If you have ever trained a model on medical images, you know the uncomfortable truth: the model is rarely the problem. The data is.
Cone-beam CT (CBCT) is a good example. It is widely used in implant planning, orthodontics, and maxillofacial surgery, and it produces rich 3D volumes of teeth, jaw bones, nerve canals, and airways. But a folder full of DICOM files is not a dataset you can train on. In this post, I’ll walk through what makes training-ready CBCT datasets, which annotation types to choose, the problems you will likely hit, and a checklist you can reuse.
I work with the team at Pariedolia System LLp on medical image annotation, so this comes from practical experience with these pipelines.
TL;DR
Decide the clinical task first, then choose the annotation type
Write a versioned annotation guideline and run a pilot before scaling
Measure annotator agreement with real metrics
De-identify metadata and split data by patient
Treat the dataset like code: version it and iterate
What does “training-ready” actually mean?
A CBCT dataset is training-ready when a model can learn from it without being misled by noise, inconsistent labels, or leakage. That means:
Labels follow a written guideline
File formats, voxel spacing, and orientation are consistent
Patient data is fully de-identified
Scanners, protocols, and patient groups are varied
Annotator agreement is measured, not assumed
Train, validation, and test splits do not share patients
Pick the right annotation type
Choose based on the task, not on what is easiest to label.
Annotation type What it labels Common use
Semantic segmentation Every voxel by structure Implant planning, airway analysis, nerve canal
Instance segmentation Each tooth separately Tooth numbering, charting
Landmarks 3D anatomical points Cephalometric measurements
3D bounding boxes Cuboid regions Cysts, impacted teeth, implants
Classification Scan or region tags Pathology presence, bone density
Surface and mesh Smooth 3D surfaces Prosthetics, aligners, surgical guides
Segmentation gives the richest supervision but costs the most time. Bounding boxes are faster when you only need localization.
Problems you will probably hit
Artifacts. Metal fillings, braces, and patient motion create streaks that hide anatomy. Your guideline needs rules for unclear regions and criteria for rejecting scans.
Inconsistent intensities. CBCT grayscale values vary between scanners, so basic thresholding often breaks across devices.
Scanner variability. Voxel size, field of view, and reconstruction settings differ. A model trained on one clinic’s device can fail on another’s.
Annotator disagreement. Two experts may draw the nerve canal differently. Unmanaged, that disagreement becomes label noise.
Class imbalance. Small structures take up a tiny part of the volume, so models learn to ignore them.
Privacy. DICOM headers carry identifiers, and facial features can be visible in reconstructions.
Data leakage. If scans from one patient land in both training and test sets, your metrics look great and your real-world performance does not.
Best practices that work
Define the clinical goal first. Every label class should trace back to a decision the model supports.
Write a detailed annotation guideline. Include definitions, boundary rules, artifact handling, and example images. Version it.
Run a pilot batch. Around 20 scans is enough to expose unclear rules before you scale.
Mix annotators and experts. Trained annotators handle volume, and dental or radiology specialists resolve disputes and set standards.
Measure quality. Use Dice, IoU, and surface distance for segmentation, and Cohen’s kappa for categorical labels.
Double-read critical structures. High-risk labels like the inferior alveolar canal deserve consensus review.
Use pre-labeling carefully. A baseline model can speed things up, but humans must verify every output, or you reinforce the model’s own mistakes.
Standardize formats. Use NIfTI or DICOM-SEG, harmonize orientation and spacing, and document each preprocessing step.
De-identify thoroughly. Clean metadata and confirm compliance with your region’s rules and consent terms.
Split by patient. Never let one patient appear in more than one split.
Version your data. Track releases and annotator history like you would track code.
A quick sanity check in code
Patient-level splitting is the one step people most often get wrong. Here is a minimal Python example using scikit-learn:
from sklearn.model_selection import GroupShuffleSplit
scan_paths: list of CBCT volumes
patient_ids: matching list of patient identifiers
splitter = GroupShuffleSplit(n_splits=1, test_size=0.2, random_state=42)
train_idx, test_idx = next(splitter.split(scan_paths, groups=patient_ids))
train_scans = [scan_paths[i] for i in train_idx]
test_scans = [scan_paths[i] for i in test_idx]
Confirm no patient appears in both sets
assert set(patient_ids[i] for i in train_idx).isdisjoint(
set(patient_ids[i] for i in test_idx)
)
Grouping by patient ID guarantees that all scans from one patient stay together.
The workflow in short
Define the task and label schema
Collect and de-identify diverse scans
Standardize formats and metadata
Write guidelines and run a pilot
Annotate with pre-labeling and expert review
Run quality checks and adjudicate disagreements
Split by patient and freeze a versioned release
Train, evaluate, and trace model errors back to the data
The last step is the one that pays off most. When the model fails, look at the labels and the scan coverage before you touch the architecture.
Final thoughts
Building training-ready CBCT datasets takes clinical insight, process design, and disciplined quality control. Teams that invest in guidelines, measurable agreement, and privacy-safe pipelines get models that generalize better and earn clinical trust sooner.
If you are planning a CBCT project and want help with the data foundation, Pariedolia System LLp offers medical image annotation and dataset preparation. You can read the original version of this article at pareidolia.in.
Top comments (0)