Genetic mouse models are a powerful tool to study congenital defects in tooth enamel. Comparison of micro-computed tomography (micro-CT) reconstructions between mutant mice and wild-type (WT) controls requires an automated method of segmenting the mineralized tissues of the jaw. However, traditional segmentation methods are easily frustrated by mineral gradients, especially those in forming enamel in the continuously growing incisor. Herein, we demonstrate that convolutional neural networks (CNNs) with UNET architecture can be trained to accurately segment WT jaws into seven classes (molar enamel, incisor dentin, molar dentin, bone, background, Al wire, and either incisor enamel or ectopic mineral). We then evaluate the performance of CNNs trained on WT jaws on two mutant phenotypes, with and without additional cross-training, using CNNs trained only on mutant data for comparison, and find that cross-training provides significant improvement. Finally, we demonstrate that segmentation using a cross-trained network enables extracting metrics for quantitative comparison of mineral gradients in the incisor. Our results show that the CNN-based segmentation and quantification pipeline is a versatile tool that will empower enamel researchers, help delineate mechanisms of disease, and enable the development of new approaches of intervention.