Anatomical Landmark Localization for Knee X-ray Images via Heatmap Regression Refined with Graph Convolutional Network

Accurate detection of anatomical landmarks for knee X-ray images holds paramount significance for the comprehensive assessment of knee osteoarthritis. Nevertheless, prevailing heatmap regression methodologies often fall short in fully leveraging the holistic structural information of landmarks, leading to potential inaccuracies in offsets owing to predicted integer coordinates. In this paper, we present a sophisticated end-to-end differential landmark localization model which amalgamates heatmap regression with a graph convolution network to precisely pinpoint anatomical landmarks in knee joint X-ray images. Our innovative approach represents the landmarks as a graph, enabling the capture of crucial structural information. It progressively refines the initially coarse integer landmark coordinates derived from heatmaps using cascade GCNs. Additionally, we incorporate the attention layer within the feature sampling module to augment the precision of regression outcomes. Our method attains remarkable results on the OAI dataset, achieving the Mean Radial Error of 0.84mm and the Successful Detection Rate of 71.18% at 1mm. These outcomes surpass the performance of alternative heatmap regression methods, significantly contributing to the facilitation of osteoarthritis assessment.