|
|
import os |
|
|
import json |
|
|
import datasets |
|
|
import pandas as pd |
|
|
|
|
|
_DESCRIPTION = """\ |
|
|
MedIAnomaly is a benchmark for evaluating anomaly detection methods on seven diverse medical imaging datasets: |
|
|
RSNA, VinCXR, BrainTumor, LAG, ISIC2018_Task3, Camelyon16, and BraTS2021. It supports both image-level |
|
|
classification and pixel-level segmentation tasks. |
|
|
|
|
|
All datasets follow a consistent one-class learning protocol: the training set contains only normal (non-anomalous) |
|
|
images, while the test set includes both normal and abnormal cases. This setting is designed to reflect real-world |
|
|
scenarios where anomalous samples are rare or unavailable during training. MedIAnomaly provides standardized preprocessing, train/test splits, and label formats to facilitate fair comparison |
|
|
across methods. |
|
|
""" |
|
|
|
|
|
_HOMEPAGE = "https://github.com/caiyu6666/MedIAnomaly/tree/main" |
|
|
|
|
|
_CITATION = """\ |
|
|
@article{cai2024medianomaly, |
|
|
title={MedIAnomaly: A comparative study of anomaly detection in medical images}, |
|
|
author={Cai, Yu and Zhang, Weiwen and Chen, Hao and Cheng, Kwang-Ting}, |
|
|
journal={arXiv preprint arXiv:2404.04518}, |
|
|
year={2024} |
|
|
} |
|
|
""" |
|
|
|
|
|
_BASE_URL = "https://huggingface.co/datasets/randall-lab/medianomaly/resolve/main" |
|
|
|
|
|
_URLS = { |
|
|
"rsna": f"{_BASE_URL}/rsna.tar", |
|
|
"brats2021": f"{_BASE_URL}/brats2021.tar", |
|
|
"braintumor": f"{_BASE_URL}/braintumor.tar", |
|
|
"camelyon16": f"{_BASE_URL}/camelyon16.tar", |
|
|
"isic2018_task3": f"{_BASE_URL}/isic2018.tar", |
|
|
"lag": f"{_BASE_URL}/lag.tar", |
|
|
"vincxr": f"{_BASE_URL}/vincxr.tar", |
|
|
} |
|
|
|
|
|
config_names = {"rsna": "RSNA", "vincxr": "VinCXR", |
|
|
"brats2021": "BraTS2021", "braintumor": "BrainTumor", |
|
|
"camelyon16": "Camelyon16", "isic2018_task3": "ISIC2018_Task3", |
|
|
"lag": "LAG"} |
|
|
|
|
|
class Medianomaly(datasets.GeneratorBasedBuilder): |
|
|
BUILDER_CONFIGS = [ |
|
|
datasets.BuilderConfig(name="rsna", version=datasets.Version("1.0.0"), description="RSNA Pneumonia dataset."), |
|
|
datasets.BuilderConfig(name="brats2021", version=datasets.Version("1.0.0"), description="BraTS2021 brain tumor dataset."), |
|
|
datasets.BuilderConfig(name="braintumor", version=datasets.Version("1.0.0"), description="BrainTumor MRI dataset."), |
|
|
datasets.BuilderConfig(name="camelyon16", version=datasets.Version("1.0.0"), description="Camelyon16 histopathology dataset."), |
|
|
datasets.BuilderConfig(name="isic2018_task3", version=datasets.Version("1.0.0"), description="ISIC 2018 melanoma classification dataset."), |
|
|
datasets.BuilderConfig(name="lag", version=datasets.Version("1.0.0"), description="LAG (glaucoma detection) fundus dataset."), |
|
|
datasets.BuilderConfig(name="vincxr", version=datasets.Version("1.0.0"), description="VinCXR chest X-ray dataset."), |
|
|
] |
|
|
|
|
|
def _info(self): |
|
|
config_name = self.config.name.lower() |
|
|
if config_name in ["rsna", "vincxr", "braintumor", "lag", "camelyon16"]: |
|
|
return datasets.DatasetInfo( |
|
|
description=_DESCRIPTION, |
|
|
features=datasets.Features({ |
|
|
"image": datasets.Image(), |
|
|
"label": datasets.ClassLabel(names=["normal", "abnormal"]), |
|
|
}), |
|
|
supervised_keys=("image", "label"), |
|
|
homepage=_HOMEPAGE, |
|
|
license="apache-2.0", |
|
|
citation=_CITATION, |
|
|
) |
|
|
elif config_name == "brats2021": |
|
|
return datasets.DatasetInfo( |
|
|
