|
from datasets import GeneratorBasedBuilder, DownloadManager, DatasetInfo, Features, Value, Sequence, ClassLabel, Image, BuilderConfig, SplitGenerator, Version |
|
import datasets |
|
import pandas as pd |
|
import json |
|
import zipfile |
|
import os |
|
|
|
_DESCRIPTION = """\ |
|
This dataset is curated from the original “The MultiCaRe Dataset” to focus on the chest tuberculosis patients and can be used to develop algorithms of the segmentation of chest CT images and the classification of tuberculosis positive or control. |
|
""" |
|
|
|
_CITATION = """\ |
|
Nievas Offidani, M. and Delrieux, C. (2023) “The MultiCaRe Dataset: A Multimodal Case Report Dataset with Clinical Cases, Labeled Images and Captions from Open Access PMC Articles”. Zenodo. doi: 10.5281/zenodo.10079370. |
|
""" |
|
|
|
class TuberculosisDataset(GeneratorBasedBuilder): |
|
|
|
BUILDER_CONFIGS = [ |
|
BuilderConfig(name="tuberculosis_dataset", version=Version("1.0.0")) |
|
] |
|
|
|
def _info(self): |
|
return DatasetInfo( |
|
description=_DESCRIPTION, |
|
features=Features({ |
|
"case_id": Value("string"), |
|
"gender": Value("string"), |
|
"age": Value("int8"), |
|
"case_text": Value("string"), |
|
"keywords": Value("string"), |
|
"image_files": Sequence(Image()), |
|
"caption": Value("string"), |
|
}), |
|
supervised_keys=None, |
|
homepage="https://zenodo.org/api/records/10079370/files-archive", |
|
citation=_CITATION, |
|
) |
|
|
|
def _split_generators(self, dl_manager): |
|
base_url = "https://raw.githubusercontent.com/zhankai-ye/tuberculosis_dataset/main/" |
|
urls = { |
|
"cases_csv": f"{base_url}cases.csv", |
|
"keywords_json": f"{base_url}article_metadata.json", |
|
"caption_json": f"{base_url}image_metadata.json", |
|
"images_zip": "https://github.com/zhankai-ye/tuberculosis_dataset/raw/main/images/PMC.zip" |
|
} |
|
downloaded_files = dl_manager.download_and_extract(urls) |
|
|
|
return [ |
|
SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs=downloaded_files), |
|
] |
|
|
|
def _generate_examples(self, cases_csv, keywords_json, caption_json, images_zip): |
|
|
|
cases_df = pd.read_csv(cases_csv) |
|
cases_df.dropna(subset=['age'], inplace=True) |
|
|
|
|
|
with open(keywords_json, 'r') as f: |
|
keywords_json_data = json.load(f) |
|
keywords = pd.json_normalize(keywords_json_data) |
|
keywords['keywords'] = keywords['keywords'].apply(lambda x: ', '.join(x) if isinstance(x, list) else x) |
|
|
|
|
|
caption_json_data = [] |
|
with open(caption_json, 'r') as f: |
|
for line in f: |
|
caption_json_data.append(json.loads(line)) |
|
caption = pd.json_normalize(caption_json_data) |
|
|
|
|
|
merged_df = pd.merge(cases_df, keywords[['pmcid', 'keywords']], left_on='case_id', right_on='pmcid', how='left') |
|
merged_df = pd.merge(merged_df, caption[['case_id', 'caption']], on='case_id', how='left') |
|
merged_df = merged_df.where(pd.notnull(merged_df), None) |
|
merged_df['age'] = merged_df['age'].astype('int8') |
|
|
|
|
|
image_file_paths = self._prepare_image_file_paths(images_zip) |
|
|
|
|
|
for idx, row in merged_df.iterrows(): |
|
image_files = image_file_paths.get(row["case_id"], []) |
|
yield idx, { |
|
"case_id": row["case_id"], |
|
"gender": row["gender"], |
|
"age": int(row["age"]), |
|
"case_text": row["case_text"], |
|
"keywords": row["keywords"], |
|
"image_files": image_files, |
|
"caption": row["caption"], |
|
} |
|
|
|
def _prepare_image_file_paths(self, images_zip_path): |
|
image_file_paths = {} |
|
for root, _, files in os.walk(images_zip_path): |
|
for file in files: |
|
if file.endswith('.jpg'): |
|
key = file.split('_')[0] |
|
if key not in image_file_paths: |
|
image_file_paths[key] = [] |
|
image_file_paths[key].append(os.path.join(root, file)) |
|
return image_file_paths |
|
|
