives are initially reduced by removing regions with an area less than a predefined threshold Tarea. We choose Tarea = 122 pixels, thus any region less than 5 mm in diameter is removed. This approach is used here. From the remaining suspicious regions, features are extracted, and using a trained ANN classifier, a region is labelled as abnormal or normal.

Feature extraction: For those regions that remain following the application of the area test, the 316-dimensional feature vector described in [42] is extracted using the pixels comprising the region. To improve classifier generalization [32] unbiased PCA is used to map the 316-dimensional feature vector into a lower dimensional feature space. PCA is used on a per breast type basis, so that the number of principal components is selected independently for each breast type. Using this approach, for each predicted breast type, the number of principal components selected are as follows: (type 1, 37 components; type 2, 33 components; type 3, 35 components; type 4, 41 components). From this table it can be seen that the highest dimensional feature space results from the densest breast types (type 4), which are the generally the hardest to interpret by an expert radiologist [37].

Model order selection: In order to maximize the performance of each ANN for each predicted breast type, model order selection of the ANN classifier is performed. By varying the number of hidden nodes and performing a classification on all suspicious region, ROC can be performed and the area under the ROC curve (AZ) computed. The optimal number of hidden nodes is determined as that maximising the AZ value.

0 0

Post a comment