Learning Spatial Configuration Feature for Landmark Localization in Hand X-rays

Abstract

Medical landmark localization is crucial for treatment planning. Although FCN-based heatmap regression methods have made significant progress, there is a lack of FCN-based research focused on features that can learn spatial configuration between medical landmarks, notwithstanding the well-structured patterns of these landmarks. In this paper, we propose a novel spatial-configuration-feature-based network that effectively learns the anatomical correlation between the landmarks. Specifically, we focus on a regularization method and a spatial configuration loss that capture the spatial relationship between the landmarks. Each heatmap, generated using U-Net, is transformed into an embedded spatial feature vector using the soft-argmax method and spatial feature maps, here, Cartesian and Polar coordinates. A correlation map between landmarks based on the spatial feature vector is generated and used to calculate the loss, along with the heatmap output. This approach adopts an end-to-end learning approach, requiring only a single feedforward execution during the test phase to localize all landmarks. The proposed regularization method is computationally efficient, differentiable, and highly parallelizable. The experimental results show that our method can learn global contextual features between landmarks and achieve state-of-the-art performance. Our method is expected to significantly improve localization accuracy when applied to healthcare systems that require accurate medical landmark localization.