Some theoretical considerations concerning texture analysis of radiographic images

A significant amount of the diagnostic information contained in radiographic images is of a textural nature. This is indicated by the presence of the adjectives "fluffy, honeycombed, spotted, etc." in the radiologist descriptive vocabulary. Unfortunately, the textural patterns representing disease entities On radiographs are fairly complex. If one couples this fact with the obvious need for correct classification of disease entities, then the problem of choosing the "best" texture algorithm becomes one of concern. The purpose of this paper is to present a theoretical comparison of the "innate" abilities of four texture analysis algorithms to do automated texture discrimination. The algorithms to be examined are the Spatial Gray Level Dependence Method, the Gray Level Run Length Method, the Power Spectral Method, and the Gray Level Difference Method.