Computer-assisted diagnosis of CT pulmonary images

Purpose: to produce a robust algorithm as the basis of a computer program for diagnosing pulmonary images from computed tomography (CT) scans. Materials and Methods: single slices are extracted from complete CT data-sets and normalized for radio-density / greyscale relationship. Normalization is discussed in a companion paper; it is an essential image pre-processing procedure. Modal greyscale values in regions of the parenchyma provide indices of physiological or pathological conditions. The application of thresholds focuses attention around three characteristic values that correspond to below-normal, normal and above-normal tissue densities for given locations. Plots of the thresholded pixels form patterns that map the corresponding range of densities; their number and distribution characteristics assessed using information dimensions providing the diagnostic criteria. Parameters are empirically derived by analysis of a variety of data-sets. Results: an algorithm using pixel counts and information dimensions of patterns derived from standard greyscale thresholds gives consistent and reliable diagnoses for all slices within the training sets: 97% accuracy, 100% sensitivity and 96% specificity. Conclusion: a promising approach to the task of formulating computer-based automatic, or semi-automatic diagnosis of the lungs, uses algorithms based on quantifying the pixel plots that result from appropriate greyscale thresholding. The method may prove applicable to other organs that have a fractal-like structure.