LIDC-IDRI — lung nodule segmentation on chest CT (complete TCIA collection, NIfTI)

The Lung Image Database Consortium / Image Database Resource Initiative collection: 1,010 patients / 1,018 chest CT scans / 243,958 slices with the four-radiologist nodule (≥ 3 mm) contours, re-hosted as NIfTI with per-annotation masks and a documented default ground truth. Source images: TCIA collection LIDC-IDRI (doi:10.7937/K9/TCIA.2015.LO9QL9SX). Source annotations: the QIICR DICOM-SEG/SR re-encoding DICOM-LIDC-IDRI-Nodules (doi:10.7937/TCIA.2018.h7umfurq). Both CC BY 3.0.

This mirror supersedes and absorbs a planned separate DICOM-LIDC-IDRI-Nodules dataset: same annotations, same TrackingUID clustering, over the 883-scan subset that carries contours. Never count the two as independent benchmarks.

What is here

path	content
images/<case>.nii.gz	CT in Hounsfield units, int16, native geometry (1,018)
masks/<case>.nii.gz	default ground truth, uint8 {0,1}: nodules with ≥ 3 annotations at ≥ 50 % consensus of their annotations
masks_instance/<case>.nii.gz	uint16 nodule id (per scan) for every clustered nodule at 50 % consensus
masks_annotation_count/<case>.nii.gz	uint8 number of annotations covering each voxel
masks_annotation/<case>/<case>_nodNN_annK.nii.gz	one uint8 mask per DICOM-SEG object (6,859) — lossless w.r.t. the source
scans.csv	one row per scan: geometry, manufacturer, transfer syntax, duplicate-z slices dropped, nodule/annotation counts, in_luna16, luna16_subset, in_qin_lungct_seg
nodules.csv	one row per clustered nodule (2,651): n_annotations, in_default_gt, consensus/union/intersection voxels, bbox, centroid, median SR characteristics (subtlety … malignancy, volume, diameter)
annotations.csv	one row per SEG object with its SR characteristics and mask path
train.jsonl	loader manifest, one row per scan (single train split — no official split exists)
aux/	LIDC-XML-only.zip (the original XML reads incl. < 3 mm and non-nodule point marks), LIDC-IDRI_MetaData.csv, nodule-count and diagnosis sheets, verbatim from TCIA
preview config	one PNG per scan (the slice with the most default-GT foreground, or the middle slice) with mask and overlay, for the Dataset Viewer

<case> = <PatientID>_<10-char SeriesInstanceUID suffix>; eight patients have two scans (LIDC-IDRI-0132/0151/0315/0332/0355/0365/0442/0484) — group on patient_id.

Not mirrored: the collection's 290 chest radiographs (DX/CR, no annotations) and the 90 3D Slicer Community SEG objects served under the same collection name (QIN CT-challenge algorithm outputs on 10 patients, already in MedOtter/QIN-LungCT-Seg).

Ground truth

Each scan was read by four thoracic radiologists (blinded, then unblinded; no forced consensus). Only nodules ≥ 3 mm were outlined. Upstream ships no consensus mask — each SEG object is one annotation of one nodule, and reader identity "was intentionally not exposed by the LIDC initiative" (Fedorov et al. 2020), so counts here are of annotations (a nodule carries 1–8; distribution over 2,651 nodules: 1 × 771, 2 × 488, 3 × 481, 4 × 897, 5 × 8, 6 × 2, 7 × 3, 8 × 1).

• Default (masks/): nodule included when ≥ 3 annotations exist (the LUNA16 reference-standard rule); voxel included when ≥ 50 % of that nodule's annotations include it (pylidc consensus(clevel=0.5)). Keeps 1,392 nodules; 1–2-annotation nodules are background here. 135 scans have no contour at all (real all-zero masks).
• Any other tier is reconstructible from masks_annotation/ (+ nodules.csv) or from masks_annotation_count/ + masks_instance/.

Overlap with other datasets

LUNA16 is a strict subset: all 888 LUNA16 scans are LIDC-IDRI series (the LUNA16 .mhd name is the SeriesInstanceUID) — see in_luna16 / luna16_subset. LUNA16's nodule masks are sphere approximations; these are the real contours. QIN-LungCT-Seg shares 10 series (in_qin_lungct_seg). Also LIDC-derived elsewhere: SAROS, CADS 0016_lidc, Pulmonary-Nodules-Segmentation, ULS23 (upstream), LUNA22-ISMI, NoduleMNIST3D. Dedup on series_uid.

Provenance and quirks

• DICOM instances pulled from NCI Imaging Data Commons (v24, public S3) — byte-identical to TCIA's; every series verified against the NBIA ImageCount.
• Five series carried duplicate-z slices (0085, 0146, 0418, 0572, 0979): the first slice at each z is kept and SEG frames referencing the twin are composited onto it (n_duplicate_slices_dropped).
• z_gap_min/z_gap_max expose irregular slice spacing; transfer_syntax records JPEG-compressed series (decoded with GDCM).

License and attribution

CC BY 3.0 (https://creativecommons.org/licenses/by/3.0/) — stated on the TCIA collection page for the images and every annotation row, on the DICOM-LIDC-IDRI-Nodules page, and in the LicenseName of every series. Mirrored under TCIA's Data Usage Policy (https://www.cancerimagingarchive.net/data-usage-policies-and-restrictions/): attribute, link the policy, do not attempt re-identification.

Please cite:

• Armato SG III, McLennan G, Bidaut L, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A completed reference database of lung nodules on CT scans. Medical Physics 38(2):915–931, 2011. doi:10.1118/1.3528204
• Armato SG III, et al. Data From LIDC-IDRI. The Cancer Imaging Archive, 2015. doi:10.7937/K9/TCIA.2015.LO9QL9SX
• Fedorov A, Hancock M, Clunie D, et al. DICOM re-encoding of volumetrically annotated Lung Imaging Database Consortium (LIDC) nodules. Medical Physics 47(11):5953–5965, 2020. doi:10.1002/mp.14445; data doi:10.7937/TCIA.2018.h7umfurq
• Clark K, et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. J Digit Imaging 26(6):1045–1057, 2013.

Acknowledgement required by the collection: "The authors acknowledge the National Cancer Institute and the Foundation for the National Institutes of Health, and their critical role in the creation of the free publicly available LIDC/IDRI Database used in this study."