Ontology-Grounded Structured Prediction for Dental CBCT Reporting

Code for the MICCAI paper "Ontology-Grounded Structured Prediction for Dental CBCT Reporting".

Contributions

• A dental CBCT ontology for structured report labels: ontology/ontology.ttl.
• A conversion pipeline from raw Italian reports to patient-level TTL labels.
• An ontology-driven prediction model that adapts its heads to the ontology schema.

Steps

1. Put raw reports in data/ToothFairy3/reportsTr/.

2. Convert reports to structured TTL with the main script, ontology/extract.py:

   OPENAI_API_KEY=... python ontology/extract.py

   This combines all reports for each patient into one TTL. ontology/extract_single.py is only for one-report-per-TTL inter-agreement checks.

3. Check outputs:

   • valid TTLs: data/ttl_reports/
   • generated single-report TTLs: data/ttl_reports_single/
   • SHACL/API errors: data/shacl_errors/
   • schema files: ontology/ontology.ttl, ontology/shapes.ttl

4. Preprocess CBCT volumes into regional tokens:

   python main.py preprocess --dataset-root data/ToothFairy3 --output-dir data/preprocessed_tokens

5. Train:

   python -m src.train --wandb_name experiment_name

   For a quick local run:

   python -m src.train --use_kfold false --max_epochs 5 --no_wandb

Notes

• TTL filenames must match preprocessed token patient IDs.
• Training only uses samples with both .pt tokens and .ttl labels.
• Inspect label distribution before training:

  python -m src.scripts.histogram_ttl