---
license: apache-2.0
task_categories:
- visual-question-answering
language:
- en
- zh
tags:
- medical-imaging
- pet-ct
- volumetric-vqa
- benchmark
pretty_name: MetaStructAtlas
configs:
- config_name: publication_preview
  default: true
  data_files:
  - split: preview
    path:
    - preview/example_2a.json
    - preview/example_2b.json
    - preview/example_2c.json
---

# MetaStructAtlas

**Whole-body FDG PET/CT annotations and the MetaStructVQA benchmark.** MetaStructAtlas links anatomical masks to report-derived morphology and FDG-uptake descriptions. MetaStructVQA presents these records as parallel Chinese and English questions.

[Preview](#preview) · [Availability](#availability) · [Reproduction code](https://github.com/withtst/MetaStructVQAbenchmark)

## Preview

Select **`publication_preview`** in the Dataset Viewer, or open one of the three JSON examples below. Each example contains four question views across two languages and two mask settings.

| Task | Guided / mask-free format | Example |
|---|---|---|
| Morphology selection | `2a` / `2a-mf` | [View example](preview/example_2a.json) |
| Binary morphology | `2b` / `2b-mf` | [View example](preview/example_2b.json) |
| FDG description | `2c` / `2c-mf` | [View example](preview/example_2c.json) |

These training-set records show the anatomical target, declared visual inputs, question, options and reference answer. The viewer displays the preview question records. Image volumes and clinical reports belong to the separate full release.

## Benchmark at a glance

| Component | Count |
|---|---:|
| Patients | 490 |
| Train / validation / test patients | 323 / 81 / 86 |
| Standard anatomical classes | 103 |
| Standard anatomical mask files | 50,470 |
| Structured annotation records | 7,275 |
| Questions per language | 100,241 |
| Test questions per language | 17,911 |
| Mask-guided / mask-free test pairs | 4,011 |

The nine question formats cover anatomical and modality recognition, morphology and FDG description, and report integration.

## Availability

The public preview contains the three examples listed above and is available under the Apache License 2.0. The examples are drawn from the training split of the publication dataset described on this card.

The complete publication dataset is planned for release after manuscript acceptance, in accordance with applicable data-access permissions.

## Reproduction

The [MetaStructVQAbenchmark repository](https://github.com/withtst/MetaStructVQAbenchmark) provides the manuscript reproduction package, quick start and reference-result index.