Random field models in the textural analysis of ultrasonic images of the liver

Conventional two-dimensional (2-D) texture parameters serve as the ''gold standard'' of texture analysis. We compared a new stochastic model, based on autoregressive periodic random field models (APRFM) with conventional texture analysis (CTA) parameter, which were defined as measures of the cooccurrence matrix, i.e., entropy, contrast, correlation, uniformity, and maximum frequency. By fitting the model to a given texture pattern, the estimated model parameters are suitable texture features. In 81 patients, divided into patients without (N = 19) and with (N = 62) microfocal lesions of the liver, a set of 24 CTA and 16 APRFM parameters were calculated from ultrasonic liver images. To ensure simple computation the APRFM parameters were based on the unilateral type of pixel neighborhood. Regenerated texture by APRFM was visually comparable with the original texture. Reclassification analysis using the classification and regression tree (CART) discriminant analysis system and the area under the receiver operating characteristic (ROC) curve was used to assess the texture classification potency of APRFM- and CTA-parameters. Discriminating between liver with and without microfocal lesions, the best results were seen for the APRFM parameter.