CoNNS: Concept-Guided Noisy Negative Suppression for Chest X-Ray Zero-Shot Learning

A new framework called CoNNS (Concept-Guided Noisy-Negative Suppression) has been proposed to improve zero-shot classification and grounding of chest X-ray findings. The method addresses a key limitation in standard contrastive learning for vision-language alignment: treating radiographs and reports from different patients as negative pairs introduces noise because different patients often share similar findings. CoNNS uses a hierarchical concept ontology built with large language models, structuring 41 key clinical concepts by modeling presence, attributes (location and characteristics), and evidential text segments. This ontology supports a negative suppression mechanism that reduces semantic ambiguity, enhancing performance in zero-shot understanding tasks. The approach differs from previous methods that rely on raw reports or templatized texts. The work is published on arXiv under identifier 2605.19374.