Internship for Engineers/Master's Students: Training a DeepSDF Model on Optical Impressions to Evaluate Dental Wear

Context

Optical impressions and 3D acquisition technologies are now essential tools for capturing dental anatomical shapes with unparalleled precision. Various machine learning models trained on extracted teeth have already been developed and have shown good performance in modeling dental shapes (Binvignat et al., 2024). However, in vivo optical impressions are not limited to a single tooth and exhibit greater variability in shapes, making the development of a precise model more complex. Recently, the DeepSDF approach has been specifically adapted for the reconstruction of fractured incisors from a set of 80 anatomies acquired by cone-beam computed tomography (CBCT), demonstrating its potential for highly precise dental reconstructions (Chen et al., 2024). DeepSDF works by learning a continuous signed distance function representing the set of shapes (Park et al., 2019). This method allows for precise and faithful reconstruction of complex shapes from partial data, making it particularly suitable for detailed dental applications. The hypothesis is that DeepSDF could handle a larger set of incisors scanned with high-quality intraoral scanners, potentially surpassing techniques based on Principal Component Analysis.

This internship proposes to explore this issue by using a DeepSDF model to learn the shapes of intact or worn teeth from optical impressions of dental arches and to evaluate the severity of this wear.

Project Objective:

The objective of this project is to evaluate the ability of a DeepSDF model to capture variations in dental shapes from optical impressions. A sub-objective will be to compare the performance of different classification methods based on latent codes to identify levels of dental wear.

Project Steps:

  • Model Alignment: Use registration techniques to align dental impressions. Use already developed PointNet techniques to segment dental shapes.
  • Training the DeepSDF Model: Generate a set of 3D points from STL files, then train the model (work already done on extracted teeth, to be adapted to the specifics of an open optical impression file).
  • Evaluation of Learning: Calculate performance metrics (average Euclidean distance, surface difference, Hausdorff distance).
  • Comparison of Classification Methods: Use techniques such as Random Forest on latent codes to classify dental shapes according to their wear.

Required Skills:

Knowledge in geometric modeling, mesh processing, and machine learning.
Proficiency in Python and deep learning libraries (PyTorch or TensorFlow).
Experience with 3D software (e.g., Rhino, MeshLab) and managing 3D files.
Perspectives:
The internship will allow the development of skills in digital modeling applied to dentistry and machine learning, with opportunities in research or R&D engineering in the health sector.

 

Contact:

References :

Binvignat P, Chaurasia A, Lahoud P, Jacobs R, Pokhojaev A, Sarig R, Ducret M, Richert R. 2024. Isotopological remeshing and statistical shape analysis: Enhancing premolar tooth wear classification and simulation with machine learning. J Dent. 149(April):105280. https://doi.org/10.1016/j.jdent.2024.105280.

Chen D, Yu M, Li Q, He X, Liu F, Shen J. 2024. Precise tooth design using deep learning-based templates. J Dent. 144(March):104971. https://doi.org/10.1016/j.jdent.2024.104971.

Park JJ, Florence P, Straub J, Newcombe R, Lovegrove S. 2019. Deepsdf: Learning continuous signed distance functions for shape representation. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2019-June:165–174.