AIBRIEF
The new public dataset and benchmark require vision language models to answer clinical questions about combined PET/CT scans and point at the region in the image that supports each answer.
Reading a whole-body PET/CT scan means holding two images in mind at once: a CT that maps anatomy and a PET that maps metabolism, fused so a radiologist can see structure and function together. Most medical AI today is trained on only one of the two.
A new public release called MetaStructAtlas, paired with a benchmark called MetaStructVQA, asks a different question. The AI has to answer a clinical prompt about a combined scan and point at the region of the image that supports the answer, a property the paper calls "grounded" or region-anchored reasoning.
Scale: 490 co-registered 3D PET/CT volumes, 50,470 organ-level segmentation masks, and 100,565 question-answer pairs that link diagnostic queries to visual evidence across anatomy, morphology, and metabolism.
The authors identify a gap: prior 3D medical vision-language benchmarks are regional CT-only, leaving whole-body PET/CT evaluation untouched. MetaStructVQA is positioned as a test bed for that gap, with state-of-the-art 3D medical VLMs already benchmarked against it.
The "point at the evidence" requirement is the load-bearing part of the benchmark. It shifts the test from "did the model answer" to "did the model answer about the right region," surfacing a different failure mode than text-only medical VQA.
It is a research release, not a clinical tool. Clinical utility, scanner generalization, and downstream impact remain unvalidated in the preprint.