Two-stage synthetic-to-real transfer learning for automated mammography report generation using vision-language models

Gani Esen, Marat Nurtas, Jandos Amankulov, Laura La Paglia

Scientific Reports, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1038/s41598-026-76357-4
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Abstract We present a two-stage synthetic-to-real training pipeline for automated mammography report generation, together with a controlled, preregistered evaluation of whether its synthetic pretraining stage contributes to final report quality. Automated mammography report generation is constrained by the scarcity of paired mammogram-report data: public collections such as VinDr-Mammo and CBIS-DDSM provide structured annotations but no free-text reports. The pipeline converts BI-RADS annotations into synthetic reports (400 generated, 390 retained) to pretrain MedGemma-4B with multimodal LoRA, then fine-tunes on 407 real DMID reports. On the leak-free DMID test subset ( $$n = 49$$ ) it reaches ROUGE-L 0.656, METEOR 0.666, CIDEr 0.649 and BERTScore 0.917, exceeding every external baseline, including the published state of the art. The evaluation then shows that synthetic pretraining does not produce these numbers: the apparent advantage was measured between models of unequal adapter capacity, and at matched capacity the two schemes are equivalent. Across 60 training runs over three corpus sizes, three capacities and five initialisations, against a margin fixed in advance, no configuration favours two-stage training; capacity alone reproduces the gap. Two findings bear on evaluation. Under an assertion-level criterion no model fabricates findings on normal studies; the failure mode is omission. Textual and clinical quality diverge within one model: where two-stage training gains 0.058 ROUGE-L, BI-RADS agreement and sensitivity both fall. With agreement never exceeding $$\kappa = 0.14$$ , the obstacle to clinically useful reporting is diagnostic accuracy, not fluency.