Exploring Different Normalization and Classification Approaches for Mammography Analysis with CNNs

In order to improve the performance of Convolutional Neural Networks (CNN) in the classification of mammographic images, many researchers choose to apply a normalization method during the pre-processing stage. In this work, we aimed to assess the impact of 6 different normalization methods in the classification performance of 2 CNNs. We have also explored 5 classifiers, being the first one the CNN itself. The other 4 correspond to Support Vector Machine (SVM), Random Forest (RF), Simple Logistic (SL) and Voted Perceptron (VP) classifiers, all of them fed with features extracted from one of the layers - comprised between the sixteenth and the nineteenth - of the CNN. The last 3 classifiers were tested with different options for data testing presentation, according to the Weka software: Supplied Test Set (STS), 10-fold Cross Validation (10-FCV) and Percentage Split (PS). Results indicate that the effect of image normalization in the performance of the CNNs depends on which network is chosen to make the classification; besides, the normalization method that seems to have the most positive impact is the one that subtracts to each image the corresponding image mean and divide it by the standard deviation (best AUC mean values were 0.786 for CNN-F and 0.790 for Caffe; the best run AUC values were, respectively, 0.793 and 0.791. Layer 1 freezing decreased the running time and did not harm the classification performance. Regarding the different classifiers, CNNs used alone with softmax yielded the best results, with the exception of the RF and SL classifiers, both using the 10-FCV and PS options; however, with these options, we cannot guarantee that the test set images are presented for the first time to the network.