The contribution of image cytometry and artificial intelligence-related methods of numerical data analysis for adipose tumor histopathologic classification

Abstract

Thirty-five lipomatous tumors were quantitatively described using 47 variables generated by means of computer-assisted microscope analysis. Of these 47 quantitative variables, 27 were computed on Feulgen-stained specimens (25 on cytologic and 2 on histologic samples) and, of the remaining 20, 8 related to vimentin and S-100 protein immunostaining patterns and the other 12 to the glycohistochemical staining patterns of peanut agglutinin, succinylated wheat germ agglutinin, and concavalin A agglutinin. The 35 lipomatous tumors included 6 atypical lipomas and 8 well differentiated, 5 dedifferentiated, 6 myxoid, and 10 pleomorphic liposarcomas. The actual diagnostic value contributed by each of the 47 variables with respect to the 5 lipomatous tumor groups was determined by means of the decision tree technique, an artificial intelligence-related algorithm that forms part of the supervised learning algorithms. Of the 47 quantitative variables, the decision tree technique retained 8: i.e., 2 tissue architecture-, 2 DNA ploidy level-, 2 morphonuclear-, 1 lectin histochemical-, and 1 vimentin immunostain-related variables. The decision tree technique made use of these 8 variables to set up logical rules that make it possible to identify atypical lipomas from well differentiated liposarcomas, on the one hand, and dedifferentiated liposarcomas from those that are well differentiated and pleomorphic, on the other. Thus, the combination of an artificial intelligence algorithm analyzing quantitative variables generated by means of the computer-assisted microscope analysis of cytologic and histologic samples from lipomatous tumors can be considered an expert system contributing significant diagnostic information to conventional diagnosis.