AI and the World's Hard Problems · Medicine & drug discovery · Nature (computational biology) · 2026-10-23
Two-stage synthetic-to-real transfer learning for automated mammography report generation using vision-language models
The paper addresses the scarcity of paired mammogram-report data by pretraining MedGemma-4B on synthetic reports built from BI-RADS annotations, then fine-tuning on 407 real reports. A preregistered, controlled evaluation found that the apparent gain came from unequal adapter capacity; at matched capacity, the two training schemes performed equivalently. Diagnostic agreement never exceeded κ = 0.14, suggesting that diagnostic accuracy, not fluency, is the main obstacle to clinically useful reporting.