Preprocessing for classification of thermograms in breast cancer detection

Performance of binary classification of breast cancer suffers from high imbalance between classes. In this article we present the preprocessing module designed to negate the discrepancy in training examples. Preprocessing module is based on standardization, Synthetic Minority Oversampling Technique and undersampling. We show how each algorithm influences classification accuracy. Results indicate that described module improves overall Area Under Curve up to 10% on the tested dataset. Furthermore we propose other methods of dealing with imbalanced datasets in breast cancer classification.