README.md

---
license: apache-2.0
task_categories:
- visual-question-answering
language:
- en
- zh
tags:
- medical-imaging
- PET/CT
- 3D vision-language model
- multimodal learning
- anatomical reasoning
- metabolic reasoning
- radiology
- nuclear medicine
pretty_name: MetaStructAtlas
# Dataset status
private: false
---

# MetaStructAtlas: A Grounded 3D Vision-Language Dataset and Benchmark for Functional and Structural Reasoning in Whole-Body PET/CT


## Dataset Summary

MetaStructAtlas is a large-scale grounded 3D vision-language dataset designed for functional and structural reasoning in whole-body PET/CT imaging.

Unlike existing medical vision-language datasets that mainly focus on single-modality or regional imaging, MetaStructAtlas integrates:

- 3D PET metabolic information,
- 3D CT anatomical structure,
- organ-level segmentation masks,
- structured radiology reports,
- spatially grounded vision-language annotations.

The dataset provides a unified framework for developing and evaluating foundation models capable of understanding the relationship between anatomical structures and metabolic abnormalities in whole-body PET/CT.

MetaStructAtlas contains:

- **490 paired whole-body PET/CT volumetric scans**
- **50,470 organ-level segmentation masks**
- **103 anatomical structure classes**
- **305 fine-grained anatomical substructures**
- **100,565 grounded visual question-answering (VQA) pairs**

Each textual description is explicitly associated with corresponding anatomical regions and imaging evidence, enabling interpretable multimodal reasoning.


## Dataset Motivation

Whole-body PET/CT combines:

- PET-derived molecular metabolic information
- CT-derived anatomical morphology

and is widely used for oncology screening, staging, treatment response evaluation, and systemic disease assessment.

However, current medical vision-language datasets are mainly limited to:

- 2D images,
- regional CT analysis,
- ungrounded image-report pairs,
- anatomical-only reasoning.

MetaStructAtlas addresses these limitations by providing dense spatial grounding between:
PET image
|
CT image
|
Segmentation mask
|
Radiology description
|
Question-answer reasoning

This enables models to learn clinically meaningful relationships between:

- organ anatomy,
- morphological abnormalities,
- FDG metabolic activity.


## Dataset Construction

The dataset construction pipeline contains four major stages.


### 1. Whole-body PET/CT Collection

MetaStructAtlas was constructed from retrospectively collected whole-body
18F-FDG PET/CT examinations.

Each case contains:

- co-registered PET volume
- CT volume
- clinical radiology reports

All imaging data were anonymized before dataset construction.


### 2. Anatomical Segmentation and Spatial Grounding

CT volumes were processed using an automated anatomical segmentation pipeline.

The dataset provides:

- 103 standardized anatomical masks per subject
- 305 fine-grained anatomical structures

These masks provide spatial references for grounding textual findings.

The anatomical structures cover:

- brain
- thoracic organs
- abdominal organs
- pelvic organs
- cardiovascular structures
- gastrointestinal tract
- vertebrae and ribs


### 3. Clinical Report Structuring

Free-text radiology reports were transformed into structured annotations.

The extraction process identifies:

1. Anatomical entities

2. Morphological descriptions

   Examples:

   - normal morphology
   - abnormal density
   - lesion characteristics
   - structural changes

3. FDG metabolic descriptions

   Examples:

   - increased FDG uptake
   - physiological uptake
   - abnormal metabolic activity


The extracted descriptions are spatially linked to corresponding anatomical masks.


### 4. MetaStructVQA Generation

Based on grounded annotations, we generated
MetaStructVQA, a hierarchical 3D VQA benchmark containing:

**100,565 QA pairs**.

## Benchmark Structure

MetaStructVQA contains three reasoning levels.


## Level 1: Foundational Grounding

Tasks:

### Organ Identification

Input:

- CT volume
- anatomical mask
- question

Output:

- anatomical structure classification


### Modality Identification

Input:

- PET or CT volume

Output:

- modality classification


Purpose:

Evaluate basic visual perception and anatomical localization.


---

## Level 2: Targeted Clinical Characterization

Tasks include:


### Morphological Description

Based on CT:

- lesion appearance
- density
- structural abnormalities


### FDG Metabolism Description

Based on PET/CT:

- FDG uptake pattern
- metabolic abnormalities


Both tasks include:

- mask-guided versions
- mask-free versions


Purpose:

Evaluate structure-aware clinical reasoning.


---

## Level 3: Comprehensive Multimodal Reasoning

Tasks integrate:

- anatomical information
- morphological characteristics
- metabolic activity
- quantitative measurements

Examples:

- SUVmax interpretation
- lesion localization
- anatomical relationship reasoning
- negative finding identification


Purpose:

Evaluate advanced whole-body PET/CT reasoning.


## Data Format

Each sample may contain:

{
"pet": "xxx.nii.gz",
"ct": "xxx.nii.gz",
"mask": "xxx.nii.gz",
"report": "...",
"question": "...",
"options": [],
"answer": ""
}



## Intended Use

MetaStructAtlas can support research in:

- 3D medical vision-language models
- multimodal foundation models
- PET/CT representation learning
- anatomical reasoning
- metabolic reasoning
- medical VQA
- radiology AI


## Limitations

Current version limitations:

1. Organ-level grounding

The dataset focuses on anatomical structures rather than lesion-level segmentation.

2. Single PET tracer

The dataset mainly includes 18F-FDG PET/CT.

3. Spatial alignment

Segmentation masks are derived from CT and transferred to PET space through registration.

## Current Release Status

🚧 Under preparation

The dataset repository currently provides documentation only.

The full imaging dataset will become available after publication acceptance.