Project proposes development of: (i) self-supervised data adaptive algorithms for clustering of large datasets generated from (non)linear low-dimensional subspaces; (ii) algorithms for linear regression with data adaptive sparse regularization functions (surrogates of L0 quasi-norm); (iii) algorithms for low-rank tensorial models based approximation of multidimensional data for extraction of discriminative features. Algorithms developed under (i) will be applied to semantic segmentation of hyperspectral image of stained frozen section of adenocarcinoma of a colon in a liver, as well as to semantic segmentation of RGB images of stained hitopathological specimens of three frequent cancers of a liver (hepatocellular carcinoma, cholangiocellular carcinoma and adenocarcinoma of colon in a liver). These applications are aimed to quantify contribution of hyperspectral imaging in semantic segmentation, as well as to quantify the quality of various models for semantic segmentation of RGB images of specimens with diagnosed frequent carcinomas of a liver. Algorithms developed under (ii) will be applied to identification of metabolites present in small proportions in 1H NMR spectra of human urine of subjects with diabetes type 2. This approach will rely on extended library (290 to 330) of metabolites indicative for human urine with/without diabetes type 2. This application is important for non-targeted metabolic profiling with the objective to identify as much metabolites as possible present in spectra of biological specimen. Algorithms developed under (ii) and (iii) will be applied to diagnosis of frequent retina diseases through extraction of discriminative features from 3D optical coherence tomography images. This application is important as potential alternative to deep networks that are demanding for training.