description=_DESCRIPTION, |
|
|
features=datasets.Features({ |
|
|
"image": datasets.Image(), |
|
|
"label": datasets.ClassLabel(names=["normal", "abnormal"]), |
|
|
"annotation": datasets.Image(), |
|
|
}), |
|
|
supervised_keys=("image", "label"), |
|
|
homepage=_HOMEPAGE, |
|
|
license="apache-2.0", |
|
|
citation=_CITATION, |
|
|
) |
|
|
elif config_name == "isic2018_task3": |
|
|
return datasets.DatasetInfo( |
|
|
description=_DESCRIPTION, |
|
|
features=datasets.Features({ |
|
|
"image": datasets.Image(), |
|
|
"label": datasets.ClassLabel(names=["normal", "abnormal"]), |
|
|
"labels": datasets.Sequence(datasets.Value("int32")), |
|
|
"MEL": datasets.ClassLabel(names=["melanoma", "non-melanoma"]), |
|
|
"NV": datasets.ClassLabel(names=["nevus", "non-nevus"]), |
|
|
"BCC": datasets.ClassLabel(names=["basal cell carcinoma", "non-basal cell carcinoma"]), |
|
|
"AKIEC": datasets.ClassLabel(names=["actinic keratosis", "non-actinic keratosis"]), |
|
|
"BKL": datasets.ClassLabel(names=["benign keratosis", "non-benign keratosis"]), |
|
|
"VASC": datasets.ClassLabel(names=["vascular lesion", "non-vascular lesion"]), |
|
|
"DF": datasets.ClassLabel(names=["dermatofibroma", "non-dermatofibroma"]), |
|
|
}), |
|
|
supervised_keys=("image", "label"), |
|
|
homepage=_HOMEPAGE, |
|
|
license="apache-2.0", |
|
|
citation=_CITATION, |
|
|
) |
|
|
else: |
|
|
raise NotImplementedError(f"{config_name} is not implemented in Medianomaly.") |
|
|
|
|
|
|
|
|
def _split_generators(self, dl_manager): |
|
|
config_name = self.config.name.lower() |
|
|
if config_name not in _URLS: |
|
|
raise NotImplementedError(f"{config_name} is not implemented in Medianomaly.") |
|
|
archive_path = dl_manager.download_and_extract(_URLS[config_name]) |
|
|
|
|
|
if config_name in ["rsna", "vincxr", "braintumor", "lag"]: |
|
|
data_dir = os.path.join(archive_path, config_names[config_name]) |
|
|
with open(os.path.join(data_dir, "data.json"), "r") as f: |
|
|
metadata = json.load(f) |
|
|
return [ |
|
|
datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={ |
|
|
"samples": metadata["train"], "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={ |
|
|
"samples": metadata["test"], "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
] |
|
|
elif config_name == "brats2021": |
|
|
data_dir = os.path.join(archive_path, config_names[config_name]) |
|
|
return [ |
|
|
datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={ |
|
|
"samples": "train", "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={ |
|
|
"samples": "test", "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
] |
|
|
elif config_name == "camelyon16": |
|
|
data_dir = os.path.join(archive_path, config_names[config_name]) |
|
|
return [ |
|
|
datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={ |
|
|
"samples": "train", "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={ |
|
|
"samples": "test", "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
] |
|
|
elif config_name == "isic2018_task3": |
|
|
data_dir = os.path.join(archive_path, config_names[config_name]) |
|
|
return [ |
|
|
datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={ |
|
|
"samples": "train", "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={ |
|
|
"samples": "test", "base_dir": data_dir, "config": config_name |
|
|
}), |
|
|
] |
|
|
|
|
|
def _generate_examples(self, samples, base_dir, config): |
|
|
if config in ["rsna", "vincxr", "braintumor", "lag"]: |
|
|
base_dir = os.path.join(base_dir, "images") |
|
|
