[Paper Review] Chronological Contrastive Learning: Few-Shot Progression Assessment in Irreversible Diseases

Clemens Watzenböck, Daniel Aletaha+10 more|arXiv (Cornell University)|Mar 23, 2026

Rheumatoid Arthritis Research and Therapies0 citations

TL;DR

ChronoCon learns progression-aware representations from longitudinal medical images without expert labels by enforcing chronological ordering in latent space, enabling strong few-shot progression assessment in irreversible diseases like rheumatoid arthritis.

ABSTRACT

Quantitative disease severity scoring in medical imaging is costly, time-consuming, and subject to inter-reader variability. At the same time, clinical archives contain far more longitudinal imaging data than expert-annotated severity scores. Existing self-supervised methods typically ignore this chronological structure. We introduce ChronoCon, a contrastive learning approach that replaces label-based ranking losses with rankings derived solely from the visitation order of a patient's longitudinal scans. Under the clinically plausible assumption of monotonic progression in irreversible diseases, the method learns disease-relevant representations without using any expert labels. This generalizes the idea of Rank-N-Contrast from label distances to temporal ordering. Evaluated on rheumatoid arthritis radiographs for severity assessment, the learned representations substantially improve label efficiency. In low-label settings, ChronoCon significantly outperforms a fully supervised baseline initialized from ImageNet weights. In a few-shot learning experiment, fine-tuning ChronoCon on expert scores from only five patients yields an intraclass correlation coefficient of 86% for severity score prediction. These results demonstrate the potential of chronological contrastive learning to exploit routinely available imaging metadata to reduce annotation requirements in the irreversible disease domain. Code is available at https://github.com/cirmuw/ChronoCon.

Motivation & Objective

• Motivate reducing annotation burden by exploiting longitudinal imaging data for progression assessment in irreversible diseases.
• Propose ChronoCon to learn progression-aware representations using visitation order without expert labels.
• Demonstrate label-efficient learning in RA radiographs and compare to fully supervised baselines.
• Show that ChronoCon improves performance in low-label settings and under few-shot supervision.

Proposed method

• Introduce ChronoCon loss that uses chronological negatives and anti-chronological negatives within the same patient and ROI to enforce progression order in latent space.
• Two-stage learning: Stage 1 pretraining with ChronoCon (and optional reconstruction loss) to learn embeddings; Stage 2 train a multi-headed regressor to map embeddings to ordinal scores.
• Define forward and backward chronological loss contributions to account for monotonic progression and timestamp asymmetry.
• Use double-crop augmentation to enable training on sequences with two visits.
• Stage 2 only uses MSE to predict AO/ERO/JSN scores; total SvH is derived post hoc from subscores.

Experimental results

Research questions

• RQ1 Can chronological visitation order serve as a self-supervision signal to learn disease progression representations without expert labels?
• RQ2 How does ChronoCon perform in low-label and few-shot settings compared with fully supervised baselines and other pretraining strategies?
• RQ3 Do temporal-order based losses improve longitudinal prediction of disease progression (ΔSvHS) in RA radiographs?
• RQ4 Is the learned feature space progression-aware and useful for downstream severity scoring?

Key findings

• ChronoCon with reconstruction loss outperforms the single-stage baseline across label-splits, especially in low-label settings.
• In few-shot scenarios (labels from five patients), ChronoCon achieves an ICC of 0.86 for severity prediction and RMSE of 19.9, outperforming some baselines.
• ChronoCon pretraining yields stronger longitudinal (ΔSvHS) performance and maintains gains across varying training set sizes, indicating learned patient-specific progression.
• When labels are abundant, label-based pretraining (ordinal-constrained) performs best, but ChronoCon offers substantial benefits where labels are scarce.
• Alternative temporal losses (RnC:t) can underperform ChronoCon in low-label settings, highlighting the value of separate forward/backward temporal contributions.