Learning with Label Structure and Uncertainty
Getting more out of imperfect labels, from coarse annotations to class taxonomies to RF waveforms.
Problem. Real-world labels are coarse, noisy, and structured; standard one-hot training throws that structure away.
Findings across four works:
- LEARNER (ISBI 2026): contrastive learning recovers granular labels from coarse ones, cutting annotation requirements for fine-grained medical tasks.
- Label uncertainty for ultrasound segmentation (ISBI 2026): modeling annotator uncertainty directly improves segmentation quality.
- LCA-on-the-Line (ICML 2024): class-taxonomy distance (LCA) predicts out-of-distribution generalization, turning label hierarchies into a model-selection signal.
- W-Net (Medical Image Analysis 2022) and confidence labels (arXiv 2021): incorporating ultrasound RF waveform data and probabilistic class-similarity labels improves dense segmentation of subcutaneous and breast tissue; both are patented (US20240177000A1, US20240177445A1).
Why it matters. Label structure is free supervision. Exploiting it improves accuracy, calibration, and out-of-distribution robustness without new data collection.
See the publications page for the papers.