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

• Progression modelling requires repeated imaging, but longitudinal cohorts are scarce and expensive while cross-sectional imaging is abundant.Source: §1 · p. 1
• AMD progression is naturally framed as survival analysis, yet ophthalmology AI has focused more on disease detection.Source: §1 · p. 1
• AREDS followed thousands of eyes for more than a decade, providing a longitudinal cohort for studying AMD progression.Source: §1 · p. 2
• The self-supervision objective, rather than corpus scale or domain match, chiefly explained encoder-transfer differences, concentrated in frozen, low-data evaluation.Source: §1 · p. 2
• The study compares NAKO-pretrained encoders with DINOv2 and RETFound for AREDS time-to-late-AMD prediction across objectives, protocols, and 100 to 32,250 labels.Source: §1 · p. 2

2.1 Datasets

• NAKO fundus images supply cross-sectional pre-training data, while longitudinal AREDS images support survival prediction of late-AMD progression.AREDS observations were split by patient; the analysis treated each eye as an independent observation.Source: §2.1 · pp. 2–3, §2 · p. 2

2.2 Encoder Models & Pre-training

• The comparison includes in-house SimCLR, MAE, and DINOv2 encoders, foundation models RETFound and DINOv2, and ResNet18 baselines.SimCLR uses contrastive learning; MAE uses reconstruction, while the in-house DINOv2 approach adapts the model using LoRA rather than training it from scratch.Source: §2.2 · pp. 4–5

2.3 Survival Modelling

• The Cox proportional hazards model represents risk with a shared baseline hazard and a subject-specific score, with a neural network parameterizing image-based risk.Censoring is retained as informative: it records that the event had not occurred by a subject's last observation.Source: §2.3 · p. 5

2.4 Evaluation

• Models are evaluated for discrimination with Harrell’s C-index and calibration with the inverse-censoring-weighted integrated Brier score over 1–5 years.The C-index measures risk ranking among comparable pairs; lower IBS indicates better calibration.Source: §2.4 · pp. 5–6

2.5 Experimental Details

• Each encoder is frozen or fully fine-tuned and paired with a linear or two-layer MLP survival head across seven training-set sizes from 100 to 32,250 examples.Each configuration was trained with five random seeds and assessed on the same held-out test set.Source: §2.5 · p. 6

2.6 Statistical analysis

• Generalised additive models relate C-index and IBS to training-set size, while nested-model comparisons assess objective, model-family, and pre-training-domain contributions.The analysis also compares predicted performance across objectives as training size changes.Source: §2.6 · pp. 6–7

3 Results

• As training data increased, predicted survival curves became more similar across models; linear readouts generally produced higher, less variable survival probabilities over time.Source: §3 · p. 8

3.3 Minimum training set size for clinically reasonable performance

• Frozen DINOv2 and SimCLR models reached C-index thresholds at n ≤100 absolute and n ≤1,000 relative, while ResNet-Scratch never reached 0.75.Source: §3.3 · pp. 8–9, Table 4 · p. 11

3.1 SimCLR-pretrained models had the best peak performance

• At n = 32,250, SimCLR-ImageNet-NAKO led frozen-encoder MLP results with C-index 0.920 ± 0.001 and IBS 0.054 ± 0.001.Source: §3.1 · p. 8, Table 3 · p. 9

3.3 Minimum training set size for clinically reasonable performance

• Under fine-tuning, every model crossed the absolute threshold by n ≤10,000; SimCLR and MAE crossed fastest (n ≤500), and peak performance ranged from 0.908 to 0.933.Source: §3.3 · p. 9, Table 5 · p. 11

3.4 The pre-training objective matters more than dataset size or domain

• Within objectives, in-house encoders matched or neared larger counterparts: RETFound and MAE-ImageNet-NAKO peaked at 0.870 and 0.876 despite RETFound's roughly threefold image count and more than twice the model size.Source: §3.4 · p. 9

3.2 Pre-training benefits are most pronounced in the low-data regime

• With frozen encoders, many pre-trained models retained high calibration and discrimination with a few hundred examples, while supervised baselines degraded sharply; fine-tuning narrowed differences.Source: §3.2 · p. 8, Figure 3 · p. 10

