Segmentation of ultrasonic ovarian images by texture features

Abstract

Auto-segmenting two-dimensional images of the ovary into non-ovarian, normal ovarian, and abnormal ovarian regions is required when using ultrasonic image to detect ovarian cancer. The texture-based segmentation method presented here is a pixel classifier based on four texture energy measures associated with each pixel in the images. The 25 two-dimensional feature masks are derived from 3 basic one-dimensional vectors to evaluate the classification results. Four of those features are selected as the bases for the automated clustering procedure. The segmented images produced as the result of applying the algorithm to an example image are presented and discussed. The automated clustering algorithm with these texture-feature masks has been found to hold promise as an automated segmentation method for ultrasonic ovarian images.