Reproduce “Chronological Contrastive Learning: Few-Shot Progression Assessment in Irreversible Diseases” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Clemens Watzenböck, Daniel Aletaha, Michaël Deman, Thomas Deimel, Jana Eder, Ivana Janickova, Robert Janiczek, Peter Mandl and 4 more

ChronoCon uses the order of repeated radiographic examinations—rather than expert disease scores—to learn representations of rheumatoid arthritis progression. On hand and foot X-rays, the method improved few-shot prediction of Sharp–van der Heijde scores, reaching a cross-sectional ICC of 0.86 and longitudinal ICC of 0.64 with labels from only five patients, while fully supervised label-based pretraining remained strongest when abundant annotations were available.

Key insights

  • Visit order can act as weak supervision for predominantly irreversible diseases without treating the time between visits as a quantitative measure of progression.
  • Chronological representations are especially valuable in low-label settings and provide stronger estimates of score changes between visits than generic autoencoding or augmentation-based contrastive learning.
  • When many expert labels are available, ordinal and label-based contrastive objectives outperform chronology alone, positioning ChronoCon as a label-efficient pretraining strategy rather than a replacement for supervised learning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.