The iron ore pellet forming process consists of preparing the raw materials, forming the raw pellet and hardening by firing. The end product must be a porous material which allows gas to diffuse in the blast furnace and at the same time resists compression, which is a relevant feature during transport and loading of the furnace. However, during heat treatment and transport cracks may appear that compromise the integrity of the pellets. The discrimination of pores and cracks is therefore an important factor for microstructural analysis and material quality control. X-ray microtomography is a non-destructive technique that generates three-dimensional images, allowing a full view of the pellet. However, the usual methodology of digital image processing, based on extraction of size and shape attributes, has limitations to discriminate crack from pores. Deep Learning Neural Networks are a powerful alternative to classifying object types in images, using as input the pixel intensities and attributes automatically determined by the network. After training a model with the patterns corresponding to each class, it is possible to assign each pixel of the image to one of the classes present, allowing a semantic segmentation. In this dissertation, a Deep Learning network with U-Net architecture was optimized, using as a training set a few 2D layers of the original 3D image. Applying the model to the pellet used in training it was possible to discriminate cracks pores properly. Application of the model to other pellets required the incorporation of layers of these pellets into the training and optimization of model parameters. The results were adequately classified, despite the difficulty of creating a general model for discrimination between pores and cracks in iron ore pellets.