The purpose of texture characterization is to produce a set of measurements which identify different types of textures. It is desirable to use measurements which produce unique characteristics for each type of texture. This allows a texture class to be described by a set of parameters or features which can be used by a segmentation algorithm to partition an image into homogeneous texture regions. Since texture features ultimately find their use in segmentation algorithms, it is more appropriate to consider those features which can effectively characterize the textures in the images being segmented. In many applications, particularly medical imaging, the types of textures are known a priori. It is unnecessary to make invalid assumptions about the types of textures which the features can effectively detect. In this paper, a new framework is presented to evaluate a given set of texture features, based on the set of textures they are to characterize. The textures are sampled from “real” images which are to be segmented, rather than the commonly used Brodatz textures. Using robust statistics, the number of outliers produced by each feature is used to determine its effectiveness in characterizing the set of textures. The proposed framework is applied to a set of textures sampled from a sequence of diagnostic ultrasound images of an ovary in vitro. Thirty-two spatial and frequency domain texture features are evaluated. Numerous experiments are performed to evaluate the ability of the texture characterization framework to characterize the textures present in ultrasound images. It is found that the texture classes present in the ultrasound images could be uniquely characterized at all resolutions, independent of orientation. The experimental results suggest that the proposed framework, together with the use of robust statistics, can provide a ranking on the effectiveness of the features and thus can determine better features for a latter segmentation process.