|
|
import os |
|
|
import pandas as pd |
|
|
from datasets import GeneratorBasedBuilder, SplitGenerator, Split, Value, Features, Image, DatasetInfo, Sequence |
|
|
|
|
|
|
|
|
class NovaDataset(GeneratorBasedBuilder): |
|
|
def _info(self): |
|
|
return DatasetInfo( |
|
|
description="NOVA benchmark: anomaly localization and clinical reasoning in brain MRI.", |
|
|
features=Features({ |
|
|
"image": Image(), |
|
|
"filename": Value("string"), |
|
|
"caption": Value("string"), |
|
|
"case_id": Value("string"), |
|
|
"scan_id": Value("string"), |
|
|
"title": Value("string"), |
|
|
"publication_date": Value("string"), |
|
|
"clinical_history": Value("string"), |
|
|
"differential_diagnosis": Value("string"), |
|
|
"final_diagnosis": Value("string"), |
|
|
"link": Value("string"), |
|
|
"bbox_gold": Sequence({ |
|
|
"x": Value("float32"), |
|
|
"y": Value("float32"), |
|
|
"width": Value("float32"), |
|
|
"height": Value("float32"), |
|
|
}), |
|
|
"bbox_raters": Sequence({ |
|
|
"rater": Value("string"), |
|
|
"x": Value("float32"), |
|
|
"y": Value("float32"), |
|
|
"width": Value("float32"), |
|
|
"height": Value("float32"), |
|
|
}), }), |
|
|
supervised_keys=None, |
|
|
) |
|
|
|
|
|
def _split_generators(self, dl_manager): |
|
|
archive_path = dl_manager.download_and_extract("test/Images.zip") |
|
|
image_files = list(dl_manager.iter_files(archive_path)) |
|
|
|
|
|
|
|
|
csvs = { |
|
|
"captions": dl_manager.download("captions.csv"), |
|
|
"bboxes_gold": dl_manager.download("bboxes_gold.csv"), |
|
|
"bboxes_raters": dl_manager.download("bboxes_raters.csv"), |
|
|
"case_metadata": dl_manager.download("case_metadata.csv"), |
|
|
} |
|
|
|
|
|
return [ |
|
|
SplitGenerator( |
|
|
name=Split.TEST, |
|
|
gen_kwargs={"image_files": image_files, "csvs": csvs}, |
|
|
) |
|
|
] |
|
|
|
|
|
def _generate_examples(self, image_files, csvs): |
|
|
captions = pd.read_csv(csvs["captions"]) |
|
|
gold = pd.read_csv(csvs["bboxes_gold"]) |
|
|
raters = pd.read_csv(csvs["bboxes_raters"]) |
|
|
meta = pd.read_csv(csvs["case_metadata"]) |
|
|
|
|
|
captions["case_id"] = captions["filename"].str.extract(r"^(case\d+)_") |
|
|
|
|
|
gold_grouped = gold.groupby("filename").apply( |
|
|
lambda df: df[["x", "y", "width", "height"]].astype(float).to_dict(orient="records") |
|
|
).to_dict() |
|
|
|
|
|
raters_grouped = raters.groupby("filename").apply( |
|
|
lambda df: df[["rater", "x", "y", "width", "height"]].astype({"x": float, "y": float, "width": float, "height": float}).to_dict(orient="records") |
|
|
).to_dict() |
|
|
meta = meta.set_index("case_id").to_dict("index") |
|
|
|
|
|
images_map = {os.path.basename(f): f for f in image_files} |
|
|
|
|
|
for idx, row in captions.iterrows(): |
|
|
filename = row["filename"] |
|
|
case_id = row["case_id"] |
|
|
record = { |
|
|
"image": images_map.get(filename, None), |
|
|
"filename": filename, |
|
|
"caption": row["caption"], |
|
|
"case_id": case_id, |
|
|
"scan_id": str(row["scan_id"]), |
|
|
"bbox_gold": gold_grouped.get(filename, []), |
|
|
"bbox_raters": raters_grouped.get(filename, []), |
|
|
} |
|
|
if case_id in meta: |
|
|
record.update(meta[case_id]) |
|
|
else: |
|
|
record.update({k: "" for k in ["title", "publication_date", "clinical_history", "differential_diagnosis", "final_diagnosis", "link"]}) |
|
|
yield idx, record |
|
|
|
