hugging-science/arc-aphasia-bids

Aphasia Recovery Cohort (ARC)

Multimodal neuroimaging dataset for stroke-induced aphasia research.

Dataset Summary

The Aphasia Recovery Cohort (ARC) is a large-scale, longitudinal neuroimaging dataset containing multimodal MRI scans from 230 chronic stroke patients with aphasia. This HuggingFace-hosted version provides direct Python access to the BIDS-formatted data with embedded NIfTI files.

Metric	Count
Subjects	230
Sessions	902
T1-weighted scans	444 sessions (447 runs)*
T2-weighted scans	440 sessions (441 runs)*
FLAIR scans	233 sessions (235 runs)*
BOLD fMRI (naming40 task)	750 sessions (894 runs)
BOLD fMRI (resting state)	498 sessions (508 runs)
Diffusion (DWI)	613 sessions (2,089 runs)
Single-band reference	88 sessions (322 runs)
Expert lesion masks	228

*Sessions with multiple runs of the same structural modality now include all runs as a list.

• Source: OpenNeuro ds004884
• Paper: Gibson et al., Scientific Data 2024
• License: CC0 1.0 (Public Domain)

Supported Tasks

• Lesion Segmentation: Expert-drawn lesion masks enable training/evaluation of stroke lesion segmentation models
• Aphasia Severity Prediction: WAB-AQ scores (0-100) provide continuous severity labels for regression tasks
• Aphasia Type Classification: WAB-derived aphasia type labels (Broca's, Wernicke's, Anomic, etc.)
• Longitudinal Analysis: Multiple sessions per subject enable recovery trajectory modeling
• Diffusion Analysis: Full bval/bvec gradients enable tractography and diffusion modeling
• Task-based fMRI: Naming40 and resting-state runs separated for targeted analysis

Languages

Clinical metadata and documentation are in English.

Dataset Structure

Data Instance

Each row represents a single scanning session (subject + timepoint):

{
    "subject_id": "sub-M2001",
    "session_id": "ses-1",
    "t1w": [<nibabel.Nifti1Image>, ...],       # T1-weighted structural (list of runs)
    "t2w": [<nibabel.Nifti1Image>, ...],       # T2-weighted structural (list of runs)
    "t2w_acquisition": "space_2x",             # T2w sequence type
    "flair": [<nibabel.Nifti1Image>, ...],     # FLAIR structural (list of runs)
    "bold_naming40": [<Nifti1Image>, ...],     # Naming task fMRI runs
    "bold_rest": [<Nifti1Image>, ...],         # Resting state fMRI runs
    "dwi": [<Nifti1Image>, ...],               # Diffusion runs
    "dwi_bvals": ["0 1000 1000...", ...],      # b-values per run
    "dwi_bvecs": ["0 0 0\n1 0 0\n...", ...],   # b-vectors per run
    "sbref": [<Nifti1Image>, ...],             # Single-band references
    "lesion": <nibabel.Nifti1Image>,           # Expert lesion mask
    "age_at_stroke": 58.0,
    "sex": "M",
    "race": "w",
    "wab_aq": 72.5,
    "wab_days": 180.0,
    "wab_type": "Anomic"
}

Data Fields

Field	Type	Description
subject_id	string	BIDS subject identifier (e.g., "sub-M2001")
session_id	string	BIDS session identifier (e.g., "ses-1")
t1w	Sequence[Nifti]	T1-weighted structural MRI runs
t2w	Sequence[Nifti]	T2-weighted structural MRI runs
t2w_acquisition	string	T2w acquisition type: space_2x, space_no_accel, turbo_spin_echo (nullable)
flair	Sequence[Nifti]	FLAIR structural MRI runs
bold_naming40	Sequence[Nifti]	BOLD fMRI runs for naming40 task
bold_rest	Sequence[Nifti]	BOLD fMRI runs for resting state
dwi	Sequence[Nifti]	Diffusion-weighted imaging runs
dwi_bvals	Sequence[string]	b-values for each DWI run (space-separated)
dwi_bvecs	Sequence[string]	b-vectors for each DWI run (3 lines, space-separated)
sbref	Sequence[Nifti]	Single-band reference images
lesion	Nifti	Expert-drawn lesion segmentation mask (nullable)
age_at_stroke	float32	Subject age at stroke onset in years
sex	string	Biological sex ("M" or "F")
race	string	Self-reported race: "b" (Black), "w" (White), or null
wab_aq	float32	Western Aphasia Battery Aphasia Quotient (0-100)
wab_days	float32	Days since stroke when WAB was administered
wab_type	string	Aphasia type classification

Data Splits

Split	Sessions	Description
train	902	All sessions (no predefined train/test split)

Note: Users should implement their own train/validation/test splits, ensuring no subject overlap between splits for valid evaluation.

Dataset Creation

Curation Rationale

The ARC dataset was created to address the lack of large-scale, publicly available neuroimaging data for aphasia research. It enables:

• Development of automated lesion segmentation algorithms
• Machine learning models for aphasia severity prediction
• Studies of brain plasticity and language recovery

Source Data

Data were acquired under studies approved by the Institutional Review Board at the University of South Carolina (per the OpenNeuro ds004884 dataset_description.json).

