Hierarchical Discovery of Patterns of Infections in Chest Radiographs Using Non-Negative Matrix Factorization

Chest radiography is a screening tool for the detection and the primary assessment of abnormalities consistent with the infections of the lower respiratory tract. Most commonly, the radiographic manifestations of bacterial infection are areas of consolidation with a segmental or lobar distribution which appear as areas of increased pulmonary opacity. This paper addresses the problem of the discovery of consolidation patterns in chest radiographs, by a methodology for automatic segregation of lung consolidations from normal lung parenchyma. The proposed methodology is based on non-negative matrix factorization (NMF) of pulmonary radiographic patterns represented by intensity histograms and Gabor textural features. NMF is considered as a soft clustering algorithm which is iteratively applied according to an hierarchical cluster-merging scheme. This scheme reckons with the resulting NMF bases so as to overcome the limitations associated with the geometry of the clusters. The experimentation results validate the effectiveness of the proposed methodology and demonstrate its comparative advantage over the conventional hierarchical and NMF partitional clustering algorithms.