Five conditions to fill the gap between medical 3D vision and the clinic
Shared Data · Reliable Supervision · Robust Models · Actionable Outputs · Trusted Evaluation
Shared cohorts, controlled evaluation and standardized evidence — the ToothFairy challenge series and what it taught us about benchmarking.
IEEE Transactions on Medical Imaging, Dec 2024
The first ToothFairy challenge: inferior alveolar canal segmentation in CBCT.
Medical Image Analysis, April 2026
ToothFairy2: multi-structure segmentation, 42 classes.
Medical Image Computing and Computer Assisted Intervention – MICCAI 2026, May 2026
ToothFairy3: scaling to 77 classes with U-Mamba2.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Mar 2025
The nnU-Net lesson: once the benchmark is controlled, the pipeline matters as much as the backbone.
Challenge portal — ToothFairy, ToothFairy2 and ToothFairy3 (ToothFairy4 runs as Task 1 of ODIN 2026)
Code, evaluation scripts and baselines for the whole series
Where our datasets are versioned, documented and shared
Dense 3D masks are the exception; sparse labels are what clinics already produce. Propagating them, and encoding the annotation protocol in the tooling.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2022
Deep label propagation: turning sparse annotations into dense supervision.
European Conference on Computer Vision (ECCV), June 2026
Dataset, platform and baselines for intraoral dental data — the work behind Yggdrasil.
Patient-centric ingestion, joint multi-modal visualization, protocol encoded in the platform
Missing modalities, domain shift and a task that keeps changing after deployment — three ways the input assumption breaks in practice.
Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, May 2025
IM-Fuse: graceful degradation when MRI sequences are missing.
Proceedings of the British Machine Vision Conference, August 2026
Cond-UNet: a lightweight conditioned U-Net that beats a foundation model on multicenter testicular ultrasound.
Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, May 2025
U-Net Transplant: add a structure by merging specialists instead of retraining.
Segmentation describes anatomy; reports communicate clinical meaning. ToothFairy4 moves the benchmark from masks to structured findings.
Medical Image Computing and Computer Assisted Intervention – MICCAI 2026, June 2026
Ontology-grounded structured prediction: 893 reports over 529 public CBCT volumes, a clinician-designed OWL ontology and six stage-aware metrics.
Task 1 of the ODIN 2026 challenge on Grand Challenge — the benchmark moves from masks to structured findings
Lexical metrics score how a report is phrased, not whether it is true — and a published Dice does not tell a hospital which model to deploy.
European Conference on Computer Vision (ECCV), June 2026
Ranking segmentation models with no labels, no source data and no weights — only the predictions.
None of this is single-author work. Code for the papers above lives in our group GitHub organization.
Our group GitHub organization — code for the works above and more
AImageLab, University of Modena and Reggio Emilia
Something missing or a broken link? Let me know. A complete list of publications is available on the publications page.