Tissue characterization from ultrasound B-scan data.
An approach to ultrasonic tissue characterization, using textural features of the B-scan image, is described. Portions of a B-scan image, 64 X 64 pixels spatially by 8 bits deep, are acquired from regions of interest and subjected to computer analysis. A systematic approach to defining a set of 93 textural features of a B-scan is described and methods and criteria for selecting optimum combinations of these are discussed. As a test of its power, the approach has been applied to the discrimination between the B-scan textures corresponding to livers and spleens of normal humans and various measures of "success" have been quantified both on a "training set only" and on a "training set plus test set" basis. The overall test probability of success of 82% on a single image and 94% on a subject yielding multiple images indicates the potential of the techniques for conditions where a subtle but uniform change in parenchymal texture may be present.