Dataset Card for LIDC-IDRI (Curated Edition)

🌐 Language / 言語: English | 日本語版 (Japanese Version)

Dataset Summary

The LIDC-IDRI Curated Edition is an audited, high-quality volumetric CT collection derived from the internationally recognized reference database TCIA LIDC-IDRI.

This repository (Kotomiya07/LIDC-IDRI) hosts a clean, curated subset of DICOM CT series specifically filtered to remove geometric slice defects, variable slice spacing, thick-slice scans ($\ge 3.0$ mm), prone acquisitions, and contrast-enhanced CTs. It provides an artifact-free, anatomically consistent benchmark optimized for x-ray-to-CT (2D projection to 3D volumetric reconstruction), DRR synthesis, and generative 3D medical vision tasks.

All scans are cataloged in the unified metadata.csv and organized under the LIDC-IDRI/ directory structure.

Key Dataset Statistics

Statistic	Curated Edition (This Repository)	Upstream TCIA LIDC-IDRI	Description
Total DICOM Files	184,617	234,584	Clean, verified DICOM slices
Patient Cases	629	1,010	De-identified cases with high-quality CT
CT Series	630	1,018	Guaranteed uniform z-spacing, complete thorax
Slice Thickness	0.6 mm – 2.5 mm	0.6 mm – 5.0 mm	Thick slices ($\ge 3.0$ mm) strictly excluded
Z-Spacing Uniformity	100% Equidistant	Inconsistent (4 variable, 5 duplicate)	Zero duplicate slices, zero missing slice gaps
Patient Position	100% Supine (FFS/HFS)	Mixed (14 Prone / HFP)	No gravity-induced anatomical inversion
Contrast Agent	0% (Pure Non-Contrast)	~30.5% (310 series)	Predictable tissue HU for frontal CXR pairing
Thoracic Coverage	Min 200 mm (Median: 312.5 mm)	Incomplete scans (<165 mm)	Full adult lung field and dome coverage
Primary Scanner Vendors	GE, SIEMENS, Philips, TOSHIBA	GE, SIEMENS, Philips, TOSHIBA	Multi-vendor scanner diversity preserved

The source collection is provided by The Cancer Imaging Archive (TCIA), funded by the National Cancer Institute (NCI) and the National Institutes of Health (NIH).

Supported Tasks and Use Cases

• Pulmonary Nodule Detection & Segmentation: Benchmarking 2D/3D convolutional architectures and vision transformers to detect and segment pulmonary nodules and suspicious lesions.
• 3D Anatomical Reconstruction & Point Cloud Generation: Converting volumetric CT Hounsfield Units (HU) to clean 3D skeletal and body point clouds ((N, 3) coordinates) for geometric deep learning and anatomical modeling.
• Multi-Modality Chest Radiography Analysis: Studying cross-modal representation learning between 3D volumetric CT scans and corresponding 2D projection chest radiographs (DX and CR).
• Hounsfield Unit (HU) Calibration and Preprocessing: Evaluating windowing, tissue density thresholding, and bed/artifact removal algorithms.
• 2D-to-3D X-ray to CT Synthesis (x-ray2CT & DRR-to-CT): Training neural radiance fields, 3D diffusion models, and cross-view volumetric transformers from single/bi-planar radiographs to 3D volumetric CT.

Dataset Quality Curation & Filtering (for x-ray2CT & 3D Reconstruction)

To ensure robust model convergence, artifact-free 3D volume synthesis, and reliable evaluation in x-ray-to-CT (2D projection to 3D volumetric reconstruction) and DRR-based deep learning workflows, the original LIDC-IDRI dataset was audited across all 1,308 series (234,000+ DICOM slices).

Low-quality, physically inconsistent, or artifact-inducing scans were identified and filtered into a Curated High-Quality Subset.

Filtering Criteria and Exclusions

Defect Category	Series Excluded	Technical Rationale & Impact on Model Training
Geometric Slice Defects (Variable Spacing & Duplicates)	9	Non-uniform z-spacing (dz jitter/gaps up to 38mm in LIDC-IDRI-0514, missing slices in 0672) and duplicate slice locations (dz = 0 in 0418, 0979, 0085, 0146, 0572) cause fatal geometric distortion, zero-division, or non-invertible spatial transforms when stacking into 3D tensors.
Prone Acquisition Position (PatientPosition == HFP)	14	Scanned in prone position (lying face down, [-1, 0, 0, 0, -1, 0] orientation in 0218, 0262, 0291, etc.). Gravity-induced pulmonary atelectasis and mediastinal shifts cause severe structural mismatch against standard supine/erect chest X-rays.
Incomplete Thoracic Coverage (Z-extent < 200 mm)	2	Partial or localized scans (e.g. LIDC-IDRI-0403 with 165mm coverage for airway study) lack lower lung fields and diaphragmatic domes, leaving missing ground-truth anatomy for frontal chest radiographs.
Thick Slices (SliceThickness >= 3.0 mm)	121	Scans with 3.0 mm, 4.0 mm, and 5.0 mm slice thickness suffer from severe partial-volume effects and staircase artifacts along the z-axis. Isotropic resampling cannot recover lost rib cage contours, fine vascular arborization, or micro-structures.
Contrast-Enhanced CT Scans (ContrastBolusAgent)	310	Iodinated contrast agents artificially hyper-attenuate the pulmonary vasculature, aorta, and cardiac chambers (>300 HU). Because non-contrast planar X-rays cannot predict vascular phase contrast enhancement, mixing contrast and non-contrast CT induces severe hallucination and gradient variance during training.
Total Excluded Series (Union)	388	Unique excluded CT series across all criteria.

