MammoWeave: Fidelity-First Cohort Cartography for Auditing Mammography Data and Automated Labels

Abstract

MammoWeave is a provenance-preserving visual-analytics workbench for auditing mammography datasets and automated labels before model development. We introduce three coupled design contributions: an evidence-role grammar that keeps source pixels, acquisition metadata, reference labels, automated scores/boxes, and derived audit measurements visually and semantically distinct; a fidelity-first cohort-to-pixel workflow that connects population context, interpretable quality structure, threshold-conditioned score-label discordance, and paired multi-resolution evidence; and a Casebook handoff that retains source, transformation, uncertainty, and review status for selected findings. These ideas are operationalized as a five-stage process---Survey, Stratify, Contradict, Inspect, and Record---with coordinated selection, alternative encodings, a review-threshold control, and paired-image inspection. A walkthrough on the supplied Challenge 2 archives shows how aggregate signals can be traced back to source products while exposing metadata drift, sparse standard-view context, qualified score-label discordance, and cross-resolution coordinate incongruence. The prototype intentionally bounds interpretation and interaction, but the visual contribution is the persistent evidence structure that prevents derived findings from silently becoming clinical or annotation verdicts.