Context-sensitive model learning for lung nodule detection

Nodule detection in chest radiographs is a main component of current Computer Aided Diagnosis (CAD) systems. The problem is usually approached as a supervised classification task of candidate nodule segments. To this end, a discriminative model is learnt from predefined set of features. A key concern with this approach is the fact that some normal tissues are also imaged and these regions can overlap with the lung tissue as to hide the nodules. These overlaps may reduce the discriminative ability of extracted features and increase the number of false positives accordingly. In this study, we offer to learn distinct models for bone and normal tissue regions following to the segmentation of ribs, which are often the major reason for false positives. Thus, the nodule candidates in bone and normal tissue regions can be assessed in context-sensitive way. The experiments on a common benchmark set determine that the proposed approach can significantly recue the false positives while preserving the sensitivity of detections.