TusharLab/iTrialSapce_LungsNodule

iTrialSpace — Lung Nodule Dataset

A multi-part lung-nodule resource spanning real-CT segmentation masks, nodule metadata/profiles, large-scale synthetic CT volumes (digital-twin / in-silico trial "modes"), and a vision-language evaluation set of axial QC images.

⚠️ Large dataset. The repository totals several terabytes. Most folders are plain files; the vlm_dataset/synthetic/ images are packed into tar shards (see below). Use targeted downloads — pull only the paths you need.

Repository layout

.
├── profiles/            # per-nodule profile tables (7 CSVs)
├── meta/                # per-dataset metadata tables (7 CSVs)
├── masks/               # real-CT segmentation masks (.nii.gz), 7 source datasets
├── generated_cts/       # synthetic CT volumes + masks, 13 generation "modes"
├── inserted_masks/      # inserted-nodule masks (.nii.gz), same 13 modes
└── vlm_dataset/
    └── synthetic/       # axial QC PNGs for VLM eval, packed as 27 tar shards

Source datasets referenced throughout: DLCS24, IMDCT, LNDbv4, LUNA16, LUNA25, LUNGx, NSCLCR.

---

profiles/ — nodule profile tables

Seven CSVs (one per source dataset), e.g. LUNA25_nodule_profiles.csv. Per-nodule descriptors.

meta/ — dataset metadata tables

Seven CSVs (one per source dataset), e.g. LUNA25_dataset_FEB192026_FIT_v1.csv. Case/series-level metadata.

masks/ — real-CT segmentation masks

NIfTI masks (*.nii.gz) for the 7 source datasets, 41,804 files total. Each dataset contains several mask types as subfolders (e.g. nodule_seg/, organ_seg/, refined_seg/, combined_seg/, radiomics_seg/).

Dataset	Mask files
LUNA25	18,363
IMDCT	10,127
DLCS24	7,312
LUNA16	2,985
LNDbv4	1,461
NSCLCR	1,263
LUNGx	293
Total	41,804

generated_cts/ — synthetic CT volumes (in-silico trial "modes")

Synthetic CT volumes with paired masks and per-case JSON, organized into 13 generation modes, 267,882 files total (~3 TB). Each case folder typically contains synthetic_ct.nii.gz, input_mask*.nii.gz, and metadata JSON (dataset.json, pipeline_summary.json, nodmaisi_audit.json).

Mode	Files
mode1_controlled_prevalence	5,057
mode2_size_detection_curve	3,003
mode3_location_sensitivity	2,510
mode4_demographic_stratification	4,813
mode5_counterfactual	15,081
mode6_cross_dataset	9,029
mode7_bootstrap_confidence	24,047
mode8_algorithm_comparison	3,011
mode9_screening_simulation	9,021
mode10_multi_nodule_realism	2,999
mode11_digital_twin_isolation	78,134
mode12_digital_twin_complete	54,099
mode13_digital_twin_cross	57,078
Total	267,882

inserted_masks/ — inserted-nodule masks

NIfTI masks (*_mask.nii.gz) marking the inserted/synthesized nodules used to generate the generated_cts/ volumes — one mask per case, organized into the same 13 modes. 45,018 files total (~130 GB), plus a small JSON per mode. Filenames encode provenance, e.g. iTS--<run>--C0252--host-<hostcase>--src-<source_dataset>--nod-<nodule_id>_mask.nii.gz.

Mode	Files
mode1_controlled_prevalence	1,001
mode2_size_detection_curve	606
mode3_location_sensitivity	505
mode4_demographic_stratification	804
mode5_counterfactual	2,505
mode6_cross_dataset	1,505
mode7_bootstrap_confidence	4,020
mode8_algorithm_comparison	501
mode9_screening_simulation	1,503
mode10_multi_nodule_realism	502
mode11_digital_twin_isolation	13,094
mode12_digital_twin_complete	9,010
mode13_digital_twin_cross	9,462
Total	45,018

