Adapting neural networks for rapid segmentation of mineralized tissues in mouse jaws (Invited Paper)

Quantitative comparison of mineral in rodent incisors requires segmentation of relevant tissues, which can be accomplished quickly and accurately with convolutional neural networks. Here we describe a protocol to adapt base networks to new data, thereby creating segmentation tools broadly useful for the diverse datasets necessary to compare rodent models. Specifically, we used μCT images collected for 18 mouse lines from both synchrotron and laboratory X-ray sources with a variety of voxel dimensions and image artifacts. We demonstrate the success of our adapted networks and show the results of a subsequent data processing pipeline for quantitatively comparing stages of mineral formation in the incisor. We envision enamel researchers adapting base networks with modest amounts of their own data, following the protocol shown here, to suit their needs and rapidly extract quantitative measures. The insights gained from these methods will greatly contribute to our understanding of pathologies in dental tissues.