The present paper proposes a machine learning (ML) framework for image segmentation of micro-CT scans of ceramic matrix composites (CMCs). Specifically, we study CMC minicomposites with continuous silicon carbide fibers embedded in the silicon carbide matrix. Fibers were coated with boron nitride. To isolate the features of the microstructure, such as matrix, fibers, fiber coatings, and pores, we develop a sequential two-class semantic image segmentation model based on 2.5D U-Net deep learning architecture, which was implemented in the Dragonfly software package. The obtained ML models are then employed to construct and train a five-class semantic image segmentation ML platform for simultaneous recognition of fibers, fiber coatings, pores and the matrix. The introduced methodology is illustrated on image segmentation of CMC minicomposites; including analysis and estimation of volume fractions of composite phases and comparisons with the existing in the literature results.