Bone anatomical landmark localization with cascaded spatial configuration network

Abstract

Accurate localization of bone anatomical landmarks in medical images is of great significance for decision-making in surgical plans and orthopedic surgery visual navigation systems development. However, due to the high dimensionality and large size of medical images, current automatic anatomical landmarks detection methods still have issues in terms of accuracy, robustness, and efficiency. To improve the accuracy of bone anatomical landmark detection in medical images, a new network model with a two-step strategy (coarse localizing and fine localizing) was proposed in this study. By combining fully convolutional neural networks and the heatmap regression model, a cascaded spatial configuration network was designed to combine the global and local features of 3D image features and localize anatomical landmarks stepwise. The model was evaluated with a collected knee CT image dataset and a published spine CT image dataset. The results were compared to existing state-of-the-art models. The proposed method outperformed other models with an average error of 1.31 mm for knee landmarks and 5.31 mm for the localization of spine landmarks. The outlier rates at error radius 3 mm, 5 mm, and 7 mm are also smaller compared with other models, indicating good robustness of the model. Our proposed method provides a new neural network model with reasonably good accuracy and robustness with a limited computational cost for landmark localization tasks.