Summary
A medRxiv preprint describes CARE-MVLM, a mammography vision-language model designed to provide region-grounded evidence and abstain when an answer is not supported. Its authors report stronger lesion-sensitive abstention than the comparison models in an evaluation using 7,358 curated triplets from three public datasets.
A medRxiv preprint describes CARE-MVLM, a vision-language model designed to answer mammography questions with supporting image evidence or abstain when it cannot answer reliably. The researchers report improved lesion-sensitive abstention compared with other systems in an evaluation using 7,358 curated triplets from three public mammography datasets.
Separating answers from evidence and abstention
CARE-MVLM is built on Qwen2.5-VL-7B-Instruct. Its design uses separate modules for predicting an answer, generating evidence conditioned on that answer, and deciding whether the question is answerable. The approach treats abstention as part of the task: a system should be able to withhold a response when the available image evidence is insufficient, rather than always producing an answer.
For mammography, connecting a finding to a specific image region can help make an answer more interpretable. The paper’s abstract describes the evidence as region-grounded, but does not detail the grounding method in the supplied summary.
Evaluation on public mammography datasets
The team evaluated CARE-MVLM on 7,358 triplets curated from CBIS-DDSM, MIAS and VinDr-Mammo. The authors report that it improved lesion-sensitive selective behaviour over baselines using the same vision-language-model backbone. They also report stronger lesion-sensitive abstention than GPT-5.6 and Gemini-3.5-flash, and the highest combined reliability for answers, evidence grounding and abstention among the systems compared.
The preprint’s authors describe the assessment as using existing, de-identified data that were publicly available before the study. No new patient data were collected. This is a model evaluation on curated datasets, rather than a prospective study of the system in a clinical workflow. The reported results therefore concern benchmark performance, not demonstrated effects on patient diagnosis or care.
The work focuses on a practical design question for medical AI: whether a model can pair an answer with image evidence and recognise when it should decline to answer. The abstract reports the direction of the comparisons but does not provide numerical effect estimates, so the size of the reported improvements cannot be judged from that summary.