3D Geometric Tooth Alignment Planning via Deep Reinforcement Learning

3D geometric tooth alignment planning is a critical, labor-intensive task in orthodontics that involves calculating the sequential movement of teeth from an initial malocclusion to a target alignment. Current clinical practice relies heavily on manual planning, while existing automated methods—often based on particle swarm optimization—struggle to handle complex spatial constraints like inter-dental collisions or to generalize across diverse dental morphologies. This paper asks whether a deep reinforcement learning (DRL) framework can learn to generate safe, efficient, and clinically viable alignment trajectories by treating the planning process as a sequential decision-making problem.

The researchers formulate tooth alignment as a Markov Decision Process (MDP). In this setup, an agent observes the current state of the dentition (pose, target pose, and shape embeddings) and outputs movement actions for each tooth. To handle the high-dimensional complexity of dental arrangements, the authors introduce three key innovations to the Deep Deterministic Policy Gradient (DDPG) algorithm:

1. Transformer-based Agent: A Transformer architecture with multi-head attention models the spatial relationships between teeth, using both absolute and learnable relative positional encodings to prioritize interactions between adjacent teeth.
2. Dynamic Masking: To mimic clinical logic where only a subset of teeth are moved at any given time, the model employs a masking scheme that restricts movement to a sparse subset of teeth per step.
3. Two-Stage Curriculum Learning: The training process is split into two phases. The first phase focuses on finding a general path to the target with relaxed constraints, while the second phase introduces stricter penalties for collisions and path inefficiency to refine the trajectory.

Findings and Significance

The proposed method was evaluated against a dataset of 10,000 expert-designed clinical treatment plans. The results demonstrate that the DRL-based agent outperforms traditional optimization baselines in both path safety (collision avoidance) and geometric efficiency. By framing the problem as a sequential decision-making task rather than a single-pass generation problem, the model better captures the clinical reality of orthodontic treatment. This work provides a scalable, automated solution that could significantly reduce the time and labor required for orthodontic treatment planning.