X-ray Microtomography is a proven tool for phase fraction analysis of multi-phase systems, provided that each phase is adequately partitioned by some means of data processing. For porosity in materials containing low-density ceramic phases, differentiation between pores and the low-density phase(s) can be intractable due to low scattering in the low-density phase, particularly if small pores necessitate low binning. We present a novel, combined methodology for accurate porosity analysis—despite these shortcomings. A 3-stage process is proposed, consisting of (1) Signal/noise enhancement using non-local means denoising, (2) Phase segmentation using a convolutional neural network, and (3) Quantitative analysis of the resulting 3D pore metrics. This particular combination of denoising and segmentation is robust against the fragmentation of common segmentation algorithms, while avoiding the volitional aspects of model selection associated with histogram fitting. We discuss the procedure applied to ternary phase SiC–TiC-diamond composites produced by reactive spark plasma sintering with porosity spanning 2–9 vol%.