VQA-RAD

The first manually constructed radiology visual question answering dataset, pairing clinician questions about head, chest and abdominal images with reference answers.

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorymultimodal
Subcategorymedical visual question answering (radiology)
Page statusactive
MetricAccuracy (exact match for closed/yes-no questions; model-graded semantic match for open-ended questions)
Directionhigher_is_better
Unit%
Dataset size2244
Dataset licenceCC BY 4.0 (paper); the Hugging Face dataset card states CC0 1.0 Universal, a discrepancy not resolved by a source read for this page
PublisherLister Hill National Center for Biomedical Communications, U.S. National Library of Medicine

What it measures

VQA-RAD shows a model a radiology image (a head CT/MRI, chest X-ray, or abdominal CT) together with a naturally occurring clinical question about it, such as its imaging modality, anatomical plane, organ system, or whether an abnormality is present, and asks the model to answer. It measures whether a model can combine basic radiology image interpretation with reading comprehension of a clinically phrased question, across 11 question categories spanning both yes/no (closed) and free-text (open) answer types.

Task format

The model receives an image and a question in a single multimodal turn and must answer in text. inspect_evals' implementation scores closed-ended (yes/no) questions by exact string match and open-ended questions with a tool-calling grader model that judges semantic equivalence to the reference answer, then reports closed, open, and overall accuracy.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub