Meta-CLIP introduces DICOM acquisition metadata into the contrastive pretraining of medical image‑text models. A kernel over numerical and categorical fields defines a bounded pairwise margin that grows with metadata similarity, thereby re‑weighting cross‑modal negatives that share similar scanning conditions. The kernel stays fixed during image‑text optimization, and inference uses the usual dual‑encoder without requiring metadata. We evaluate the method on chest radiography and volumetric chest CT through zero‑shot classification, appearance perturbations, grounding, and representation analysis. On the CT‑RATE benchmark Meta‑CLIP improves macro‑AUROC over CT‑CLIP by 1.41 points with a 95% paired bootstrap confidence interval of 0.96–1.85, and achieves the highest macro‑AUROC on four of five CXR datasets. An exploratory BrgSA study extends the analysis to another CT backbone, confirming acquisition metadata as a practical training‑time prior. Code will be released upon publication.
Review