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

Clemens Watzenböck; Daniel Aletaha; Michaël Deman; Thomas Deimel; Jana Eder; Ivana Janickova; Robert Janiczek; Peter Mandl

cs.CVMar 23, 2026

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

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

Organizations: Computational Imaging Research Lab, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria · Comprehensive Center for Artificial Intelligence in Medicine, Medical University of Vienna, Vienna, Austria · Division of Rheumatology, Department of Medicine III, Medical University of Vienna, Vienna, Austria · Johnson & Johnson · Ludwig Boltzmann Institute of Arthritis and Rehabilitation, Vienna, Austria · Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, Massachusetts, USA

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 & tables

Figure 3

Dataset size				Cross sect. (SvHS)		Longit. ( $\Delta$ SvHS)
scores [%]	scores [ $N$ ]	images	patients	RMSE $\downarrow$	ICC $\uparrow$	RMSE $\downarrow$	ICC $\uparrow$
1	2 442	82	5	$19.9^{{\color[rgb]{0,0.3906,0}\;\mathbf{-7.0}}}_{\;\pm 2.5}$	$86^{{\color[rgb]{0,0.3906,0}\;\mathbf{+17}}}_{\;\pm 2}$	$9.5^{{\color[rgb]{0,0.3906,0}\;\mathbf{-1.9}}}_{\;\pm 0.7}$	$64^{{\color[rgb]{0,0.3906,0}\;\mathbf{+30}}}_{\;\pm 3}$
2	4 475	152	10	$19.3^{{\color[rgb]{0,0.3906,0}\;\mathbf{-4.5}}}_{\;\pm 2.6}$	$86^{{\color[rgb]{0,0.3906,0}\;\mathbf{+10}}}_{\;\pm 2}$	$9.0^{{\color[rgb]{0,0.3906,0}\;\mathbf{-2.1}}}_{\;\pm 0.6}$	$63^{{\color[rgb]{0,0.3906,0}\;\mathbf{+24}}}_{\;\pm 3}$
4	9 413	319	20	$18.7^{{\color[rgb]{0,0.3906,0}\;\mathbf{-4.7}}}_{\;\pm 2.7}$	$87^{{\color[rgb]{0,0.3906,0}\;\mathbf{+7}}}_{\;\pm 2}$	$8.1^{{\color[rgb]{0,0.3906,0}\;\mathbf{-2.3}}}_{\;\pm 0.5}$	$67^{{\color[rgb]{0,0.3906,0}\;\mathbf{+19}}}_{\;\pm 3}$
6	14 719	499	31	$17.2^{{\color[rgb]{0,0.3906,0}\;\mathbf{-2.1}}}_{\;\pm 2.7}$	$90^{{\color[rgb]{0,0.3906,0}\;\mathbf{+3}}}_{\;\pm 2}$	$7.7^{{\color[rgb]{0,0.3906,0}\;\mathbf{-1.4}}}_{\;\pm 0.6}$	$72^{{\color[rgb]{0,0.3906,0}\;\mathbf{+11}}}_{\;\pm 2}$
10	23 546	808	46	$14.5^{{\color[rgb]{0,0.3906,0}\;\mathbf{-2.9}}}_{\;\pm 2.1}$	$93^{{\color[rgb]{0,0.3906,0}\;\mathbf{+3}}}_{\;\pm 1}$	$7.8^{{\color[rgb]{0,0.3906,0}\;\mathbf{-1.1}}}_{\;\pm 0.5}$	$72^{{\color[rgb]{0,0.3906,0}\;\mathbf{+10}}}_{\;\pm 2}$

Table 1: Test-set performance of our model after pretraining with ChronoCon loss on 466 patients and fine-tuning on a fraction of labeled data. Superscripts in green/red show improvement over the single-stage baseline ; subscripts give half the 95% CI.

