Automated segmentation of dental anatomy from 3D CBCT scans using a 3D U-Net. Built during a funded research internship at NIT Calicut in collaboration with KMCT Hospital.
Segments six anatomical structures from volumetric dental CT scans:
- Edentulous (missing tooth) regions
- Mandible & Mandibular nerve
- Upper & lower teeth
- Maxilla / upper skull
Outputs are compatible with 3D Slicer for clinical visualization and implant planning.
| Structure | Dice Score |
|---|---|
| Mandible | 0.65 |
| Lower Teeth | 0.63 |
| Maxilla & Upper Skull | 0.62 |
| Upper Teeth | 0.61 |
| Other | 0.59 |
| Edentulous Region | 0.58 |
| Mandibular Nerve | 0.55 |
Mean Dice: ~0.60 across 167 manually annotated CBCT scans.
Python PyTorch MONAI SimpleITK nibabel 3D Slicer
DICOM → NIfTI conversion
→ Preprocessing (resampling, normalization, augmentation)
→ 3D U-Net training (500 epochs, Dice + CE loss, Adam lr=1e-4)
→ Prediction export → 3D Slicer visualization
167 CBCT scans with manual segmentation masks annotated in 3D Slicer.
Dataset not publicly available due to patient privacy constraints.
- Transformer-based architectures (Swin-UNETR)
- Larger, more diverse dataset
- Bone density estimation for implant planning