Radiology VLMs

Radiology vision-language models connect images such as chest X-rays, CT, and MRI with reports and clinical questions to support diagnosis, report writing, image retrieval, and tracking change over time. Common approaches include contrastive image-report pretraining, instruction-tuned multimodal language models, and anatomy-aware grounding; newer systems also use clinical rewards, retrieval, tools, and expert gaze to guide their outputs. Evaluation spans tasks and benchmarks such as MIMIC-CXR, IU X-Ray, and CT-RATE, with growing attention to localization, longitudinal reasoning, uncertainty, and performance across institutions and languages. A central challenge is ensuring that fluent answers reflect image evidence rather than text cues or learned shortcuts, so benchmark scores must be weighed against robustness tests and radiologist assessment.