← Back to Home

Making Chest X-Ray Decisions Traceable: CLEAR Breaks Black-Box Predictions Down into Clinical Concepts

This open model not only identifies imaging abnormalities but also attempts to explain the basis for its decisions: it links predictions back to hundreds of thousands of radiological observations, allowing physicians to inspect and correct the signals on which the model relies. But despite progress in transparency for research use, critical gaps remain before clinical safety and regulatory acceptance can be achieved.

By SURL BioNews

Chest X-rays are among the most common imaging examinations in clinical practice, but an artificial intelligence system that gives the right answer may not necessarily be able to explain why. This gap is more than a matter of trust: if a model mistakes lung collapse, patient positioning, or biases in the data source for evidence of cardiac enlargement, it may fail when deployed at another hospital even if its overall accuracy appears acceptable. Published in *Nature Biomedical Engineering*, CLEAR approaches the problem by asking whether the model’s decision pathway can be traced.

CLEAR is an openly accessible foundation model for biomedical imaging. According to its official code repository, it converts each chest X-ray into similarity scores against 368,294 radiological observation concepts, after which a lightweight classifier completes the specific task. As a result, a positive or negative prediction is no longer represented only by a single score: researchers can also break down which concepts pushed the result higher and which pulled it lower.

This extensive concept library is derived from radiology reports and covers semantics including signs of disease, anatomical locations, and imaging descriptions. Official model documentation indicates that CLEAR uses a DINOv2 ViT-B/14 image encoder and a radiology text encoder to place images and text in a shared 768-dimensional space. Pretraining used a total of 873,342 chest X-ray–report pairs from MIMIC-CXR, CheXpert-Plus, and ReXGradient, involving 239,391 patients.

The research team also evaluated the model on four physician-annotated datasets from the United States, Europe, and Asia to examine its performance across data sources. A more concrete benefit emerged during model auditing: for assessments of cardiomediastinal enlargement, the researchers found that the classifier relied on concepts associated with atelectasis. After removing this potentially confounding shortcut and refitting only the lightweight classifier, without retraining the image encoder, the AUROC increased from 0.727 to 0.784. This shows that interpretability may be more than a retrospective display and could also serve as a tool for correcting models.

However, “human-readable concepts” do not necessarily mean “causally reliable.” Because the concept library is derived from existing reports, it may still inherit biases in report wording, hospital workflows, and patient composition. High-weight concepts identified by the model indicate only computational associations and cannot by themselves prove that the model has captured the true pathological mechanism. The four physician-annotated datasets provide cross-regional testing, but the public abstract and model documentation remain insufficient to answer questions about stability in prospective workflows, across different equipment, for rare diseases, and among different patient subgroups.

The research team has released the code, research scripts, model weights, concept vocabulary, and precomputed embedding data, allowing other researchers to reproduce and examine the method. This increases the possibility of external auditing but does not constitute clinical approval: the official repository explicitly states that CLEAR is for research use only, is not a medical device, and has not been validated for diagnostic, treatment, or triage purposes. Before it can enter hospitals, the next step will require more than comparing accuracy. It must also be demonstrated that the conceptual explanations can be interpreted consistently in real clinical settings, can reveal errors promptly, and can provide traceable safety evidence under regulatory requirements.

References

  1. Nature Biomedical Engineering
  2. GitHub
  3. Hugging Face