CLEAR: Auditable Chest X-ray AI Uses Clinical Concepts for Zero-Shot Detection

Trained on over 0.87 million image-report pairs and externally validated across four large datasets from the US, Europe, and Asia, CLEAR achieves state-of-the-art classification performance and provides auditable zero-shot pathology detection by decomposing predictions into clin…

• CLEAR represents chest X-rays via embeddings from large language models, enabling decomposition of predictions into weighted contributions from specific radiological observations.
• External validation on physician-annotated datasets from multiple continents demonstrated state-of-the-art performance and identified confounders.
• The framework supports auditable zero-shot pathology detection and creation of concept bottleneck models.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.