Semantic segmentation of mouse jaws using convolutional neural networks

Defects in tooth enamel are associated with a multitude of health conditions. An ongoing push to improve our understanding of enamel formation is generating a large number of mutant mouse lines to map protein function and create an Enamel Atlas. Reproducible analysis of a large amount of micro-CT data to compare these mouse lines necessitates an automated, high-throughput method of segmenting enamel, bone, and dentin. Neither simple binary segmentation nor region growing algorithms are effective for enamel in continuously growing mouse incisors due to the gradient in mineral content. To overcome these limitations, we have trained and validated a 3D convolutional neural network (CNN) to semantically segment mouse jaws. The network adopted a UNet architecture and incorporated training data from synchrotron- and laboratory-based sources. We evaluated the performance of the 3D CNN for the wildtype by comparing segmented outputs to ground truth labels and to outputs from a similarly trained 2D network. Next we tested the adaptability of the network by segmenting mutant tissues displaying phenotypes ranging in severity. Finally, we will demonstrate the use of CNN-segmented datasets to calculate metrics for quantitative comparison of the 3D mineral distribution between wildtype and mutant genotypes. We will discuss segmentation of the incisor, which allows us to track changes in the mineral during each developmental stage of enamel production. 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.