ToothFairy4 Dataset
This page describes the ToothFairy4 dataset, the latest extension of the ToothFairy series developed at the University of Modena and Reggio Emilia with the collaboration of Radboud University and the Digital Research Center of Sfax. The dataset is part of the ODIN challenges included in the ODIN2026 workshop within MICCAI2026 which will take place in Strasbourg, France, from 27 September to 1 October 2026. For challenge details, visit the dedicated website or check the structured submission document.
Challenge Focus
Clinical Reasoning: systems must evaluate dental status, bone quality and quantity, and anatomical risks to generate structured reports for procedures such as implant placement, extractions, and sinus lifts.
Multimodal Integration: although this track input is CBCT, the core technical objective is to convert high-dimensional 3D medical imaging into coherent clinical documentation.
Generalization: the benchmark is designed to assess robustness across centers and acquisition protocols, with a hidden private test set from an external institution.
Dataset Structure
| Field | Value |
|---|---|
| Total CBCT cases | 625 |
| Set P | 417 volumes |
| Set F | 63 volumes |
| Set S | 52 volumes |
| Set A | 100 volumes |
| CBCT files per case | 1 volume (.nii.gz) |
| Reports (IT and EN) | 1001 total for each language |
| Reports per case (IT and EN) | 250 cases with 1 report, 371 with 2 reports, 3 with 3 reports |
Each case is organized as:
cbct/: one CBCT volume (NIfTI, compressed as.nii.gz),reports_it/: original clinician-authored report(s) in Italian,reports_en/: English-translated report(s).
The dataset is organized into four subsets: P, F, S, and A. Sets P, F, and S come from ToothFairy3, while set A is the new subset contributed by the Digital Research Center of Sfax, acquired with different CBCT machines.
Clinical Context
ToothFairy4 is designed around real reporting workflows in implantology and oral-maxillofacial surgery, where clinicians must summarize dentition status, bone characteristics, and risk anatomy into concise planning notes.
The benchmark focuses on generating clinically useful draft reports from CBCT, with emphasis on factual completeness and consistency across centers, so models can support safer and more standardized surgical planning.
Translation Pipeline
Original reports are preserved, and English versions are produced with an LLM-based translation pipeline. The current translation script uses gpt-5.2. To mitigate translation errors, random subsets of training reports are quality checked by clinicians, while test-set translations are clinically reviewed.
System prompt used for report translation:
You are an expert bilingual medical translator specializing in dental radiology and CBCT reporting. Task: translate an Italian clinical dental/CBCT report into English. Rules: 1. Preserve meaning with maximal clinical fidelity. Do not add, remove, soften, or overstate findings. 2. Use natural, native-level medical English as written by a radiologist fluent in Italian. 3. Keep the narrative form and sentence flow of the source report (no bullet points, no restructuring into templates). 4. Preserve all clinical qualifiers and uncertainty language faithfully (e.g., possible, probable, compatible with). 5. Preserve tooth numbering, side labels, measurements, anatomic references, and abbreviations exactly when clinically appropriate. 6. Keep register professional and concise. 7. Return only the final English translation text, with no preface or commentary.
