Patent Application
20230190409
The goal of the orthodontic treatment planning process is to determine where the post-treatment positions of a person's teeth (setup state) should be, given the pre-treatment positions of the teeth in a malocclusion state. This process is typically performed manually using interactive software and is a very time-consuming process. Furthermore, the course of treatment can change, requiring changes to the setup state. A need thus exists for an algorithm to generate a subset of the setup stages between the initial and final setups
A computer-implemented method for generating stages for a portion of orthodontic aligner treatment includes receiving a digital 3D model of teeth in malocclusion and generating a subset of stages of setups among a complete set of stages of setups for aligner treatment of the teeth.
A computer-implemented method for generating a setup for orthodontic aligner treatment includes receiving a digital 3D model of teeth in malocclusion. The method uses a machine learning model that has been trained using historic setups to generate a proposed final or intermediate setup for the digital 3D model of teeth in malocclusion.
Embodiments include a computerized system to generate stages for a portion of a complete aligner treatment. The system takes as input a digital three-dimensional (3D) model of a malocclusion. Optional input includes treatment guidelines such as a number of stages, amount of tooth movement, or treatment strategy, or any combination thereof. The digital 3D model then undergoes any necessary preprocessing, which may include data cleanup, tooth segmentation, and tooth coordinate system identification. Next, the first N stages of treatment are generated from the preprocessed scan. The user of the system can optionally make modifications to the treatment before sending the digital data to a manufacturing process for tray manufacturing.
Embodiments also include a deep learning model to automatically generate a digital setup from the malocclusion positions of teeth. This process can be divided into two steps: model development and training, and model deployment. During model training, many digital 3D models of patients' malocclusions and setups are input into a deep learning model, which is optimized to learn patterns that minimize the difference between predicted and actual setups. During model deployment, the trained deep learning model is used to generate a setup prediction for new case data.
A typical aligner treatment planning workflow is based on designing an ideal final position of teeth (final setup), then designing a set of stages used to manufacture trays that will move the teeth from the initial setup to final setup. In some alternative workflows, it may be preferable to design only a subset of aligner trays that achieve a certain treatment goal. For example, an orthodontist may want to use a different workflow to design a few stages of treatment to create space between teeth prior to attempting more complex movements.
The following are algorithmic methods to generate and possibly manufacture only a portion of a complete aligner treatment based upon these approaches.
Approach 1: Based on a final ideal setup (step 34).
One approach to this problem involves creating the final ideal setup and all intermediate stages. Only a subset of stages (either the first N of M stages, or stages up to a treatment goal among the final setup and intermediate stages) would then be selected and manufactured into the corresponding aligners.
Approach 2: Based on a target intermediate setup (step 36).
Rather than using a final ideal setup, a target intermediate setup that achieves a set of desired tooth movements can be created. All intermediate stages between the malocclusion and target setup could then be generated and manufactured. This target intermediate setup could either be created manually by a user (doctor or technician), or created algorithmically. Exemplary algorithmic approaches include the following:
Algorithm 1: Given a set of movements that would be desirable to achieve, an algorithm could apply these movements to any malocclusion. For example, if expansion is desired to create space, the algorithm would apply an input amount of crown torque and/or bodily expansion to the teeth. This algorithm can use a rule-based approach to generate the setups from malocclusion to the target intermediate setup. An example of a rule-based approach to generate setups is disclosed in PCT Patent Application Publication No. WO 2019/069191, which is incorporated herein by reference as if fully set forth.
Algorithm 2: One optimization algorithm that creates setups based on optimizing a set of metrics subject to some constraints in described in PCT Patent Application Publication No. WO 2020/026117, which is incorporated herein by reference as if fully set forth. These metrics and constraints can be modified to create target intermediate setups. For example, the algorithm can increase the constraint on tooth movement, which would result in a setup that moves teeth less than the amount allowed for the final setup. The algorithm can also modify the metrics to penalize certain types of movement that may be difficult to achieve at first (e.g., root torque) and promote desirable movement types (e.g., expansion). The optimization algorithm can be run with these modified constraints and metrics to create an optimal target setup. Table 1 provides exemplary pseudocode for generating final setups for this optimization-based approach.
The method for this approach can modify metrics (change the penalty term in the Scoring function) and/or constraints (change the Constrain function) to create target intermediate setups. For example:
1. Constraints: Increase the constraint on tooth movement, which would result in a setup that moves teeth less than the amount allowed for the final setup. The Constrain function would move the teeth in the current state to a position in which the movement between the maloccluded state and the current state is less than a certain amount.
