Characteristics of Density Distribution of Chest Roentgenograms and Automatic Recognition of Rib Boundaries

Abstract

Computer diagnosis of the roentgenogram is regarded as a typical two-dimensional pattern recognition problem. As the first step to approach the problem, we investigated the characteristics of the density distribution of the chest roentgenograms by comparing two samples, one is of normal lung and the other of spontaneous pneumothorax. The main results are summarized as follows : (1) The average density varies widely even in a roentgenogram depending on the shape of the lung. (2) Remarkable random components are observed, which are caused by irregular lung markings, photographic granularity and electrical noise of recording device. (3) Highly deterministic patterns (e. g., ribs and vessels) which are not the eventual recognition objects exist. (4) There exist overlaps of various kinds of images including the above three. The density at any point on the film is approximately equal to the superposition of density of each component image.
For realization of automatic processing of such a complex pattern as the roentgenogram, it should be stressed that each of the basic operations suits processing of each corresponding component mentioned above, and their successful consolidation is indispensable.
As one of such basic operations, we studied a detecting method of edge lines of ribs using a simple model hypothesized on the basis of the above results. In the model, the patterns (sets of discrete sample points) consist of regions of two different uniform densities with additive Gaussian noise. The procedure is decided into three steps : (I) The state of each sample point “on edge line” or “not” is estimated by the likelihood ratio estimated from the state of its neighborhood. (II) Two sample points decided as “on edge line” are connected if they are adjacent, thus giving connected graphs. (III) In each connected graph a line being most likely “on edge line” is determined as the recognition result.
Finally results of recognition experiment using a cutout of the middle lobe region including two ribs and tuberculosis region are shown, indicating effectivenss of the method proposed here.