Automatic extraction and categorization of lung abnormalities from HRCT data in MDR/XDR TB patients

Abstract

An ancient disease, tuberculosis (TB) remains one of the major causes of disability and death worldwide. In 2006, 9.2 million new cases of TB emerged and killed 1.7 million people. We report on the development of tools to help in the detection of lesions and nodules from High Resolution Computed Tomography (HRCT) scans and changes in total lesion volumes across a study. These automated tools are designed to assist radiologists, clinicians and scientists assess patients' responses to therapies during clinical studies. The tools are centered upon a rule-based system that initially segments the lung from HRCT scans and then categorizes the different components of the lung as normal or abnormal. A layered segmentation process, utilizing a combination of adaptive thresholding, three-dimensional region growing and component labeling is used to successively peel off outside entities, isolating lung and trachea voxels. Locating the Carina allows logical labeling of the trachea and left/right lungs. Shape and texture analysis are used to validate and label normal vascular tree voxels. Remaining abnormal voxels are clustered on density, gradient and texture-based criteria. Several practical problems that arise due to large changes in lung morphology due to TB and patients' inability to hold their breath during scan operations need to be addressed to provide a viable computational solution. Comparisons of total common volumes of lesions by size for a given patient across multiple visits are in concordance with expert radiologist's manual measurements. © 2012 Springer-Verlag GmbH.