Representation learning
Can a self-taught CT model find cancer markers with little data?
Open access · cc by · source: Europe PMC
A CT image model pretrained without labels produced features that, with just a simple linear classifier, matched or beat fully supervised deep networks and held up best when training data were cut to a tenth.
Study at a glance
- Design
- Computational / modelling — Self-supervised contrastive pretraining of a convolutional encoder on CT lesions, then linear-probe versus fine-tuned evaluation against baselines on lesion-site, nodule-malignancy and lung-cancer survival tasks, with reduced-data ablations.
- N
- No single N: pretraining used 11,467 lesions from 2,312 patients; downstream test sets were 1,221 lesions (use case 1), 170 nodules (use case 2) and two survival cohorts of 420 and 133 patients (use case 3).
- Population
- Retrospective CT scans from five datasets (DeepLesion, LUNA16, HarvardRT, LUNG1, RADIO)
- Outcome
- Balanced accuracy, mean average precision and AUC; survival stratification; test-retest stability
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
Their modified SimCLR beat the other pretraining methods, with a balanced accuracy of 0.779 on the site task versus 0.696 for standard SimCLR. The fine-tuned model reached an AUC of 0.944 for nodule malignancy, and linear classification of frozen features was the most robust option when training data were reduced, beating all baselines at every reduced size. For survival the gains were modest (AUC 0.638 in the LUNG1 cohort), but frozen features gave the best risk stratification and were very stable on repeat scans.
Methodology
The authors pretrained a 3D convolutional encoder on thousands of unlabelled CT lesions using a modified version of the SimCLR contrastive learning method. They then used it two ways — as a frozen feature extractor with a linear classifier, or fine-tuned end to end — on three tasks: classifying a lesion's body site, predicting whether a lung nodule is malignant, and predicting 2-year survival in non-small cell lung cancer. They compared against supervised training from scratch and public pretrained medical models, repeated the tasks with 50%, 20% and 10% of training data, and tested stability, saliency maps and links to gene expression.
Limitations
All data are retrospective, so the study cannot show that these biomarkers help patients in real clinical use. Survival prediction was only moderately accurate, and the small RADIO cohort made most model differences non-significant. The authors note that performance across demographic groups and under distribution shift was not tested, that saliency maps have known technical limits, and that the gene-expression link is preliminary.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Pretrained representations shine in the low-label regime.
Pretraining helped most when labels were scarce: frozen CT foundation-model features beat all baselines at every reduced training size, and a medical multimodal model with 10% of labels beat prior models trained on all labels.
Evidence for the claim as stated.
Pretrained embeddings helped for CT imaging, protein and metagenome tasks but hurt for bulk transcriptomic phenotype prediction with a few hundred labelled samples.
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Pretrained embeddings helped for CT imaging, protein and metagenome tasks but hurt for bulk transcriptomic phenotype prediction with a few hundred labelled samples.
Related papers in this topic
Same topic cluster — not a recommendation engine.
- Can a small protein language model match much bigger ones?
- Can a transformer predict which drugs bind which proteins?
- Can one language model read protein sequence and structure?
- Can random negative pairs teach a better genome-sorting model?
- Can a network trained on everyday photos search pathology slides?