Objective
Epidemiologic studies report that cracked teeth are the third most common cause of tooth loss in industrialized countries. Current diagnostic tools have a limited ability to accurately diagnose cracks. There is an imperative need to develop an objective and reliable method to detect cracks beyond information obtained from clinical and radiographic evaluation. Kitware and UT Health San Antonio School of Dentistry have developed a novel algorithm for crack detection, though it requires reliable tooth isolation, i.e. segmentation, method to work appropriately. There has been inconsistent segmentation using this algorithm when scans from different cone beam computed tomography (CBCT) machines were used. In this abstract, we present a robust convolutional neural network (CNN)-based segmentation method that works on several small field of view CBCT scans acquired by different CBCT machines.
Study Design
Data show that regular CNN segmentation models fail to generalize to new acquisitions when scanner protocols shift and upgrade, a problem known as domain shift. To overcome this, we successfully generalized 3-dimensional Fourier Domain Adaptation methods to build 3-dimensional tooth segmentation models that are robust to domain shift. The method works by finding the transformations between a source domain into an adapted target domain in the Fourier space. Applying this method to multiple small field of view CBCT scans acquired by different machines resulted in successful segmentation of the teeth, tested on multiple scans.
Results
This development enables the use of our algorithm (as well as other algorithms) on scans from a variety of CBCT machines, thus vastly improving their generalizability.
Conclusion
The access to a reliable, single tooth segmentation method will enable the early detection and localization of tooth pathology, including cracks. This along with appropriate interventions has the potential to enable effective strategies to prevent tooth loss. This technology may also be applied to other dental applications that require use of automated segmentation of teeth. Funded by National Institutes of Health/National Institute of Dental and Craniofacial Research R44DE027574