Coal workers' pneumoconiosis (CWP), also known as black lung disease, exhibits rounded or irregular small opacities in standard chest x-rays caused by a physiological reaction to the deposition of coal dust in the lung. The disease is staged with two parameters: profusion, a measure of opacity concentration, and opacity size/shape. When diagnosing an x-ray, there is a great deal of variability in among radiologists in staging the disease. Consistent diagnoses are vital, however, because of medical and legal issues such as prevention of disease progression by early detection, and equitable compensation for afflicted workers. The goal of this research is to develop a method to minimize diagnostic variability by automatically staging CWP utilizing computer analysis of digitized chest x-rays. This work is a cooperative effort with the National Institute of Occupational Safety and Health (NIOSH, Morgantown, WV) which maintains a repository of all coal workers' x-rays and corresponding diagnostic and epidemiological data for the United States. No deterministic methodology exists for solving problems of pattern recognition such as the automated staging of CWP. We plan to apply statistical pattern recognition techniques utilizing new features derived from mathematical modeling of the digitized image. Pairwise linear discriminant analysis will be employed in dimensionality reduction and the development of a classifier based on an optimal subset of features. We are currently investigating features based on the spatial gray level dependence matrix, two-dimensional Fourier spectra, image morphology, and Gauss-Markov random fields. Our analysis will examine the dependence on context (spatial location) of features in classifier development, which has not been applied to this problem previously. Training and test set films will be obtained from the standard International Labour Organization set of pneumoconiosis films and from NIOSH films, digitized at the University of Pittsburgh. Diagnostic data on the films will be obtained from previous readers (usually two) and the University of Pittsburgh Department of Radiology. From each x-ray, more than 100 tiles are selected from the inter-rib spaces in the lung fields. Image noise from rib structures is therefore eliminated. Mathematical modeling of opacities will be employed to explore the appropriateness of the features to be used, as well as their robustness to variations in exposure and digitization noise, which is unique to our work. Our utilization of an opacity model in the prediction of image covariance corresponded to results obtained from actual data. This type of analysis will give much more insight into both the value and meaning of each feature.
