Class-Oriented Features Selection Technology in Medical Images Classification Problem on the Example of Distinguishing Between Tuberculosis Sensitive and Resistant Forms

This work proposes a class oriented technology for features ensembles formation based on preliminary sequential selection by interclass and intraclass dispersion criteria for final promising ensembles selection by the combinational criterion of correlation with classes of the most pairwise independent features. The speeding up of the complete ensembles selection at the final stage is realized by the self-organization procedure based on a genetic algorithm. Depending on the primary features set structure, it is possible to obtain a design of both a single sequential selection technology and parallel processing according to given channels, which are combined on the final stage. The Random Forest algorithm was used for classification with optimization of the voting function by the group method of data handling. The provided technology was used for differentiation of sensitive and resistant tuberculosis on patient CT images, which is one of the urgent needs of modern medicine. We used matrix elements of textural features as primary features: Grayscale Colors Matching Matrix, Grayscale Spacing Matrix, Grayscale Leap Length Matrix, Grayscale Tone Difference Matrix, Grayscale Dependency Matrix, Gray Level Histogram Matrix. For the regions of interest highlighted in the images, high quality of classification was obtained: 85% accuracy in the examination sample. Data for the research were provided by the State Institution “F.G. Yanovsky National Institute of Phthisiology and Pulmonology of the NAMS of Ukraine”.