NA-MIC Project Weeks

Back to Projects List

Universal Tooth Labeling Module

Key Investigators

• Enzo Tulissi (University of Michigan, USA)
• Lucia Cevidanes (University of Michigan, USA)
• Juan Prieto (University of North Carolina, USA)
• Jonas Bianchi (University of Pacific, USA)

Project Description

The Universal Tooth Labeling module employs nnUNet to automatically label all teeth (including primary dentition) from CBCT scans.
It aims to deliver robust, anatomically correct labels for each tooth.

Currently, some outputs exhibit left/right mirroring errors (e.g., both canines labeled as “right canine”).

Objective

1. Resolve mirroring errors in the labeling output.
2. Optimize nnUNet training (patch size, batch size, learning rate) and augmentations.
3. Implement post‐processing checks to verify and correct side‐specific labels.

Approach and Plan

1. Analyze cases with mirroring errors and identify their characteristics.
2. Tune nnUNet hyperparameters for optimal label accuracy.
3. Develop a post‐processing module to enforce correct left/right assignment.
4. Evaluate performance (DSC, IoU) on a dedicated test set.

Progress and Next Steps

1. Describe specific steps you have actually done. Completed:
   • Initial nnUNet pipeline and label export implemented.
   • Prototype output integrated into 3D Slicer.

Next Steps:

• Fix mirroring bugs in output labels.
• Conduct clinical validation and benchmarking.

Illustrations

Figure 1: Example output from Universal Tooth Labeling

Background and References

• Universal Tooth Labeling (in progress)