Extracted Features

The result of the Fuzzy C-Means algorithm is the division of the breast into (only) two clusters. Subsequently, a set of features for both classes can be directly extracted from the original images (no preprocessing/filtering was applied). Here we used a set of morphological and texture features.

As morphological features, the relative area and the four first histogram moments for both clusters were calculated. Note that the four moments of the histogram are related to the mean intensity, the standard deviation, the skewness and the kurtosis of each cluster. On the other hand, a set of features derived from co-occurrence matrices [64] were used as texture features. Here we use four different directions: $0^\circ$ , $45^\circ$ , $90^\circ$ , and $135^\circ$ , and three distances equal to $1$ , $5$ , and $9$ pixels. Note that these values were empirically determined and are related to the scale of textural features found in mammographic images. Co-occurrence matrices are not generally used as features, rather a large number of textural features derived from matrices have been proposed [64]. For each co-occurrence matrix the following statistics were used: contrast, energy, entropy, correlation, sum average, sum entropy, difference average, difference entropy, and homogeneity features.

As each of these features were extracted from each class, we deal with $226$ features in total, $10$ from morphological characteristics and $216$ from textural information.