Global Characteristic Guided Landmark Detection for Genu Valgus and Varus Diagnosis

Lingfeng Ma, Chuanbin Liu, Sicheng Zhang, Yizhi Liu, Hongtao Xie

Abstract: Genu valgus and varus (GVV) are common orthopedic deformities for children. The fundamental step for GVV diagnosis is to locate and identify anatomical landmarks in X-rays. However, it is quite challenging for both humans and computers to accurately detect the landmarks, due to the lack of distinctive position clues. In this paper, we develop a deep learning method named GVV-Net to tackling this issue by fully exploiting the global information inherent in the bones, such as the stable structure features and the slender shape features. Firstly, we propose the Spatial Dependency Mining (SDM) module for capturing the long-range latent dependency and providing global structural information as a supplementary position clue. Secondly, we develop the Vertical Information Aggregation (VIA) module providing a holistic view to help the network perceive more informative regions, which coincides with the slender shape of bones. Besides, we construct the first public dataset with 1555 X-ray images for deep learning research. We achieve an accurate performance in landmark detection with 1.764 mm in point error. The experimental results verify that our method can be reliable assistance for clinical application of GVV.