NA-MIC Project Weeks

Back to Projects List

Interpretable Deep Learning for the Detection and Classification of Impacted Canines and severity of root resorption

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 CLIC module uses a Mask R-CNN segmentation model to locate and classify impacted canines in CBCT scans.
This project extends CLIC by developing a supervised model to automatically classify the severity of root resorption in teeth adjacent to impacted canines.

Objective

1. Segment impacted canines using the existing CLIC module.
2. Extract adjacent teeth volumes for analysis.
3. Assemble an annotated dataset of adjacent teeth with clinician‐provided resorption severity scores.
4. Train a classification model to predict root resorption severity from segmented volumes.
5. Integrate and visualize segmentation plus severity scores within 3D Slicer.

Approach and Plan

1. Run CLIC across the CBCT dataset to isolate impacted canines.
2. Combine CLIC masks to extract volumes of adjacent teeth.
3. Annotate each extracted tooth volume with a severity label (mild, moderate, severe).
4. Extract geometric and morphological feature sets from each volume.
5. Train a classifier on these features.
6. Extend CLIC or create a new Slicer module for real‐time severity classification.
7. Validate model performance (accuracy, recall, precision) and integrate into the 3D visualization workflow.

Progress and Next Steps

Completed:

• CLIC module validated.
• Prototype pipeline for adjacent tooth extraction established.

Next Steps:

• Clinician annotation of extracted tooth volumes.
• Feature engineering and model training.
• UI design for severity score display in Slicer.
• Performance evaluation and final documentation.

Illustrations

Figure 1: Impacted canine segmentation results from CLIC

Background and References

No response