This dataset consists of paired CMR images and pixel-level segmentation labels, covering two core clinical sequences: CINE (steady-state free precession, SSFP) for functional analysis and LGE (Late Gadolinium Enhancement) for scar tissue detection. It is structured into two main tasks to facilitate multi-task learning and cross-sequence validation. The annotations cover comprehensive cardiac structures including all four chambers and the myocardium, as well as pathological regions (scar).

Curated by: Beijing Academy of Artificial Intelligence (BAAI)
Funded by: BAAI, Beijing Anzhen Hospital
Shared by: BAAI · Beijing Anzhen Hospital
Language(s) (NLP): English
Language(s) (NLP): Not applicable (imaging dataset; metadata is in English)
License: CC BY-SA 4.0 (Creative Commons Attribution-ShareAlike 4.0 International License)

Dataset Sources [optional]

Repository: https://huggingface.co/datasets/TaipingQu/CMR-MULTI
Paper [optional]: [More Information Needed]

🧪 JOIN CMR-MULTI Challenge (MICCAI 2026)

1st Workshop on Medical World Models

This repository supports the JOIN CMR-MULTI Challenge, focusing on:

📐 Training and benchmarking universal cardiac MRI segmentation models
🌐 Multi-sequence (Cine + LGE), multi-view, and multi-center medical foundation model research
📊 Cardiac function quantification and myocardial scar analysis

⚠️ Out-of-Scope Use

The provided models, data, and code MUST NOT be used for:

❌ Clinical diagnosis or direct patient treatment
❌ Commercial purposes without explicit written permission from the organizers
❌ Any use outside of academic research and official challenge evaluation

All usage must comply with applicable ethical guidelines, data licenses, and institutional review board (IRB) requirements.

Dataset Structure

Task 1: CINE_MULTI (Functional Sequence) This task focuses on dynamic CINE scans, which are used to assess cardiac motion and function.

CINE_MULTI/
├── SAX_TR/
│   ├── image/ (Short-axis CINE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
├── 2CH_TR/
│   ├── image/ (2-chamber view CINE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
├── 4CH_TR/
│   ├── image/ (4-chamber view CINE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
└── dataset.xlsx (Metadata file with 3 sheets: SAX, 2CH, 4CH)

Label Definition (CINE Sequence)

View	Label ID	Description
SAX_TR	1	Left Ventricle Myocardium
	2	Left Ventricle Cavity
	3	Right Ventricle Cavity
2CH_TR	1	Left Ventricle Cavity
	2	Left Ventricle Myocardium
4CH_TR	1	Left Ventricle Cavity
	2	Left Ventricle Myocardium
	3	Right Ventricle Cavity
	4	Right Atrium
	5	Left Atrium

Metadata (dataset.xlsx - CINE) The Excel file contains image/label path mapping and clinical metrics. Sheets: SAX, 2CH, 4CH Common Columns: image_path, anno_path Special Columns (SAX Sheet only): LVEF (Left Ventricular Ejection Fraction, a key clinical metric for cardiac function) Task 2: LGE_MULTI (Pathological Sequence) This task focuses on Late Gadolinium Enhancement scans, which highlight myocardial scar tissue.

LGE_MULTI/
├── SAX_TR/
│   ├── image/ (Short-axis LGE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
├── 2CH_TR/
│   ├── image/ (2-chamber view LGE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
├── 4CH_TR/
│   ├── image/ (4-chamber view LGE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
├── RAS_TR/
│   ├── image/ (Right Atrium - SAX/long-axis LGE images, .nii.gz)
│   └── anno/ (Segmentation labels, .nii.gz)
└── dataset.xlsx (Metadata file with 4 sheets: SAX, 2CH, 4CH, RAS)

Label Definition (LGE Sequence)

View	Label ID	Description
SAX_TR	1	Left Ventricle Cavity
	2	Left Ventricle Myocardium
	3	Myocardial Scar (Scar)
	4	Right Ventricle Cavity
2CH_TR	1	Left Ventricle Cavity
	2	Left Ventricle Myocardium
	3	Myocardial Scar (Scar)
4CH_TR	1	Left Ventricle Cavity
	2	Left Ventricle Myocardium
	3	Myocardial Scar (Scar)
	4	Right Ventricle Cavity
RAS_TR	1	Right Atrium

Metadata (dataset.xlsx - LGE) The Excel file contains image/label path mapping and pathological metadata. Sheets: SAX, 2CH, 4CH, RAS Common Columns: image_path, anno_path Special Columns (SAX Sheet only): Scar_Quality (Quantitative metric assessing the extent/quality of scar tissue annotation)

The original images of the RAS_TR dataset are sourced from https://www.cardiacatlas.org/atriaseg2018-challenge/atria-seg-data/, and the annotations are derived from https://zenodo.org/records/15524472.

Dataset Creation

Curation Rationale

Cardiac MRI analysis requires robust segmentation across both functional (CINE) and pathological (LGE) sequences to enable comprehensive clinical assessment, including function quantification and scar detection. This dataset unifies multi-view (SAX, 2CH, 4CH) annotations for both sequences, providing a valuable resource for developing multi-task AI models that can integrate structural, functional, and pathological information.

Source Data

Data Collection & Processing: Multi-center, IRB-approved, de-identified CMR data
Views: SAX · 2CH · 4CH
Total Annotated Cases: 750+
Format: NIfTI .nii.gz
Annotations: Pixel-level segmentation by professional cardiologists & radiologists

Annotations

Standardized cardiac segmentation protocol; reviewed by senior cardiac imaging experts.

Who are the annotators?

Cardiologists and radiologists from Beijing Anzhen Hospital & collaborative clinical centers.

Personal and Sensitive Information

All data fully de-identified; IRB approved; no patient privacy information included.

Bias, Risks, and Limitations

For research & challenge only; not for clinical use
Data from 3 scanner vendors; may have domain shift between centers
Scar annotation affected by LGE image quality

Citation

BibTeX:

[More Information Needed]

Dataset Card Contact

For questions, feedback, or to report issues, please contact: tpqu@baai.ac.cn

Access & License

This dataset is available for non-commercial research and challenge purposes only. Access is granted only after manual approval by the dataset authors.

By requesting access, you agree to:

Use the data solely for academic research.
Not redistribute or share the data.
Not use the data for clinical purposes.
Cite the dataset in any publications

Downloads last month: -