Dilation-Erosion Methods for Radiograph Annotation in Total Knee Replacement

Yehyun Suh, Aleksander Mika, J. Ryan Martin, Daniel Moyer

Keywords: Label Augmentation, X-Ray, Landmark Annotation

TL;DR: We present a new training scheme for image annotation, which improves performance over standard methods for knee radiograph assessments.

Abstract: In the present work we describe a novel training scheme for automated radiograph annotation, as used in post-surgical assessment of Total Knee Replacement. As we show experimentally, standard off-the-shelf methods fail to provide high accuracy image annotations for Total Knee Replacement annotation. We instead adopt a U-Net based segmentation style annotator, relax the task by dilating annotations into larger label regions, then progressively erode these label regions back to the base task on a schedule based on training epoch. We demonstrate the advantages of this scheme on a dataset of radiographs with gold-standard expert annotations, comparing against four baseline cases.