Chronological Contrastive Learning: Few-Shot Progression Assessment in Irreversible Diseases

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.

Figures (6)

• Figure 1: Chronological contrastive learning objective illustrated using a case of monotonically worsening joint-space narrowing (JSN) in a patient’s interphalangeal (IP). Bottom: Anti-/chronological contrastive terms. The loss aligns disease trajectories in latent space, capturing severity automatically. Top right: Training stages. In stage 1, no labels beyond timestamps and patient+ROI IDs are required. In stage 2, the model is fine-tuned for score prediction.
• Figure 2: Left: Joint-level contributions to the total SvHS illustrated on a representative hand radiograph, highlighting erosions and joint spaces. Right: Regions of interest extracted during fully automatic preprocessing of hand radiographs.
• Figure 3: ICC of standard of reference and estimated SvHS as a function of training set size: (left) $SvHS$, and (right) change $\Delta SvHS$; blue: only single-stage baseline, green: pre-trained with reconstruction loss; orange: pre-trained with ChronoCon and reconstruction loss. Black cross: Pretrained with original Rank-N-Contrastive loss on time; : pre-trained with SimCLR. Error bars indicate 95% CI.
• Figure 4: Scatter plots comparing ground truth and model predictions for the single-stage baseline (blue) and our ChronoCon method ($L^{\mathrm{ChronoCon}}$ + $L_2^{\mathrm{DAE}}$; orange) trained on labels from 5, 31, and 466 patients. $L_2^{\mathrm{DAE}}$ only results are shown in green. Top: Longitudinal evaluation in terms of score differences between visits. Bottom: Cross-sectional prediction performance for total SvH scores.
• Figure 5: Feature space (PCA) of the unsupervised model pretrained with ChronoCon. Left (a): Colored by relative time $t_{\mathrm{rel}}$ (0 = first visit, 1 = last); white lines show example patient–joint trajectories. Middle (b): Same embedding colored by ground-truth scores (no score information was used during training). Right (c): Feature similarity between chronologically ordered visits compared to the corresponding joint-space–narrowing (JSN) score differences for the MCPV joint.