for label_str, items in samples.items(): |
|
|
label = int(label_str) |
|
|
for idx, item in enumerate(items): |
|
|
image_path = os.path.join(base_dir, item) |
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": label, |
|
|
} |
|
|
|
|
|
elif config == "brats2021": |
|
|
if samples == "train": |
|
|
base_dir = os.path.join(base_dir, "train") |
|
|
for idx, item in enumerate(os.listdir(base_dir)): |
|
|
image_path = os.path.join(base_dir, item) |
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 0, |
|
|
} |
|
|
elif samples == "test": |
|
|
image_dir_normal = os.path.join(base_dir, "test", "normal") |
|
|
image_dir_tumor = os.path.join(base_dir, "test", "tumor") |
|
|
annot_dir = os.path.join(base_dir, "test", "annotation") |
|
|
|
|
|
idx = 0 |
|
|
for fname in os.listdir(image_dir_normal): |
|
|
if fname.endswith(".png"): |
|
|
image_path = os.path.join(image_dir_normal, fname) |
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 0, |
|
|
"annotation": None, |
|
|
} |
|
|
idx += 1 |
|
|
|
|
|
for fname in os.listdir(image_dir_tumor): |
|
|
if fname.endswith(".png"): |
|
|
image_path = os.path.join(image_dir_tumor, fname) |
|
|
annot_name = fname.replace("flair", "seg") |
|
|
annot_path = os.path.join(annot_dir, annot_name) |
|
|
|
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 1, |
|
|
"annotation": annot_path, |
|
|
} |
|
|
idx += 1 |
|
|
|
|
|
elif config == "camelyon16": |
|
|
if samples == "train": |
|
|
base_dir = os.path.join(base_dir, "train") |
|
|
base_dir = os.path.join(base_dir, "good") |
|
|
for idx, item in enumerate(os.listdir(base_dir)): |
|
|
image_path = os.path.join(base_dir, item) |
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 0, |
|
|
} |
|
|
elif samples == "test": |
|
|
base_dir = os.path.join(base_dir, "test") |
|
|
good_dir = os.path.join(base_dir, "good") |
|
|
ungood_dir = os.path.join(base_dir, "Ungood") |
|
|
|
|
|
idx = 0 |
|
|
for item in os.listdir(good_dir): |
|
|
if item.endswith(".png"): |
|
|
image_path = os.path.join(good_dir, item) |
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 0, |
|
|
} |
|
|
idx += 1 |
|
|
|
|
|
for item in os.listdir(ungood_dir): |
|
|
if item.endswith(".png"): |
|
|
image_path = os.path.join(ungood_dir, item) |
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 1, |
|
|
} |
|
|
idx += 1 |
|
|
|
|
|
elif config == "isic2018_task3": |
|
|
if samples == "train": |
|
|
img_dir = os.path.join(base_dir, "ISIC2018_Task3_Training_Input") |
|
|
label_dir = os.path.join(base_dir, "ISIC2018_Task3_Training_GroundTruth") |
|
|
label_file = os.path.join(label_dir, "ISIC2018_Task3_Training_GroundTruth.csv") |
|
|
else: |
|
|
img_dir = os.path.join(base_dir, "ISIC2018_Task3_Test_Input") |
|
|
label_dir = os.path.join(base_dir, "ISIC2018_Task3_Test_GroundTruth") |
|
|
label_file = os.path.join(label_dir, "ISIC2018_Task3_Test_GroundTruth.csv") |
|
|
|
|
|
df = pd.read_csv(label_file) |
|
|
for idx, row in df.iterrows(): |
|
|
image_id = row["image"] |
|
|
image_path = os.path.join(img_dir, f"{image_id}.jpg") |
|
|
if not os.path.exists(image_path): |
|
|
continue |
|
|
label_vector = row.iloc[1:].astype(int).tolist() |
|
|
|
|
|
yield idx, { |
|
|
"image": image_path, |
|
|
"label": 0 if label_vector == [0, 1, 0, 0, 0, 0, 0] else 1, |
|
|
"labels": label_vector, |
|
|
"MEL": label_vector[0], |
|
|
"NV": label_vector[1], |
|
|
"BCC": label_vector[2], |
|
|
"AKIEC": label_vector[3], |
|
|
"BKL": label_vector[4], |
|
|
"DF": label_vector[5], |
|
|
"VASC": label_vector[6], |
|
|
} |
