Patient specific normalization of chest radiographs and hierarchical classification of bacterial infection patterns

The analysis of multiple chest radiographs is prone to errors mainly due to different acquisition settings used in clinical routine and due to patient specific characteristics affecting the distribution of the image intensities. The purpose of this paper is twofold: (a) it proposes a new image normalization technique based on samples extracted from the spinal cord/mediastinum region, in order to reduce the error due to the different settings and the patients' characteristics; (b) it proposes a supervised hierarchical classifier combination scheme fusing knowledge extracted from intensity histograms and Gabor textural features so as to cope with the detection of lung consolidations, and the assessment of their extent. The presence of such consolidations in a chest radiograph is usually an indication of a bacterial pulmonary infection. The experimentation results validate the advantage of the proposed normalization technique over the conventional ones and the effectiveness of the hierarchical classification of the normalized patterns.