2026 · Published
Research
Generalizable Medical Anomaly Detection
Translation-, reconstruction-, and uncertainty-based methods for detecting abnormalities across brain and spinal MRI.
Clinical motivation
Rare and heterogeneous lesions make exhaustive annotation difficult, while acquisition changes can produce false anomaly signals.
Technical challenges
- Rare and heterogeneous lesions make exhaustive pixel-level annotation difficult.
- Reconstruction can reproduce an anomaly or generate errors in complex normal anatomy.
- MRI sequences and available inputs differ across datasets and clinical settings.
Research progression
- Cross-sequence translation with disentangled representations in MultiTransAD
- Dual-uncertainty guidance for spinal-cord T2 hyperintensity detection in U2AD
- Unified cyclic translation and dual-level evidence across brain diseases and sequences in UniTransAD
Current directions
- Generalization under sequence and acquisition changes
- Combining pixel-level and feature-level anomaly evidence
- Evaluation protocols for heterogeneous public datasets
Related publications
Representative publications
2026 · Published
U2AD: Uncertainty-based unsupervised anomaly detection framework for detecting T2 hyperintensity in MRI spinal cord
2025 · Published