Information theory optimization based feature selection in breast mammography lesion classification
Johanna Uthoff, Jessica C. Sieren · 2018
Quantitative imaging features of intensity, texture, and shape were extracted from breast lesions and surrounding tissue in 287 mammograms (150 malignant, 137 benign). A feature set reduction method to remove highly intra-correlated features was devised using k-medoids clustering and k-fold cross validation. A novel feature selection method using information theory was introduced which builds a feature set for classification by determining a group of class-informative features with low set co-information. An artificial neural network was built from the selected feature set using 10-hidden layer nodes and the tanh activation function. The resulting computer-aided diagnosis tool achieved a training accuracy of 96.2%, sensitivity of 97.6%, specificity of 95.2%, and area-under-the-curve of 0.971 along with 97.1% sensitivity and 94.9% specificity a blinded validation set.