Neuro-VL beats OpenAI and Google models on brain-scan reading

Neuro-VL, a vision-language model trained on nearly one million routine clinical neuroimaging studies, outperformed frontier general-purpose and specialized medical models on brain-scan tasks, according to the authors of a preprint on arXiv.

The model "can interpret complete imaging examinations and engage in natural language conversations about their findings," the abstract states. Across internal and international datasets, the authors say Neuro-VL "outperformed frontier general-purpose and specialized medical models in diagnosing, localizing, and describing abnormalities, as well as generating comprehensive radiology reports."

It also compared serial examinations to assess disease progression.

A separate claim concerns prompts that carry wrong assumptions. Neuro-VL "more reliably resisted misleading clinical assumptions than leading frontier models from OpenAI and Google," the abstract says. That failure mode matters when a clinician queries an assistant with a working hypothesis that turns out to be wrong.

The abstract publishes no per-task accuracy numbers, does not name the health system that supplied the studies, and does not specify which OpenAI and Google models served as the comparison. The authors position the work as evidence that "the collective clinical experience of health systems can be used to develop medical AI with specialized expertise and broad clinical capabilities."