A deep learning framework for the localization of landmarks on the lateral semi circular canals

This paper introduces a Deep Learning (DL) framework to localize landmark coordinates within the semicircular canals in Computed Tomography (CT) scans of the temporal bone. These landmarks can be consistently defined across patients and imaging modalities and as such can serve as a means of forming a common coordinate system. We propose a DL based framework for automating the landmark selection process. We establish the accuracy of the methods using Bone Beam CT scans of the temporal bone of 20 patients and landmarks selected by 3 human experts as the ground truth. We show that the error rates are similar to the levels of variation in landmark selection achieved by human experts. We further validated the method on CT scans from 14 additional patients, demonstrating that the accuracy remains within clinically acceptable parameters.