Self-supervised pre-training improves eye disease models with little data
Self-supervised Pre-training Helps Retinal Disease Progression Modelling Most When Data Is Scarce
Computer Vision and Pattern Recognition
Summary
It's hard to study how eye diseases like age-related macular degeneration (AMD) get worse over time because the right kinds of patient images are rare. The authors found that training a computer program first on many single images taken at one time (instead of changes over time) can help it learn useful information. When there isn’t much data showing disease progression, this pre-training helps the model predict AMD worsening better than starting from scratch. This improvement depends more on how the model is trained beforehand, rather than on the size or similarity of the training images.
What this means in practice
- •For medical imaging teams: Use frozen self-supervised encoders with small survival models to predict retinal disease progression from limited longitudinal data.
- •For clinical software developers: Incorporate pre-trained encoders trained on abundant single-time-point images to improve disease progression tools when patient follow-up data is scarce.$Commercial implications: Enables more accurate prognostic tools for AMD based on limited follow-up images, which can be marketed to eye care providers.
Tested on one dataset.
Authors
Ifeoma Veronica Nwabufo, Julius Gervelmeyer, Sarah Müller, Philipp Berens
Abstract
Modelling how a disease progresses over time requires longitudinal imaging cohorts, which are scarce and small, whereas cross-sectional data -- one image per participant -- is abundant. Self-supervised pre-training on such data offers a way to bridge this gap, but it is unclear which strategy best supports progression modelling, or how that answer depends on the amount of labelled longitudinal data. We study this for age-related macular degeneration (AMD), pre-training encoders on the large cross-sectional NAKO cohort and predicting time to late AMD on the longitudinal AREDS dataset. We compare in-house self-supervised encoders against a general-purpose (DINOv2) and a domain-specific (RETFound) foundation model, across contrastive, masked-autoencoding, and self-distillation objectives, under frozen and fine-tuned protocols, and across labelled training sets from 100 to 32,250 examples. Which model performs best depends on how the encoder is used. When the encoder is frozen and labels are few -- the regime typical of longitudinal cohorts -- pre-trained representations reach clinically reasonable discrimination from a few hundred labelled samples, while models trained from scratch do not; this advantage fades under fine-tuning. Transfer is governed by the self-supervision objective rather than corpus scale or domain match, so that an encoder pre-trained on a modest cross-sectional cohort matches or exceeds a far larger in-domain foundation model. Together, these results offer a practical recipe for building progression models where longitudinal data is scarce: a frozen self-supervised encoder with a lightweight survival head.