Unsupervised Domain Adaptation for Human and Animal Chest X-Ray Bone Suppression

Bone suppression of chest X-ray images is an effective method for improving the reliability of diagnosis of lung diseases by X-ray reading. Deep learning-based methods for bone suppression, which learn the difference between a pair of images with and without bone, have attracted much attention due to their high performance. However, boneless images are not always available especially for animals, making it hard to use existing training methods. Therefore, we propose an unsupervised bone suppression method that does not require any boneless X-ray images. Unsupervised domain adaptation is performed at the input and feature levels using CT data as the source domain and X-ray images as the target domain. We verified the effectiveness of the proposed method by qualitative evaluation on an original animal dataset and quantitative evaluation on public human dataset.