A visual question dataset to test AI on brain tumor MRI scans

Brain tumor diagnosis requires radiologists to interpret thousands of MRI images—a time-consuming task that could benefit from AI assistance. This paper introduces UCSF-PDGM-VQA, a clinical VQA dataset with 2,387 question-answer pairs from 473 glioma MRI studies, and benchmarks six state-of-the-art vision-language models plus one LLM on the task. The evaluation reveals a critical gap: current models struggle with multi-sequence 3D MRI scans, suppressing visual features and over-relying on language priors—a failure mode called modality collapse. The work is intended for researchers developing medical AI and clinicians evaluating VLMs for neuro-oncology; the dataset and baseline results are provided to enable future development.