brain-structure.py

import os
import json
import logging
import datasets
from pathlib import Path

logger = logging.getLogger(__name__)

_DESCRIPTION = """
A collection of T1-weighted .nii.gz structural MRI scans in a BIDS-like arrangement,
with JSON sidecar metadata indicating train/validation/test splits.
"""

_CITATION = """
@dataset{Radiata-Brain-Structure,
  author    = {Jesse Brown and Clayton Young},
  title     = {Brain-Structure: Processed Structural MRI Brain Scans Across the Lifespan},
  year      = {2025},
  url          = {https://huggingface.co/datasets/radiata-ai/brain-structure},
  note      = {Version 1.0},
  publisher    = {Hugging Face}
}
"""

_HOMEPAGE = "https://huggingface.co/datasets/radiata-ai/brain-structure"
_LICENSE = "ODC-By v1.0"

# The "resolve/main/data.zip" part ensures it grabs data.zip from your 'main' branch.
_DATA_URL = "https://huggingface.co/datasets/radiata-ai/brain-structure/resolve/main/data.zip"


class BrainStructureConfig(datasets.BuilderConfig):
    """Configuration for Brain-Structure dataset (if you need multiple, define them here)."""
    def __init__(self, **kwargs):
        super().__init__(**kwargs)


class BrainStructure(datasets.GeneratorBasedBuilder):
    """
    A dataset loader for T1 .nii.gz files plus JSON sidecars stored in a single ZIP.

    Usage:
        ds_train = load_dataset("radiata-ai/brain-structure", split="train", trust_remote_code=True)
    """

    VERSION = datasets.Version("1.0.0")
    BUILDER_CONFIGS = [
        BrainStructureConfig(
            name="default",
            version=VERSION,
            description="Structural MRIs with sidecar metadata. Splits (train/val/test) indicated in the sidecars.",
        )
    ]
    DEFAULT_CONFIG_NAME = "default"

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "nii_filepath": datasets.Value("string"),
                    "metadata": {
                        "split": datasets.Value("string"),
                        "participant_id": datasets.Value("string"),
                        "session_id": datasets.Value("string"),
                        "study": datasets.Value("string"),
                        "age": datasets.Value("int32"),
                        "sex": datasets.Value("string"),
                        "clinical_diagnosis": datasets.Value("string"),
                        "scanner_manufacturer": datasets.Value("string"),
                        "scanner_model": datasets.Value("string"),
                        "field_strength": datasets.Value("string"),
                        "image_quality_rating": datasets.Value("float"),
                        "total_intracranial_volume": datasets.Value("float"),
                        "license": datasets.Value("string"),
                        "website": datasets.Value("string"),
                        "citation": datasets.Value("string"),
                        "t1_file_name": datasets.Value("string"),
                        "radiata_id": datasets.Value("int32"),
                    },
                }
            ),
        )

    def _split_generators(self, dl_manager: datasets.DownloadManager):
        """
        Downloads and extracts 'data.zip', then defines train/validation/test splits
        by matching sidecars with 'split': 'train'/'validation'/'test'.
        """
        # Download and extract your single ZIP containing all subfolders
        extracted_dir = dl_manager.download_and_extract(_DATA_URL)
        # The ZIP will typically unzip into a folder named "data" or similar. We'll just scan everything inside.

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"data_dir": extracted_dir, "desired_split": "train"},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"data_dir": extracted_dir, "desired_split": "validation"},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"data_dir": extracted_dir, "desired_split": "test"},
            ),
        ]

    def _generate_examples(self, data_dir, desired_split):
        """
        Recursively find sidecar JSONs with 'split' matching desired_split. 
        For each, yield an example containing the .nii.gz path + metadata.
        """
        id_ = 0
        data_path = Path(data_dir)

        for json_path in data_path.rglob("*_scandata.json"):
            with open(json_path, "r") as f:
                sidecar = json.load(f)

            # Only yield if sidecar["split"] matches
            if sidecar.get("split") == desired_split:
                # Build a base prefix from the JSON filename (minus "_scandata")
                # e.g. "msub-OASIS20133_ses-03"
                base_prefix = json_path.stem.replace("_scandata", "")

                # Search for a NIfTI that starts with that prefix and includes '_T1w'
                nii_path = None
                for potential_nii in json_path.parent.glob(f"{base_prefix}*_T1w*.nii.gz"):
                    nii_path = potential_nii
                    break

                if not nii_path or not nii_path.is_file():
                    logger.warning(f"No .nii.gz found for {json_path}")
                    continue

                yield id_, {
                    "id": str(id_),
                    "nii_filepath": str(nii_path),
                    "metadata": {
                        "split": sidecar.get("split", ""),
                        "participant_id": sidecar.get("participant_id", ""),
                        "session_id": sidecar.get("session_id", ""),
                        "study": sidecar.get("study", ""),
                        "age": sidecar.get("age", 0),
                        "sex": sidecar.get("sex", ""),
                        "clinical_diagnosis": sidecar.get("clinical_diagnosis", ""),
                        "scanner_manufacturer": sidecar.get("scanner_manufacturer", ""),
                        "scanner_model": sidecar.get("scanner_model", ""),
                        "field_strength": sidecar.get("field_strength", ""),
                        "image_quality_rating": float(sidecar.get("image_quality_rating", 0.0)),
                        "total_intracranial_volume": float(sidecar.get("total_intracranial_volume", 0.0)),
                        "license": sidecar.get("license", ""),
                        "website": sidecar.get("website", ""),
                        "citation": sidecar.get("citation", ""),
                        "t1_file_name": sidecar.get("t1_file_name", ""),
                        "radiata_id": sidecar.get("radiata_id", 0),
                    },
                }
                id_ += 1