Abstract

A orthodontic treatment simulation is normally made up of several cycles of treatment, that regularly covers for more than 12 months. Thus, the fundamental determinant in orthodontic treatment is to simulate medically reasonable teeth position in long-term progress. However, existing orthodontic treatment simulation system heavily rely on duplication of efforts from dentists or only estimate final tooth arrangement without orthodontically intermediate procedure. Toward clinically reasonable simulate 3D orthodontic treatment progress, we present DeepOrtho, a deep learning based novel system to simulate medically 3D tooth position for orthodontic treatment planning. Our system takes 3D tooth meshes from patients with malocclusion as input, and sequentially simulates the orthodontically proper 3D rotation and translation for each tooth within the long-term treatment. Notably, we formulate the 3D orthodontic treatment simulation as a reverse process of iteratively denoising teeth arrangements, where DeepOrtho gradually reduces medically uncertain sequences from all the teeth adjustable positions until reaching the desired positions. To the best of our knowledge, we are the first medical simulation system to explore progress 3D orthodontic treatment. Extensive experiments demonstrate that the proposed DeepOrtho outperforms existing solutions in terms of performance and clinical feasibility.