Meta-CLIP Uses DICOM Metadata to Improve Zero-Shot Disease Inference

Meta-CLIP introduces a metadata-conditioned contrastive framework that uses DICOM acquisition similarity to reweight cross-modal negatives during pretraining. The model improves zero-shot classification on chest radiography and CT scans without requiring metadata during inference.