Annotations

Expert-drawn lesion segmentation masks are provided in derivatives/lesion_masks/.

Personal and Sensitive Information

• Anonymized: OpenNeuro ds004884 dataset_description.json states the final dataset is fully anonymised.

Considerations for Using the Data

Social Impact

This dataset enables research into:

• Improved stroke rehabilitation through better outcome prediction
• Automated clinical tools for aphasia assessment
• Understanding of brain-language relationships

Potential Biases

• Site: Data were acquired under University of South Carolina IRB approval (per OpenNeuro metadata)
• Age: Adult cohort (age_at_stroke ranges from 27 to 80 years in participants.tsv)

Known Limitations

• Not all sessions have all modalities (check for None/empty lists)
• Lesion masks available for 228/230 subjects
• Longitudinal follow-up varies by subject (1-30 sessions)

Usage

from datasets import load_dataset

ds = load_dataset("hugging-science/arc-aphasia-bids", split="train")

# Access a session
session = ds[0]
print(session["subject_id"])  # "sub-M2001"
print(session["t1w"][0])      # nibabel.Nifti1Image
print(session["wab_aq"])      # Aphasia severity score

# Access BOLD by task type
for run in session["bold_naming40"]:
    print(f"Naming40 run shape: {run.shape}")

for run in session["bold_rest"]:
    print(f"Resting state run shape: {run.shape}")

# Access DWI with gradient information
for i, (dwi_run, bval, bvec) in enumerate(zip(
    session["dwi"], session["dwi_bvals"], session["dwi_bvecs"]
)):
    print(f"DWI run {i+1}: shape={dwi_run.shape}")
    print(f"  b-values: {bval[:50]}...")
    print(f"  b-vectors: {bvec[:50]}...")

# Filter by T2w acquisition type (for paper replication)
space_only = ds.filter(
    lambda x: (
        x["lesion"] is not None
        and len(x["t2w"]) > 0
        and x["t2w_acquisition"] in ("space_2x", "space_no_accel")
    )
)
# Returns 223 SPACE samples (115 space_2x + 108 space_no_accel)

# Clinical metadata analysis
import pandas as pd

# Select only scalar columns to avoid loading NIfTI columns into RAM
df = ds.select_columns([
    "subject_id", "session_id", "age_at_stroke",
    "sex", "race", "wab_aq", "wab_days", "wab_type"
]).to_pandas()
print(df.describe())

Technical Notes

Multi-Run Modalities

Functional and diffusion modalities support multiple runs per session:

• Empty list [] = no data for this session
• List with items = all runs for this session, sorted by filename

DWI Gradient Files

Each DWI run has aligned gradient information:

• dwi_bvals: Space-separated b-values (e.g., "0 1000 1000 1000...")
• dwi_bvecs: Three lines of space-separated vectors (x, y, z directions)

These are essential for diffusion tensor imaging (DTI) and tractography analysis.

Memory Considerations

NIfTI files are loaded on-demand. For large-scale processing:

for session in ds:
    process(session)
    # Data is garbage collected after each iteration

Original BIDS Source

This dataset is derived from OpenNeuro ds004884. The original BIDS structure is preserved in the column naming and organization.

Additional Information

Dataset Curators

• Original Dataset: Gibson et al. (University of South Carolina)
• HuggingFace Conversion: The-Obstacle-Is-The-Way

Licensing

This dataset is released under CC0 1.0 Universal (Public Domain). You can copy, modify, distribute, and perform the work, even for commercial purposes, all without asking permission.

Citation

@article{gibson2024arc,
  title={The Aphasia Recovery Cohort, an open-source chronic stroke repository},
  author={Gibson, Makayla and Newman-Norlund, Roger and Bonilha, Leonardo and Fridriksson, Julius and Hickok, Gregory and Hillis, Argye E and den Ouden, Dirk-Bart and Rorden, Christopher},
  journal={Scientific Data},
  volume={11},
  pages={981},
  year={2024},
  publisher={Nature Publishing Group},
  doi={10.1038/s41597-024-03819-7}
}

Contributions

Thanks to @The-Obstacle-Is-The-Way for converting this dataset to HuggingFace format with native Nifti() feature support.

Changelog

v4 (December 2025)

• BREAKING: t1w, t2w, flair changed from Nifti() to Sequence(Nifti()) for full data fidelity
• FIX: 6 sessions with multiple structural runs now include all files (previously set to None)
• NOTE: Most sessions have exactly 1 structural scan; access via session["t2w"][0]

v3 (December 2025)

• RETRACTED: Attempted fix for 222 → 223 SPACE samples was incorrect diagnosis
• NOTE: The missing sample is caused by a schema design flaw (see v4 fix above), not upload issues

v2 (December 2025)

• BREAKING: bold column split into bold_naming40 and bold_rest for task-specific analysis
• NEW: dwi_bvals and dwi_bvecs columns for diffusion gradient information
• NEW: race column from participants.tsv
• NEW: wab_days column (days since stroke when WAB administered)
• NEW: t2w_acquisition column for T2w sequence type filtering

v1 (December 2025)

• Initial release with 13 columns