SGMRI-VQA: 41,307 Expert-Annotated MRI Visual Question Answering Pairs

SGMRI-VQA is a 41,307-pair benchmark for spatially grounded reasoning on multi-frame MRI scans. It was built by SpatialGroundingVQA from expert radiologist annotations in the fastMRI+ dataset, covering brain and knee studies. Each question-answer pair includes a clinician-aligned chain-of-thought reasoning trace and frame-indexed bounding-box coordinates.

Use Cases

• Train visual question answering models for medical images based on the 41,307 expert-annotated QA pairs.
• Develop models for spatially grounded reasoning in MRI scans based on the provided bounding-box coordinates.
• Benchmark AI model performance on multi-frame medical image analysis based on the dataset's train/test splits for brain and knee studies.
• Study chain-of-thought reasoning in clinical settings based on the clinician-aligned reasoning traces accompanying each QA pair.

Strengths

• Contains 41,307 expert-annotated question-answer pairs, providing a substantial benchmark.
• Includes clinician-aligned chain-of-thought reasoning traces for interpretability.
• Provides frame-indexed bounding-box coordinates for spatially grounded analysis.
• Built from the established fastMRI+ dataset, suggesting a foundation in real clinical data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for specific splits (e.g., brain vs. knee) is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and file formats require manual inspection after download.

Provenance

Source

SpatialGroundingVQA on Hugging Face, built from the fastMRI+ dataset.

Collection Method

Built from expert radiologist annotations.

Freshness

Last updated 2026-05-08 04:20:25; freshness should be verified.