Characterize Skeletal Malocclusion and Make Treatment Decisions via a 3D Geometric Features Extraction Enhanced Neural Network

Abstract

Background and objective: Simple malocclusion cases only require orthodontic treatment to align the teeth, while patients with skeletal malocclusions need additional orthognathic surgery to correct their jaw deformities. Neglecting skeletal abnormalities would lead to incorrect diagnosis and treatment planning. Accurate measurement and analysis of a three-Dimensions(3D) skull is critical to true diagnosis and optimal treatment planning. We hypothesized that an end-to-end deep learning-based decision support systems could be developed to automatically characterize skeletal deformities using 3D skull data and suggest appropriate surgical methods if orthognathic surgery is needed.

Methods: This paper presents a point cloud multi-label classification neural network, in which a multi-resolution feature learning module was designed to extract features that contain both high-semantic and high-resolution information. Five variables of diagnosis and three variables of surgery type were included in prediction tasks. 766 real-world, well-annotated skulls in point cloud format (half normal subjects, half skeletal malocclusion) were collected for training and validation. Several training strategies and innovative data augmentation methods were also applied to enhance efficient geometric feature extraction as well as tackle imbalanced dataset. The performance of proposed neural network, Point-Malo, was evaluated with multi-classification accuracy, ablation experiments and compared with competing methods on both our dataset and two public datasets.

Results: Our proposed model, Point-Malo, obtained higher prediction accuracy than other competing methods on our dataset. The Overall Accuracy (OA) in validation set for each label were 0.92 (Skeletal type), 0.71(Maxilla), 0.73 (Mandible), 0.69 (Chin), 0.73 (Occlusion), 0.93 (Le Fort I), 0.91 (BSSRO) and 0.93 (Genioplasty). Ablation experiments showed the multi-resolution feature learning module greatly improved model performance. Our model also presented exceptional classification performance on two public datasets ModelNet40 and ScanObjectNN.

Conclusions: Our model outperforms existing methods based on our dataset and two public point cloud datasets ModelNet40 and ScanObjectNN. The developed methods can serve as a powerful screening and diagnosis tool for patients with skeletal or simple malocclusion.

Note:
Funding Declaration: This work was supported by National Natural Science Foundation of China (81571022), Natural Science Foundation of Shanghai Municipality (23ZR1438100), Multi-center clinical research project of Shanghai Jiao Tong University School of Medicine (DLY201808).

Conflicts of Interest: None

Ethical Approval: This research was approved by the Research Ethics Committee in Shanghai Ninth People’s Hospital (IRB No. SH9H-2022-TK12-1).

Keywords: geometric deep learning, Classification, skeletal malocclusion, point cloud, 3D shape analysis, multi-resolution features learning.