3.4 The pre-training objective matters more than dataset size or domain

• Adding pre-training objective raised explained deviance from 0.375 to 0.798 for C-Index and 0.640 to 0.901 for IBS; model family and domain added no improvement beyond it.Source: §3.4 · p. 10
• SimCLR stayed strong across frozen and fine-tuned use, whereas DINOv2 was strong frozen but weakest when fine-tuned at small sample sizes, and MAE showed the reverse pattern.Source: §3.4 · pp. 10–11

4 Discussion

• Pre-training helps most in frozen, low-data settings; its advantage largely disappears with fine-tuning as data grows.SimCLR was strong throughout; DINOv2 was strong frozen but initially weak when fine-tuned, while MAE lagged frozen but was stable under fine-tuning.Source: §4 · pp. 11, 13, Figure 4 · p. 12
• Transfer differences are explained mainly by objective, not corpus scale or domain match: modest-cohort MAE matched RETFound, and LoRA adaptation left DINOv2 unchanged.For SimCLR, natural-image initialisation before NAKO pre-training also made almost no difference.Source: §4 · pp. 11–12
• With fine-tuning, a standard ImageNet-initialised ResNet became competitive with every foundation model at far lower computational cost.The authors attribute frozen MAE weakness possibly to pixel-level reconstruction objectives, which fine-tuning can reshape toward useful representations.Source: §4 · p. 13
• For scarce longitudinal labels, the proposed recipe is a frozen self-supervised encoder with a lightweight survival head, reaching clinically reasonable discrimination from a few hundred samples.The authors say this avoids the labelled data, compute, and engineering demands of full fine-tuning and does not require a large or domain-matched pre-training corpus.Source: §4 · p. 13
• Generalisability remains open because the study evaluated only AMD with colour fundus photography; absolute performance is not directly comparable to prior AREDS models.Prior models used different inputs, label definitions, and evaluation protocols; the contribution is comparison across pre-training strategies and data regimes, not a new AREDS state of the art.Source: §4 · p. 13

Data availability

• AREDS access is through dbGaP and NAKO access is by application; neither dataset is redistributed, while derived per-run C-index and IBS values are in the code repository.AREDS accession: phs000001.v3.p1; NAKO access is subject to its data-use and governance requirements.Source: Data availability · p. 14

S1 Methods Extension

• For DINOv2-NAKO, the trainable-parameter count covers LoRA adapters; its ≈86M-parameter ViT-B backbone remains frozen during adaptation.Source: Table S7 · p. 19

S2 Additional Figures

• Figure S5 plots 1–5-year late-AMD progression-free survival from a year-2 fundus image, contrasting frozen models at 1,000/32,250 and fine-tuned models at 1,000/30,000 examples.The grey triangle marks observed conversion time.Source: Figure S5 · p. 20

S3 Additional Tables

• Table S8 reports fine-tuned-encoder, MLP-head performance at n = 32,250 and n = 1,000.Values are mean ± SD across five training-run repetitions.Source: Table S8 · p. 21
• Table S9 reports frozen-encoder, linear-head performance at n = 32,250 and n = 1,000.Values are mean ± SD across five training-run repetitions.Source: Table S9 · p. 21
• Table S10 reports fine-tuned-encoder, linear-head performance at n = 32,250 and n = 1,000.Values are mean ± SD across five training-run repetitions.Source: Table S10 · p. 22

S4 Further Analysis

• Figure S6 compares predicted IBS spread and discrimination across self-supervision objectives, training sizes, encoder protocols, and survival heads.The spread panels distinguish frozen and fine-tuned protocols; discrimination panels cover linear and MLP heads under both protocols.Source: Figure S6 · p. 23
• Table S11 decomposes deviance by regime for the full five-objective dataset, reporting baseline and objective-added deviance explained and ΔAIC improvement.Source: Table S11 · p. 23
• Table S12 gives the same regime-specific deviance decomposition for the three-objective SimCLR/MAE/DINOv2 subset.It reports baseline and objective-added deviance explained and ΔAIC improvement.Source: Table S12 · p. 24