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.

Key Statistics:

Metric	Count
Subjects	230
Sessions	902
T1-weighted scans	447
T2-weighted scans	441
FLAIR scans	235
BOLD fMRI runs	1,402
Diffusion (DWI) runs	2,089
Single-band reference	322
Expert lesion masks	228

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

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 (256, 256, 176)
    "t2w": <nibabel.Nifti1Image>,           # T2-weighted structural
    "flair": <nibabel.Nifti1Image>,         # FLAIR structural
    "bold": [<Nifti1Image>, ...],           # List of BOLD fMRI runs (4D)
    "dwi": [<Nifti1Image>, ...],            # List of diffusion runs
    "sbref": [<Nifti1Image>, ...],          # List of single-band references
    "lesion": <nibabel.Nifti1Image>,        # Expert lesion mask (binary)
    "age_at_stroke": 58.0,                  # Age at stroke onset
    "sex": "M",                             # Biological sex (M/F)
    "wab_aq": 72.5,                         # WAB Aphasia Quotient (0-100)
    "wab_type": "Anomic"                    # Aphasia classification
}

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	Nifti	T1-weighted structural MRI (nullable)
t2w	Nifti	T2-weighted structural MRI (nullable)
flair	Nifti	FLAIR structural MRI (nullable)
bold	Sequence[Nifti]	BOLD fMRI 4D time-series (all runs)
dwi	Sequence[Nifti]	Diffusion-weighted imaging (all runs)
sbref	Sequence[Nifti]	Single-band reference images (all runs)
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")
wab_aq	float32	Western Aphasia Battery Aphasia Quotient (0-100)
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

Initial Data Collection

Data was collected at the University of South Carolina and Medical University of South Carolina as part of ongoing aphasia recovery research. All participants provided informed consent under IRB-approved protocols.

Who are the source language producers?

N/A - This is a neuroimaging dataset, not a language dataset.

Annotations

Annotation Process

Lesion masks were manually traced by trained neuroimaging experts on T1-weighted or FLAIR images, following established stroke lesion delineation protocols.

Who are the annotators?

Trained neuroimaging researchers at academic medical centers with expertise in stroke neuroanatomy.

Personal and Sensitive Information

• De-identified: All data has been de-identified per HIPAA guidelines
• Defaced: Structural MRI images have been defaced to prevent facial reconstruction
• No PHI: No protected health information is included
• Consent: All participants consented to public data sharing

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

Discussion of Biases

• Geographic bias: Data collected primarily from Southeastern US medical centers
• Age bias: Stroke predominantly affects older adults; pediatric cases underrepresented
• Severity bias: Very severe aphasia cases may be underrepresented due to consent requirements

Other 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 (1-12 sessions)

Usage

Loading the Dataset

from datasets import load_dataset

# Load full dataset
ds = load_dataset("hugging-science/arc-aphasia-bids")

# Access a session
session = ds["train"][0]
print(f"Subject: {session['subject_id']}, Session: {session['session_id']}")

# Access structural imaging
if session["t1w"] is not None:
    t1_data = session["t1w"].get_fdata()
    print(f"T1w shape: {t1_data.shape}")

# Access multi-run functional data
for i, bold_run in enumerate(session["bold"]):
    print(f"BOLD run {i+1}: shape={bold_run.shape}")

Filtering by Modality

# Get only sessions with lesion masks
sessions_with_lesions = ds["train"].filter(lambda x: x["lesion"] is not None)

# Get sessions with BOLD fMRI
sessions_with_bold = ds["train"].filter(lambda x: len(x["bold"]) > 0)

Clinical Metadata Analysis

import pandas as pd

# Extract clinical metadata
df = ds["train"].to_pandas()[["subject_id", "session_id", "age_at_stroke", "sex", "wab_aq", "wab_type"]]
print(df.describe())

Additional Information

Dataset Curators

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

Licensing Information

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 Information

@article{gibson2024arc,
  title={A large-scale longitudinal multimodal neuroimaging dataset for aphasia},
  author={Gibson, M. and others},
  journal={Scientific Data},
  volume={11},
  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.

---

Technical Notes

Multi-Run Support

Functional and diffusion modalities (bold, dwi, sbref) support multiple runs per session. These are stored as lists:

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

Memory Considerations

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

# Stream without loading all into memory
for session in ds["train"]:
    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.