GapDream: Active Synthetic Evidence Acquisition for Long-Tailed Medical Imaging

Abstract

Synthetic medical data are commonly generated from manually chosen prompts, inverse frequency sampling, or fixed subgroup balancing rules. These strategies increase sample count but do not answer a more consequential question: which missing observations would most improve the current model? We introduce GAP DREAM, a closed-loop framework that treats a medical image generator as a queryable source of candidate evidence rather than an unlimited source of augmentation. At each round, GAP DREAM maps model failures into an evidence gap score that combines scarcity, predictive uncertainty, local representation density, and gradient novelty. A query policy translates high-value gaps into structured clinical requests, a modality specific generator proposes candidate imagetext pairs, and an independent selector retains only diagnostically consistent, diverse, non-duplicative candidates. The downstream model is then updated and the gap map recomputed. The central objective is therefore not image realism in isolation, but reduction of measurable evidence gaps under a fixed synthetic budget. Across a multi-modality protocol spanning long-tailed chest radiography, dermoscopy, clinical dermatology, and histopathology, GAPDREAM reaches stronger held-out real-data performance with substantially fewer generated samples than random,class-balanced, uncertainty-only, and static targeted augmentation. Results show gains in CXRLT tail mAP, HAM10000 macroAUROC, Came lyon17 out-of-domain accuracy, and worst group dermatology accuracy while reducing redundant synthetic samples. GAPDREAM reframes medical generation as active evidence acquisition: generate the examples the learner needs, not simply more examples of what is rare.