Texture analysis of ultrasonic images of the prostate by means of co-occurrence matrices

Abstract

As speckle on ultrasonic B-scan may reveal information relative to tissue structure, the present work attempts to discriminate the various prostatic tissues (normal tissue, benign prostatic hypertrophy and cancer) by means of texture analysis. We select regions of interest by their homogeneous appearance. A pre-processing stage is required to obtain stationary samples. The method used measures the second-order statistics, namely co-occurrence matrices. Fairly good tissue signatures have been obtained with parameters derived from these matrices. Of 37 images processed, 78 percent of the samples were classified with success, which is a high score considering that the images cannot be discriminated visually. However, while such results are obtained when wide regions of interest are investigated (64 x 64 pixels), they are not as good with smaller sample sizes, i.e., when the pathological area is very small.