OAIZIB-CM.py

import os
import pandas as pd
import nibabel as nib
import datasets
from datasets import (
    GeneratorBasedBuilder,
    SplitGenerator,
    Split,
    DatasetInfo,
    Features,
    Array3D,
    Value,
)

logger = datasets.logging.get_logger(__name__)

EXPECTED_SHAPE = (160, 384, 384)

_CITATION = """\
@article{YAO2024103035,
  title = {CartiMorph: A framework for automated knee articular cartilage morphometrics},
  journal = {Medical Image Analysis},
  author = {Yongcheng Yao and Junru Zhong and Liping Zhang and Sheheryar Khan and Weitian Chen},
  volume = {91},
  pages = {103035},
  year = {2024},
  issn = {1361-8415},
  doi = {https://doi.org/10.1016/j.media.2023.103035}
}

@InProceedings{10.1007/978-3-031-82007-6_16,
  author = {Yao, Yongcheng and Chen, Weitian},
  editor = {Wu, Shandong and Shabestari, Behrouz and Xing, Lei},
  title = {Quantifying Knee Cartilage Shape and Lesion: From Image to Metrics},
  booktitle = {Applications of Medical Artificial Intelligence},
  year = {2025},
  publisher = {Springer Nature Switzerland},
  address = {Cham},
  pages = {162--172}
}
"""

_DESCRIPTION = """\
This is the official release of the OAIZIB-CM dataset.

OAIZIB-CM is based on the OAIZIB dataset.
OAIZIB paper: Automated Segmentation of Knee Bone and Cartilage combining Statistical Shape Knowledge 
and Convolutional Neural Networks: Data from the Osteoarthritis Initiative.
In OAIZIB-CM, tibial cartilage is split into medial and lateral tibial cartilages.
OAIZIB-CM includes CLAIR-Knee-103R, consisting of:
  - a template image learned from 103 MR images of subjects without radiographic OA,
  - a corresponding 5-ROI segmentation mask for cartilages and bones, and
  - a corresponding 20-ROI atlas for articular cartilages.

This dataset is released under the CC BY-NC 4.0 license.
"""

_HOMEPAGE_URL = "https://github.com/YongchengYAO/CartiMorph"


class OAIZIBCMDataset(GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    def _info(self):
        # Define dataset information including feature schema
        return DatasetInfo(
            description=_DESCRIPTION,
            features=Features({
                "image": Array3D(shape=EXPECTED_SHAPE, dtype="float32"),
                "mask": Array3D(shape=EXPECTED_SHAPE, dtype="float32"),
                "image_path": Value("string"),
                "mask_path": Value("string"),
            }),
            homepage=_HOMEPAGE_URL,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        # Download dataset metadata and data files
        base_url = "https://huggingface.co/datasets/YongchengYAO/OAIZIB-CM/resolve/load_dataset-support"
        train_csv_url = f"{base_url}/train.csv"
        test_csv_url = f"{base_url}/test.csv"
        csv_paths = dl_manager.download({"train": train_csv_url, "test": test_csv_url})
        logger.info(f"Downloaded CSV paths: {csv_paths}")
        
        # Extract main dataset archive
        data_root_dir = dl_manager.download_and_extract(
            "https://huggingface.co/datasets/YongchengYAO/OAIZIB-CM/resolve/load_dataset-support/data/OAIZIB-CM.zip"
        )
        data_dir = os.path.join(data_root_dir, "OAIZIB-CM")
        logger.info(f"Data directory set to {data_dir}")
        
        # Load split metadata
        train_df = pd.read_csv(csv_paths["train"])
        test_df = pd.read_csv(csv_paths["test"])

        # Define split generators for training and test sets
        return [
            SplitGenerator(
                name=Split.TRAIN,
                gen_kwargs={"df": train_df, "split": "train", "data_dir": data_dir},
            ),
            SplitGenerator(
                name=Split.TEST,
                gen_kwargs={"df": test_df, "split": "test", "data_dir": data_dir},
            ),
        ]

    def _generate_examples(self, df, split, data_dir):
        # Set up directory paths based on the split
        if split == "train":
            img_dir = os.path.join(data_dir, "imagesTr")
            mask_dir = os.path.join(data_dir, "labelsTr")
        elif split == "test":
            img_dir = os.path.join(data_dir, "imagesTs")
            mask_dir = os.path.join(data_dir, "labelsTs")
        else:
            raise ValueError(f"Unknown split: {split}")
        
        # Log directories and ensure they exist
        logger.info(f"Looking for {split} images in: {img_dir}")
        logger.info(f"Looking for {split} masks in: {mask_dir}")
        os.makedirs(img_dir, exist_ok=True)
        os.makedirs(mask_dir, exist_ok=True)
        
        # Process and yield examples
        count = 0
        skipped = 0
        for idx, row in df.iterrows():
            img_file = row["image"]
            mask_file = row["mask"]
            img_path = os.path.join(img_dir, img_file)
            mask_path = os.path.join(mask_dir, mask_file)

            # Skip files that don't exist
            if not os.path.exists(img_path):
                logger.warning(f"Image not found: {img_path}")
                skipped += 1
                continue
            if not os.path.exists(mask_path):
                logger.warning(f"Mask not found: {mask_path}")
                skipped += 1
                continue

            try:
                # Load and process image data
                img_nib = nib.load(img_path)
                image = img_nib.get_fdata().astype("float32")
                mask_nib = nib.load(mask_path)
                mask = mask_nib.get_fdata().astype("float32")

                yield idx, {
                    "image": image,
                    "mask": mask,
                    "image_path": img_path,
                    "mask_path": mask_path,
                }
                count += 1
            except Exception as e:
                logger.error(f"Error processing {img_path} and {mask_path}: {e}")
                skipped += 1
                
        logger.info(f"Successfully yielded {count} examples for {split} split, skipped {skipped}")