Evaluating Deep Learning Uncertainty Measures In Cephalometric Landmark Localization

Cephalometric analysis is a key step in the process of dental treatment diagnosis, planning and surgery. Localization of a set of landmark points is an important but time-consuming and subjective part of this task. Deep learning is able to automate this process but the model predictions are usually given without any uncertainty information which is necessary in medical applications. This work evaluates three uncertainty measures applicable to deep learning models on the task of cephalometric landmark localization. We compare uncertainty estimation based on final network activation with an ensemble-based and a Bayesian-based approach. We conduct two experiments with elastically distorted cephalogram images and images containing undesirable horizontal skull rotation which the models should be able to detect as unfamiliar and unsuitable for automatic evaluation. We show that all three uncertainty measures have this detection capability and are a viable option when landmark localization with uncertainty estimation is required.