Datasetpublic

UngLong/openm3chest-labels

OpenM3Chest Labels (OM3C) JSON label files and Series UIDs from the OpenM3Chest dataset, prepared for fine-tuning medical vision-language models such as MedGemma. Raw imaging data (DICOM) can be downloaded from IDC (Imaging Data Commons) using the Series Instance UIDs provided in unique_keys.txt. Dataset Summary OpenM3Chest is a medical multimodal multitask dataset for diagnosing chest abnormalities with a focus on lung cancer screening. The original raw data… See the full description on the dataset page: https://huggingface.co/datasets/UngLong/openm3chest-labels.

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes48downloads

Dataset Card

OpenM3Chest Labels (OM3C)

JSON label files and Series UIDs from the OpenM3Chest dataset, prepared for fine-tuning medical vision-language models such as MedGemma.

Raw imaging data (DICOM) can be downloaded from IDC (Imaging Data Commons) using the Series Instance UIDs provided in unique_keys.txt.

Dataset Summary

OpenM3Chest is a medical multimodal multitask dataset for diagnosing chest abnormalities with a focus on lung cancer screening. The original raw data comes from NLST (National Lung Screening Trial) and MIDRC.

This repository contains the label files only (JSON format). Each sample includes:

—keys — DICOM Series Instance UID (use with IDC to download imaging data)
—pids — Patient ID
—lung_bboxes — Bounding boxes for left/right lung regions
—labels — Ground truth label (0 = No, 1 = Yes)
—pixel_size — Voxel spacing
—clinical_data — Patient demographics and clinical history (age, smoking, etc.)
—questions — List of natural language questions for VQA tasks
—data_name — Task identifier

Tasks / Configs

Config	Description
`chest_abn_54`	Chest abnormality: Atelectasis (segmental or greater)
`chest_abn_55`	Chest abnormality: Pleural thickening or effusion
`chest_abn_56`	Chest abnormality: Type 56
`chest_abn_57`	Chest abnormality: Type 57
`chest_abn_58`	Chest abnormality: Type 58
`chest_abn_59`	Chest abnormality: Type 59
`chest_abn_61`	Chest abnormality: Type 61
`CVD_diagnosis`	Cardiovascular disease diagnosis
`CVD_mortality`	Cardiovascular disease mortality prediction
`lung_cancer_risk`	Lung cancer risk assessment
`nodule_attenuation`	Lung nodule attenuation classification
`nodule_location`	Lung nodule location classification
`nodule_margin`	Lung nodule margin classification
`nodule_presence`	Lung nodule presence detection
`nodule_size`	Lung nodule size estimation

Usage

python

from datasets import load_dataset

# Load a specific task
ds = load_dataset("UngLong/openm3chest-labels", name="chest_abn_54")

print(ds["train"][0])
print(ds["test"][0])

Download DICOM Images

Use the keys field (DICOM Series Instance UID) to download imaging data from IDC:

python

from idc_index import IDCClient

client = IDCClient()

series_uid = "1.2.840.113654.2.55.229650531101716203536241646069123704792"
client.download_dicom_series(seriesInstanceUID=series_uid, downloadDir="./dicom")

All 116,488 unique Series Instance UIDs are listed in unique_keys.txt.

Citation

If you use this dataset, please cite the original OpenM3Chest paper and dataset:

@dataset{openm3chest,
  author = {Niu, Chuang},
  title  = {OpenM3Chest},
  year   = {2024},
  doi    = {10.5281/zenodo.14363994},
  url    = {https://zenodo.org/records/14363994}
}

License

MIT — see LICENSE for details. Original data sourced from NLST and MIDRC; please refer to their respective data use agreements.