Landmarks Detection with Anatomical Constraints for Total Hip Arthroplasty Preoperative Measurements

Total hip arthroplasty (THA) is a valid and reliable treatment for degenerative hip disease, and an elaborate preoperative planning is vital for such surgery. The key step of planning is to localize several anatomical landmarks in X-ray images for preoperative measurements. Conventionally, this work is almost conducted by surgeons manually that is labor-intensive and time-consuming. In this paper, we propose an automatic measurement method by detecting anatomical landmarks with the latest deep learning approaches. However, locating these landmarks automatically with high precision in X-ray images is challenging since image features of a certain landmark are subject to the variations of imaging postures and hip appearances. To this end, we impose the relative position constraints on each landmark by defining edges among landmarks according to the clinical significance. With multi-task learning, our method predicts the landmarks and edges simultaneously. Thus the correlations among these landmarks are exploited to correct the detection deviations implicitly in the network training. Extensive experiment results on two datasets have indicated the superiority of the anatomical constrained method and its potential for clinical applications.