Datasets:

chehablab
/

ACDC

dataset_info:
  features:
    - name: volume_id
      dtype: int32
    - name: time_id
      dtype: int32
    - name: slice_id
      dtype: int32
    - name: hight
      dtype: float32
    - name: weight
      dtype: float32
    - name: group
      dtype: string
    - name: image
      dtype: image
  splits:
    - name: train
      num_bytes: 1100824556.33
      num_examples: 38346
  download_size: 1091284591
  dataset_size: 1100824556.33
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-nc-sa-4.0
task_categories:
  - image-classification
language:
  - en
tags:
  - mri
  - medical
  - cardiac
  - imaging
pretty_name: Automated Cardiac Diagnosis Challenge
size_categories:
  - 10K<n<100K

Automated Cardiac Diagnosis Challenge (ACDC)

This dataset contains materials from the Automated Cardiac Diagnosis Challenge (ACDC) introduced during MICCAI 2017 by Bernard et al., designed to advance research in cardiac MRI analysis, representation learning, and automated cardiac disease understanding.

The dataset includes cine cardiac MRI acquisitions from healthy subjects and patients with multiple cardiac pathologies.

🫀 Dataset Description

This repository provides a 2D slice-based version of the original ACDC cardiac MRI dataset, designed for efficient deep learning and representation learning workflows.

Each entry corresponds to a single 2D cardiac MRI slice extracted from a cine cardiac MRI sequence.

📦 Dataset Structure

Each dataset entry contains:

volume_id → Unique identifier for the MRI volume/patient
time_id → Temporal frame index within the cardiac cycle
slice_id → Slice index within the volume
image → 2D cardiac MRI slice
width → Image width
height → Image height

Split

The original dataset marks patients from 1 to 100 as training and 101 to 150 as testing. In this dataset the train split contains all the patients and it is for the user to decide the split.

❤️ Clinical Categories

The original ACDC dataset includes subjects from the following groups:

Healthy controls (NOR) (20 Training + 10 Testing)
Myocardial infarction (MINF) (20 Training + 10 Testing)
Dilated cardiomyopathy (DCM) (20 Training + 10 Testing)
Hypertrophic cardiomyopathy (HCM) (20 Training + 10 Testing)
Abnormal right ventricle (RV) (20 Training + 10 Testing)

Use

from datasets import load_dataset
import matplotlib.pyplot as plt

dataset = load_dataset("chehablaborg/acdc_2d", split="train")

sample_id = 314

image = dataset[sample_id]["image"]
time_id = dataset[sample_id]["time_id"]
slice_id = dataset[sample_id]["slice_id"]

plt.imshow(image, cmap="gray")
plt.title(f"time={time_id}, slice={slice_id}")
plt.axis("off")
plt.show()

📚 Citation

If you use this dataset, please mention us https://chehablab.com in an acknowledgement and cite the original publication:

@article{bernard2018deep,
  title={Deep Learning Techniques for Automatic MRI Cardiac Multi-structures Segmentation and Diagnosis: Is the Problem Solved?},
  author={Bernard, Olivier and Lalande, Alain and Zotti, Cl{\'e}ment and Cervenansky, Fr{\'e}d{\'e}ric and others},
  journal={IEEE Transactions on Medical Imaging},
  volume={37},
  number={11},
  pages={2514--2525},
  year={2018},
  month={nov},
  doi={10.1109/TMI.2018.2837502}
}

License

This work is licensed under a Creative Commons CC BY-NC-SA 4.0 License.

Chehab Lab @ 2026