CLEAR maps chest X-rays to 368,294 clinical concepts and beats CheXzero on external tests
bravo_abad · x · 2026-07-22
CLEAR: a chest X-ray foundation model grounded in 368,294 clinical concepts
Tianyu Han and coauthors introduce CLEAR, a radiology foundation model that predicts through a very large concept bank mined from free-text reports.
- They extract 368,294 radiological observations from 227,835 reports using an LLM, instead of relying on a small hand-written concept list.
- A DINOv2 image encoder and a text encoder are aligned with CLIP-style contrastive learning on 0.87 million image-report pairs from 239,391 patients.
- At inference, each X-ray is mapped to similarities over the concept bank, then projected into an LLM embedding space; zero-shot classification uses softmax over positive and negative prompts.
- On external datasets from Vietnam, Spain, and the US, CLEAR beats CheXzero, BiomedCLIP, and OpenAI CLIP on zero-shot AUROC.
- Its linear probe averages 87.0% versus 71.8% for BiomedCLIP.
- Three blinded radiologists judged 89.8% of the top concepts as diagnostically relevant, with no dataset artifacts flagged.
The paper also shows the model can be debugged by editing the concept space: when enlarged cardiomediastinum was probed, many top concepts were actually about atelectasis, reflecting label co-occurrence in MIMIC-CXR. Filtering the bank to mediastinal concepts and refitting only the linear head raised AUROC from 0.727 to 0.784 without retraining the encoder or collecting new data.
The broader takeaway is that a large, text-derived concept bank can make vision or omics models more inspectable and easier to patch by changing the feature space instead of retraining from scratch.
More from Research
- Itshi shows a full embodied AI model driving an automotive wiring line at WAIC 2026 — 量子位 · 2026-07-22
- A new RL finding says the training harness may induce generalization — inductionheads · 2026-07-22
- XBOW paper breaks insecure LLM code generation into fixable failure categories — moyix · 2026-07-22
- Gemini 3.6 Flash “disproves” a 43-year-old conjecture in a joke post — georgemillo · 2026-07-22
- AI Security Institute tests lie detectors across 31 open-weight models — geoffreyirving · 2026-07-22
- DeltaNet notes unpack Kimi Delta Attention with equations and a state-update diagram — nrehiew_ · 2026-07-22