Parameter optimization of a computer-aided diagnosis scheme for the segmentation of microcalcification clusters in...

Our purpose in this study is to develop a parameter optimization technique for the segmentation of suspicious microcalcification clusters in digitized mammograms. In previous work, a computer-aided diagnosis (CAD) scheme was developed that used local histogram analysis of overlapping subimages and a fuzzy rule-based classifier to segment individual microcalcifications, and clustering analysis for reducing the number of false positive clusters. The performance of this previous CAD scheme depended on a large number of parameters such as the intervals used to calculate fuzzy membership values and on the combination of membership values used by each decision rule. These parameters were optimized empirically based on the performance of the algorithm on the training set. In order to overcome the limitations of manual training and rule generation, the segmentation algorithm was modified in order to incorporate automatic parameter optimization. For the segmentation of individual microcalcifications, the new algorithm used a neural network with fuzzy-scaled inputs. The fuzzy-scaled inputs were created by processing the histogram features with a family of membership functions, the parameters of which were automatically extracted from the distribution of the feature values. The neural network was trained to classify feature vectors as either positive or negative. Individual microcalcifications were segmented from positive subimages. After clustering, another neural network was trained to eliminate false positive clusters. A database of 98 images provided training and testing sets to optimize the parameters and evaluate the CAD scheme, respectively. The performance of the algorithm was evaluated with a FROC analysis. At a sensitivity rate of 93.2%, there was an average of 0.8 false positive clusters per image. The results are very comparable with those taken using our previously published rule-based method. However, the new algorithm is more suited to generalize its performance on a larger population, depends on two monotonic outputs making its evaluation much easier and can be trained in an automatic way making practical its application on a large database.