Porous ceramics are promising materials for the energy transition, for use in lightweight structures, as support materials for catalysts or heat exchangers. The surface to volume ratio gives them an advantage over dense materials. The most commonly used industrial process is the replication process based on heterogeneous polymer foams or templates. Burning out the struts leaves hollow struts with the typical triangular structure of polymer foams. Using microtomography (µCT) measurements, the microstructure can be displayed as a stack of 2D cross-sectional images, which are used for evaluation. However, current software solutions cannot separate the different types of pores. Using machine learning, a neural network based on the Eff-Net architecture is employed. The necessary training network is generated by Liner.AI software using 2,800 training images of a 30 ppi Al2O3 foam with up to 10,000 training iterations and data augmentation.  The recognition accuracy for the strut pores was over 87%, while that for the foam pores was 99%. This is then used in the next stage of the Tensorflow-based neural network to identify the strut pores in arbitrary ceramic replica foams. Here, 1200 images of another ceramic foam and a 30ppi PU foam are analysed. The network identified 16.86% material, 81.91% foam pores and 1.23% strut pores for the ceramic. For the PU foam only material (3.9%) and foam pores (96.1%) were identified. For the ceramic, the values obtained are consistent with the total porosity. In addition, the amount of coating on the replica foam can be deduced from the µCT images.