Abstract
Screening mammography is read comparatively: radiologists hold one breast against the other, and several finding categories have no meaning outside that comparison. Most automated pipelines nonetheless encode a single breast at a time. We describe Bilateral MV-CLIP, a contrastive image-text framework in which all four projections of a patient enter one forward pass and the two lateralities exchange information through a Cross-Lateral Attention block, with a learned gate and a placeholder token preserving behaviour when a projection or an entire side is unavailable. Training combines image-text and image-image objectives, graded similarity targets, and a symmetry term tying embedding geometry to recorded asymmetry status. We train on VinDr-Mammo with reports synthesised from structured annotations and evaluate by linear probing of frozen features, cross-modal retrieval and report decoding, against an independently trained single-breast model matched in backbone, optimiser, schedule and epoch budget, with checkpoints selected on a held-out validation partition. On a pre-specified composite endpoint pooling the three comparison-defined categories, fusion did not improve on that baseline: area under the curve 0.781 against 0.792, a difference of -1.1 points with a 95% confidence interval from -9.6 to +7.5. Per-category effects follow the direction the design predicts, but the partition holds 49 positive comparison-defined breasts and no difference survived correction. Retrieval is limited by weak instance-level image-text alignment rather than by report redundancy, with a pool-size-independent ranking accuracy of 0.691. On an external cohort of 11,913 patients labelled by biopsy the representations transfer at reduced discrimination, reaching 0.65 for any abnormality. We therefore report explicit bilateral comparison as plausible but, on this dataset, unconfirmed, and give confidence intervals throughout.