Data representation and reduction for chromatin texture in nuclei from premalignant prostatic, esophageal, and colonic lesions

Abstract

To identify nuclei and lesions with great specificity, a large set of karyometric features is arranged in the form of a linear profile, called a nuclear signature. The karyometric feature values are normalized as z-values. Their ordering along the profile axis is arbitrary but consistent. The profile of the nuclear signature is distinctive; it can be characterized by a new set of variables called contour features. A number of data reduction methods are introduced and their performance is compared with that of the karyometric features in the classification of prostatic, colonic, and esophageal lesions. Contour characteristics were reduced to descriptive statistics of the set of z-values in the nuclear signature and to sequence information. The contour features derived were (1) relative frequencies of occurrence of z-values and of their differences and (2) co-occurrence statistics, run lengths of z-values, and statistics of higher-order dependencies. Performance was evaluated by comparing classification scores of diagnostic groups. Rates for correct classification by karyometric features alone and contour features alone indicate equivalent performance. Classification by a combined set of features led to an increase in correct classification. Image analysis and subsequent data reduction of nuclear signatures of contour features is a novel method, providing quantitative information that may lead to an effective identification of nuclei and lesions.