rsna-atd.py

import datasets
import pandas as pd

_CITATION = """\
@InProceedings{huggingface:dataset,
title = {rsna-atd},
author = {Yeow Zi Qin},
year = {2023}
}
"""

_DESCRIPTION = """\
The dataset is the processed version of Kaggle Competition: RSNA 2023 Abdominal Trauma Detection.
It comprises of segmentation of 205 series of CT scans with 5 classes (liver, spleen, right_kidney, 
left_kidney, bowel).
"""

_NAME = "rsna-atd"

_HOMEPAGE = f"https://huggingface.co/datasets/ziq/{_NAME}"

_LICENSE = "MIT"

_DATA = f"https://huggingface.co/datasets/ziq/{_NAME}/resolve/main/data/"


class RSNAATD(datasets.GeneratorBasedBuilder):
    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "image_path": datasets.Value("string"),
                    "mask_path": datasets.Value("string")
                    # "patient_id": datasets.Value("int64"),
                    # "series_id": datasets.Value("int64"),
                    # "aortic_hu": datasets.Value("float64"),
                    # "incomplete_organ": datasets.Value("int64"),
                    # "bowel_healthy": datasets.Value("int64"),
                    # "bowel_injury": datasets.Value("int64"),
                    # "extravasation_healthy": datasets.Value("int64"),
                    # "extravasation_injury": datasets.Value("int64"),
                    # "kidney_healthy": datasets.Value("int64"),
                    # "kidney_low": datasets.Value("int64"),
                    # "kidney_high": datasets.Value("int64"),
                    # "liver_healthy": datasets.Value("int64"),
                    # "liver_low": datasets.Value("int64"),
                    # "liver_high": datasets.Value("int64"),
                    # "spleen_healthy": datasets.Value("int64"),
                    # "spleen_low": datasets.Value("int64"),
                    # "spleen_high": datasets.Value("int64"),
                    # "any_injury": datasets.Value("int64"),
                    # "image": datasets.Array3D(shape=(None, 512, 512)),
                    # "mask": datasets.Array3D(shape=(None, 512, 512)),
                }
            ),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        train_images = dl_manager.download(f"{_DATA}images.tar.gz")
        train_masks = dl_manager.download(f"{_DATA}masks.tar.gz")

        metadata = dl_manager.download(f"{_DATA}metadata.csv")
        train_images = dl_manager.iter_archive(train_images)
        train_masks = dl_manager.iter_archive(train_masks)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "images": train_images,
                    "masks": train_masks,
                    "metadata": metadata,
                },
            ),
        ]

    def _generate_examples(self, images, masks, metadata):
        df = pd.read_csv(metadata)

        for idx, ((image_path), (mask_path)) in enumerate(zip(images, masks)):
            # row = df.loc[df["path"] == image_path.lower().replace("images/", "")]

            yield idx, {
                "image_path": image_path,
                "mask_path": mask_path
                # "patient_id": row["patient_id"].values[0],
                # "series_id": row["series_id"].values[0],
                # "frame_id": row["frame_id"].values[0],
                # "image": {"path": image_path, "bytes": image.read()},
                # "mask": {"path": mask_path, "bytes": mask.read()},
            }