Loading…
Chest radiograph and report encoder pretrained contrastively on a million pairs, whose frozen embeddings train classifiers from a few hundred cases.
No providers recorded yet. Browse all providers
Training an AI system to read chest radiographs has conventionally meant paying radiologists to annotate tens of thousands of images for each condition, and the resulting classifier can only ever report the conditions on that list. Every one of those radiographs, meanwhile, already arrived with a free-text report attached — a description a board-certified radiologist wrote as part of routine care, which a label column throws away.
PASTER (Pretraining Approach for Small-sample Training Employing Radiographs) treats that report as the supervision signal. Built at National Defense Medical University in Taipei and published in Journal of Medical Systems in September 2025, it is a dual-encoder model trained with a contrastive objective on 1,004,314 chest radiograph–report pairs from the archives of Tri-Service General Hospital. An image encoder and a text encoder project into a shared 512-dimensional space in which a radiograph is pulled toward its own report and pushed away from everyone else's. Nothing in that objective names a disease, so the representation is not bounded by a fixed label vocabulary.
What pretraining buys is sample efficiency downstream. The encoder stays frozen and adaptation is an elastic-net logistic regression over its embeddings, so a new task needs a few hundred examples rather than a few hundred thousand. PASTER initializes from OpenAI's CLIP weights and reuses the zero-shot text-prompting strategy of CheXzero, but is pretrained entirely on the group's own Taiwanese corpus rather than on another model's checkpoints — distinguishing it from broad biomedical vision-language models such as BiomedCLIP.
The image encoder is a ViT-B/32 operating on 256×256 radiographs divided into 32×32 patches, with 12 transformer layers at hidden size 768; the text encoder is a 12-layer transformer with hidden size 512, 8 attention heads and a 256-token context. A [CLS] token from each branch is projected to a 512-dimensional embedding, and a softmax-normalized contrastive loss over the inner products aligns matched pairs. Pretraining ran for 50 epochs with SGD (batch size 64, learning rate 1e-4, momentum 0.9) on a 90/10 split of the 1,004,314-radiograph corpus. The 256×256 input size follows the primary paper and the released code; the npj Digital Medicine companion study's methods instead describe resizing radiographs to 224×224. In that 2026 study, the same frozen encoder supplied embeddings for 1074 phecodes across four cohorts totalling 443,697 patients; 554 prevalent and 457 incident phenotypes were significantly predicted in all three validation sets, and 60 prevalent and 42 incident phenotypes cleared high-discrimination thresholds (AUC ≥ 0.85, C-index ≥ 0.80) in the held-out and external cohorts and slightly relaxed ones in MIMIC-CXR.
The practical audience is a hospital informatics or radiology research group that has a report archive but no labelling program. Because adaptation is a regression over frozen features, a clinically useful classifier for a new indication can be produced from the few hundred confirmed cases such a group can realistically assemble — the sample scale regulators expect for a single indication. The prompted-feature mechanism extends this to interpretation: scoring a radiograph against 57 radiologist-curated findings such as cardiomegaly, atherosclerosis and ground-glass opacity reconstructs most phenotype predictions with R² ≥ 0.85, letting a clinician see which classical radiographic signs drive a given risk estimate rather than a saliency blob.
PASTER is a demonstration that the label bottleneck in medical imaging is largely self-imposed, and that a single institution's report archive is enough to build a transferable encoder. Its evidence for generalization is unusually direct: coefficients fit on Taiwanese outpatient radiographs held up on a US ICU database with a different case mix and demographics. The honest constraints are equally clear. Pretrained weights are not publicly distributed — the authors offer access under IRB-approved collaboration agreements — so the released code reproduces the method rather than the model. Discrimination degrades on AP-view images, which dominate inpatient and ICU settings, and the authors state that PA-derived performance should not be extrapolated there without further validation. All results to date are retrospective; no prospective or clinical-utility evaluation has been reported.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.