Non-negative Dimensionality Reduction for Mammogram Classification

Directly classifying high dimensional datamay exhibit the ``curse of dimensionality'' issue thatwould negatively influence the classificationperformance with an increase in the computationalload, depending also on the classifier structure. Whenworking with classifiers not affected by this issue (suchas Support Vector Machines, for instance), thecomputational load still exists due to the required timein computing the kernel matrix. Moreover, theperformance is affected when a few samples per classis available for the training procedure. One commonsolution is to carry out a feature extraction step forreducing the data dimension prior to classification.The paper describes the application of NonnegativeMatrix Factorization (NMF) for extracting featuresfrom mammogram medical images with differentresolution, further used for recognizing breast tumors.For comparison, Principal Component Analysis (PCA)and Independent Component Analysis (ICA) wereexplored. Experiments show that NMF methodoutperforms PCA and ICA, leading to higherclassification accuracy.