Evaluating five features descriptors in classification of mammography images by artificial neural network

Abstract

Purpose: Comparison of five features descriptors in terms of representation of tissues in mammographies. Method: Images had features extracted for producing five features datasets used for training an Artificial Neural Network (ANN), all the feature descriptors were submitted to the very same ANN configuration. The interest is to rank the features descriptor according to ANN’s performance in classification of tissues. Results: The best descriptor is Pyramid of Histogram of visual Words (PHOW), the second group composed by Pyramid of Histogram of Colors (PHOC), Pyramid of Wavelets (PWAV) and Pyramid of Histograms of Gradients (PHOG), at third place there is Pyramid of Gabor (PGABOR).  Conclusion: PHOW presents the best performance. Nevertheless, an application of PHOW in Computer Aided Diagnosis would need be funded in a very representative “visual vocabulary”, based on a very large mammography database. Although PHOC presents a very simple approach, surprisingly, it takes the second-best performance.