2. Metrics: Penalize certain types of movement that may be difficult to achieve at first (e.g., root torque). The penalty term in the scoring function would measure the amount of tooth movement for these types of movement. Promote desirable movement types (e.g., expansion). The penalty term would penalize movements that are less than a threshold amount by measuring how much less the current movement is compared to an ideal amount.
The optimization algorithm can be run with these modified constraints and metrics to create an optimal target setup.
Algorithm 3: Given a set of intermediate target setups from previously treated patients, a machine learning model can be trained to generate intermediate target setups. Given a malocclusion for a new patient case, this trained model can then be used to generate a custom intermediate target setup for the new case.
Approach 3: Sequential stage generation (step 38).
Given a set of teeth in a malocclusion position, a subsequent set of teeth that are displaced from the initial malocclusion (Stage 1) may be generated. From Stage 1, Stage 2 can be generated, and more stages generated until the desired number of stages are generated. Exemplary algorithmic approaches to generate stages sequentially are detailed below.
Algorithm 1: Given a set of movements that would be desirable to achieve and that respect per-stage tooth movement limits, an algorithm could apply these movements to the malocclusion as well as any subsequent stage that has been generated.
Algorithm 2: The constraints on tooth movement detailed in the optimization algorithm above (Approach 2, Algorithm 2, Constraints) can be modified to reflect per-stage tooth movement limits. Specifically:
Constraints: Increase the constraint on tooth movement, which would result in a new state that moves teeth no more than the amount allowed between consecutive states. The Constrain function would move the teeth in the current state to a position in which the movement between the previous state and the current state is less than a certain amount.
The optimization algorithm may then be run on the malocclusion or any stage to generate the next stage.
Algorithm 3: Given a setup of intermediate stages from previously treated patients, a machine learning model can be trained to generate the next intermediate stage from the current stage. For this Algorithm 3, a target setup need not be generated; rather, the stages are generated in sequence from one to the next.
An optimization based approach for determining final setups is described in PCT Patent Application Publication No. WO 2020/026117. This approach includes a method of arriving at a final setup by trying to optimize scores of quality metrics related to a good final setup such as midline, class relationship, alignment, etc. This method can be altered more directly to suit the needs of a particular protocol change, need or preference. For example, if it is desired that the root movement should be reduced, a penalty or a weight on the cost function related to root movement can be increased. However, programming more complex movements using this algorithm can be challenging.
As more data is acquired, machine learning methods and particularly deep learning methods start performing on par or exceed the performance of explicitly programmed methods. Deep learning methods have the significant advantage of removing the need for hand-crafted features as they are able to infer useful features using a combination of non-linear functions of higher dimensional latent or hidden features, directly from the data through the process of training While trying to solve the final setup problem, directly operating on the malocclusion 3D mesh can be desirable. Methods such as PointNet, PointCNN, and MeshCNN can address this problem.
Alternatively, deep learning from the methods of
Customization of these models to perform different types of treatment plans can be achieved easily by training the model with data belonging to that category, for example data of a particular doctor or data from cases where anterior teeth only were expanded.
In generating the setup, it is often required that certain teeth not be moved. If a tooth is marked as fixed, it may not be moved from its original position in the patient's mouth. If it is marked as pinned, it may not be moved from a certain position. The deep learning algorithm described herein learns to generate setups that are similar to setups made by others, with no guarantee that the fixed and pinned teeth remain unmoved. One possible approach to keeping fixed and pinned teeth in place is to adjust lambdas in the machine learning loss function so that it heavily penalizes movement of teeth that a technician has specified as either being fixed or pinned. In such a loss function, teeth are divided in to two groups—those that are fixed or pinned (indicated by values of 1.0 in the input vector), and those that are not (indicated by values of 0.0 in the input vector). During training, loss is calculated separately for each group by calculating the mean-squared-error (MSE) of the difference in tooth positions between the ground-truth positions as placed by a technician, and the positions generated by the neural network during training. The MSE pertaining to the fixed and pinned teeth is then multiplied by a lambda weighting factor when calculating total loss. The equations for this approach are provided in Table 2.
This approach does not guarantee that the fixed and pinned teeth are not moved, so after a setup is generated by the neural network, the fixed and pinned teeth are moved back to their correct positions. The desired result is that the fixed and pinned teeth have moved a small enough amount such that moving them back into position does not cause a large problem with collisions with the other teeth. Increasing the value of lambda during training should not largely affect the positions of teeth generated by the deep learning algorithm.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/IB2021/053962 | 5/10/2021 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 63033887 | Jun 2020 | US |