---

vlm_dataset/synthetic/ — tar shards

Axial QC images used for vision-language evaluation: 1,004,235 files (~1,004,176 PNGs + summary CSV/JSON). To keep the repo healthy, these are packed into 27 tar shards, organized by model × mode.

vlm_dataset/synthetic/
├── lung_axial_mode1.tar … lung_axial_mode13.tar                     # 13 shards
├── lung_axial_medgemma_mode1.tar … lung_axial_medgemma_mode13.tar   # 13 shards
├── misc.tar                                                         # 59 summary csv/json
└── _file_manifest.txt.gz                                            # full file → list index

• Two model groups: lung_axial and lung_axial_medgemma.
• One shard per (model, mode) — 13 modes each → 26 shards, plus misc.tar.
• Paths are preserved inside the tars, so extraction reconstructs the original tree (lung_axial/qc_overlays/qc_mode7_…_C0085.png, etc.).
• _file_manifest.txt.gz lists every original file path (one per line).

Approx. files per (model, mode) shard (lung_axial and lung_axial_medgemma are symmetric):

Mode	files/shard	Mode	files/shard
mode1	11,998	mode8	5,988
mode2	7,200	mode9	17,940
mode3	6,000	mode10	5,887
mode4	9,588	mode11	152,424
mode5	29,952	mode12	91,311
mode6	18,000	mode13	97,982
mode7	47,818

How to download and extract a shard

# one shard
huggingface-cli download TusharLab/iTrialSapce_LungsNodule \
  vlm_dataset/synthetic/lung_axial_mode7.tar \
  --repo-type dataset --local-dir .

tar -xf vlm_dataset/synthetic/lung_axial_mode7.tar      # restores lung_axial/.../qc_mode7_*.png

# Python: download + extract
from huggingface_hub import hf_hub_download
import tarfile

p = hf_hub_download(
    "TusharLab/iTrialSapce_LungsNodule",
    "vlm_dataset/synthetic/lung_axial_medgemma_mode11.tar",
    repo_type="dataset",
)
with tarfile.open(p) as t:
    t.extractall("synthetic_extracted/")

# stream a shard with WebDataset (no full extraction)
import webdataset as wds
url = "https://huggingface.co/datasets/TusharLab/iTrialSapce_LungsNodule/resolve/main/vlm_dataset/synthetic/lung_axial_mode3.tar"
ds = wds.WebDataset(url)
for sample in ds:
    key = sample["__key__"]          # original relative path (minus extension)
    img = sample.get("png")          # PNG bytes
    break

---

General download tips

# whole repo (very large — usually not what you want)
huggingface-cli download TusharLab/iTrialSapce_LungsNodule --repo-type dataset

# just one folder, e.g. masks for one dataset
huggingface-cli download TusharLab/iTrialSapce_LungsNodule \
  --repo-type dataset --include "masks/LUNA25/**"

Notes & provenance

• Synthetic CT volumes (generated_cts/), inserted-nodule masks (inserted_masks/), and synthetic VLM images (vlm_dataset/synthetic/) are machine-generated; real-CT-derived assets live under masks/, meta/, profiles/.
• File counts above were verified at upload time: masks/ = 41,804; generated_cts/ = 267,882; inserted_masks/ = 45,018; vlm_dataset/synthetic/ = 1,004,235 files packed into 27 tar shards.

• Intended use / scope: TODO
• Licensing of underlying source datasets (LUNA16/25, LNDb, LUNGx, NSCLC-Radiomics, etc.): confirm per-source terms — the apache-2.0 above applies to this packaging, not necessarily to third-party source data.
• Contact: Fakrul Islam Tushar (first author) — fitushar@arizoan.edu · Tushar Laboratory

Citation

If you use this dataset, please cite:

@article{tushar2026itrialspace,
  title={iTRIALSPACE: Programmable Virtual Lesion Trials for Controlled Evaluation of Lung CT Models},
  author={Tushar, Fakrul Islam and Momy, Umme Hafsa and Lo, Joseph Y and Rubin, Geoffrey D},
  journal={arXiv preprint arXiv:2605.05761},
  year={2026}
}