Table 2: Different pre-training strategies on the full training set of 466 patients. Values are shown with 95% CI. Underlined metrics indicate the best methods that do not require labels during pretraining; bold values indicate the best overall performance. Except for the baseline, all models were trained in two stages. $\star$ / $\ast$ indicate cross-sectional / longitudinal results of a two-sided paired t -test on MSE. Note: p -values are reported for the individual paired comparisons described in the text; see Appendix for details and interpretation.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

	Score distribution (counts)
Type	0	1	2	3	4	5	NS
ERO hands	160 000	11 838	3 657	1 603	395	412	2 613
ERO feet	66 581	8 733	2 472	980	221	424	2 177
JSN hands	63 165	24 048	7 043	5 248	3 093	–	1 548
JSN feet	24 279	9 221	2 338	2 225	1 640	–	1 091

Appendix

Table 3: Erosion and joint-space narrowing scores on joint-(part) level. NS = Not scoreable (e.g. surgical spacers, fused joints, missing fingers, …).

Abbreviation	Meaning
RA	Rheumatoid Arthritis
SvH / SvHS	Sharp–van der Heijde Score
$\Delta$ SvHS	Change in SvH score between visits (in chronological order)
ERO	Erosion
JSN	Joint Space Narrowing
ROI	Region of Interest

Appendix

Table 4: List of abbreviations beyond used throughout the manuscript and appendix.

Abbreviation	Description
Hand joints
PIPII–PIPV	Proximal interphalangeal joints II–V (JSN)
PIPIIED/EP … VED/EP	PIP joints II–V, distal/proximal erosion
MCPI–MCPV	Metacarpophalangeal joints I–V (JSN)
MCPIED/EP … VED/EP	MCP joints I–V, distal/proximal erosion
IPIED / IPIEP	Thumb IP joint distal/proximal erosion

Appendix

Table 5: Grouped joint and score abbreviations used in this work contributing to Sharp-van der Heijde score . Erosion applies separately to proximal (EP) and distal (ED) joint parts.

Figure 11

Explore similar work

Gradient based Severity Labeling for Biomarker Classification in OCT
Feb 23, 2026Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib +2Classification Pseudo-Labeling
OLIVES at the Center for Signal and Information Processing CSIP, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA · Retina Consultants Texas, Retina Consultants of America, Houston, Texas, USA
Temporal Inversion for Learning Interval Change in Chest X-Rays
Apr 6, 2026Hanbin Ko, Kyungmin Jeon, Doowoong Choi +1Longitudinal Medical Image Analysis Medical VLMs
Interdisciplinary Program in Bioengineering, Seoul National University Graduate School · Integrated Major in Innovative Medical Science, Seoul National University Graduate School · Dept. of Radiology, Seoul National University Hospital +2
Self-supervised Pre-training Helps Retinal Disease Progression Modelling Most When Data Is Scarce
Sep 11, 2026Ifeoma Veronica Nwabufo, Julius Gervelmeyer, Sarah Müller +1Self-Supervised Pre-Training Longitudinal Medical Image Analysis
Hertie Institute for AI in Brain Health, Faculty of Medicine, University of Tübingen, Germany · Tübingen AI Center, University of Tübingen, Germany

Feb 23, 2026J/K move · Enter open · S save

Gradient based Severity Labeling for Biomarker Classification in OCT

Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib, Stephanie Trejo Corona+1

OLIVES at the Center for Signal and Information Processing CSIP, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA · Retina Consultants Texas, Retina Consultants of America, Houston, Texas, USA

Classification Pseudo-Labeling Optical Coherence Tomography Supervised Contrastive Learning

Open paper AskPDF arXiv

In this paper, we propose a novel selection strategy for contrastive learning for medical images. On natural images, contrastive learning uses augmentations to select positive and negative pairs for the contrastive loss. However, in the medical domain, arbitrary augmentations have the potential to distort small localized regions that contain the biomarkers we are interested in detecting. A more intuitive approach is to select samples with similar disease severity characteristics, since these samples are more likely to have similar structures related to the progression of a disease. To enable this, we introduce a method that generates disease severity labels for unlabeled OCT scans on the basis of gradient responses from an anomaly detection algorithm. These labels are used to train a supervised contrastive learning setup to improve biomarker classification accuracy by as much as 6% above self-supervised baselines for key indicators of Diabetic Retinopathy.

Abstract

Explore similar work

Gradient based Severity Labeling for Biomarker Classification in OCT

Temporal Inversion for Learning Interval Change in Chest X-Rays

Self-supervised Pre-training Helps Retinal Disease Progression Modelling Most When Data Is Scarce