Matching input variables sets and feedforward neural network architectures in automatic classification of microcalcifications and microcalcification clusters

Classifying mammographic findings, microcalcifications and clusters of microcalcifications, as either benign or malignant, is a difficult task. This is mainly due to the variability of their appearance. Appropriate feature selection is probably the most critical step of an automatic classification process. This paper aimed to identify a set of features that allows for making the best automatic classification. Groups with different numbers of features were generated using the Scalar Feature Selection - SFS. Fisher´s Discriminant Ratio - FDR and the area under Receiver Operating Curve - ROC were used as auxiliary distance measurements. For classification purposes, different architectures of feedforward neural networks were employed. An attempt was made to establish the best match between a group of features and a neural network architecture. The results are evaluated through the cross validation method using measurements of accuracy, sensitivity and specificity.