Automatic Landmark Detection in Severe Craniomaxillofacial Deformities via Three-Dimensional (3D) Point Cloud Deformation Model and Deep Learning Networks

Abstract

The accurate and efficient detection of landmarks is crucial for diagnosing and treating CMF deformities, which often involve various structural abnormalities in bone structure displacement, defects, and facial asymmetry. The automated detection of CMF malformation landmarks has been extensively studied; however, these studies have been hindered by significant computational burdens and limited patient populations. Furthermore, the automated detection of landmarks in patients with bone defects and facial asymmetry has been overlooked. This research proposes a new landmark detection framework based on point cloud deformation model and deep learning networks. Firstly, to address the issue of insufficient patient data, we employ the point cloud deformation model for data augmentation. By means of deformation, severely deformative data can be transformed from normal human data for model training. The next step involves applying a 3D point cloud CNN model to all landmarks, enabling coarse localization. Subsequently, an assessment is made on whether there are any bone defects within the region where these landmarks are located. The accurate localization is achieved using two different models: 1) For landmarks in areas without bony structural defects (normal landmarks), the CNN semantic segmentation model for 3D images is used; 2) For landmarks in defect areas (defect landmarks), the 3D point cloud CNN semantic segmentation model is applied. The effectiveness of our proposed framework is demonstrated through the presentation of experimental results. Collectively, our work establishes a novel pathway for surface-based morphometry and medical shape analysis. Through two stages of localization, the accuracy rate is finally $1.19~\pm ~0.71$ mm (CT), $0.91~\pm ~0.42$ mm (CBCT) for normal landmark and $1.13\pm 0.75$ mm (CT), $0.94~\pm ~0.51$ mm (CBCT) for defect landmark, which is superior to the most advanced related methods.

Keywords

• Craniomaxillofacial (CMF) deformities
• landmark detection
• deep learning
• three-dimensional (3D) cephalometric measurements
• point cloud