Adapting neural networks for rapid segmentation of mineralized tissues in mouse jaws

Congenital defects in dental enamel are diverse in pathology and etiology, and designing treatment tools for the clinic requires fundamental research on the process of enamel formation. Rodent incisors are the model of choice, and microcomputed tomography (μCT) is often the first method of comparison between models. Quantitative comparison of μCT data requires segmentation of mineralized tissues in the jaw; previously, we demonstrated the ability of convolutional neural networks to quickly and accurately segment mineral gradients in mouse jaws in synchrotron μCT images. Here we greatly expand on that work and present a protocol for adapting base networks to new pathologies and data types. With collaborators, we have amassed a collection (~80 TB) of μCT images from laboratory machines and synchrotrons representing 18 genetic mouse lines. We demonstrate the ability of adapted networks to segment these new data without compromising accuracy. Specifically, our networks adapted well to data collected with different x-ray sources, voxel dimensions, and phenotypes. In fully segmented data, we demonstrate the ability to visualize stages during enamel formation and compare rates of change in mineral density during the process. Importantly, our work has revealed insights about how and when mineral deposition goes awry in defective enamel. We envision widespread use of these tools. Once base networks are deployed to a repository for artificial neural networks, researchers will be able to use the protocol we present here for using modest amounts of their data to adapt a network to their own analysis.