Curated High-Quality Subset Specifications

Applying the strict curation filters yields 630 high-quality CT series (across 629 unique patient cases):

• 100% Non-contrast, supine chest CT: Clean baseline tissue radiodensities matching standard chest radiography.
• Guaranteed uniform Z-spacing: No duplicate slices, no missing slice gaps, strictly equidistant spatial sampling.
• Complete adult thoracic coverage: Median Z-coverage of 312.5 mm (min: 200.0 mm, max: 415.0 mm), fully capturing lung apices down to the costophrenic angles.
• Thin & high-resolution slice thickness:
  • Thin slices (< 1.5 mm): 365 series (58%) (0.6 mm, 0.75 mm, 0.9 mm, 1.0 mm, 1.25 mm)
  • Medium slices (1.5 mm - 2.5 mm): 265 series (42%) (1.5 mm, 2.0 mm, 2.5 mm)
  • Thick slices ($\ge 3.0$ mm): 0 series (0% - completely excluded)

Critical Preprocessing Note: Overlapping Reconstruction (dz vs SliceThickness)

In 128 CT series (e.g. with SliceThickness = 1.25 mm), images were reconstructed with a 50% overlap (dz = 0.625 mm). Never rely solely on the SliceThickness (0018, 0050) DICOM tag as voxel z-spacing! Doing so will stretch the 3D volume along the z-axis by a factor of 2. Always compute the real slice spacing from the difference between consecutive ImagePositionPatient[2] coordinates:

z_coords = [float(ds.ImagePositionPatient[2]) for ds in sorted_slices]
dz = np.diff(sorted(z_coords)).mean()

Unified Metadata (metadata.csv)

The repository root metadata.csv has been unified to strictly index the 630 high-quality curated CT series. Every series and DICOM file cataloged in metadata.csv is guaranteed to exist in the repository with complete integrity, preventing any missing file errors during automated pipeline ingestion.

Languages

Metadata, medical reports, and DICOM header attributes are written in English (en).

Dataset Structure

Directory Layout

The repository files are structured as follows:

Kotomiya07/LIDC-IDRI/
├── README.md
├── metadata.csv
├── .gitattributes
└── LIDC-IDRI/
    ├── LIDC-IDRI-0001/
    │   └── 01-01-2000-NA-NA-30178/
    │       └── 3000566.000000-NA-03192/
    │           ├── 1-001.dcm
    │           ├── 1-002.dcm
    │           └── ...
    ├── LIDC-IDRI-0002/
    └── ...

Loading the Data with Python and pydicom

You can download individual series or stream the files directly using huggingface_hub and process them using pydicom:

import pydicom
from huggingface_hub import hf_hub_download
import numpy as np

# Download a specific DICOM slice
slice_path = hf_hub_download(
    repo_id="Kotomiya07/LIDC-IDRI",
    repo_type="dataset",
    filename="LIDC-IDRI/LIDC-IDRI-0065/01-01-2000-NA-NA-03808/3000536.000000-NA-68085/1-001.dcm",
)

# Read the DICOM header and pixel array
dcm = pydicom.dcmread(slice_path)
print(f"Modality: {dcm.Modality}")
print(f"Patient ID: {dcm.PatientID}")
print(f"Image Shape: {dcm.pixel_array.shape}")

# Convert raw pixel values to Hounsfield Units (HU)
slope = getattr(dcm, "RescaleSlope", 1)
intercept = getattr(dcm, "RescaleIntercept", 0)
hu_image = dcm.pixel_array.astype(np.float64) * slope + intercept

Data Fields (metadata.csv)

The root metadata.csv describes all series included in this collection:

Field	Type	Description
Series UID	string	Unique DICOM identifier for the acquisition series.
Collection	string	Collection identifier (LIDC-IDRI).
3rd Party Analysis	string	Third-party analysis tag (if applicable).
Data Description URI	string	DOI link to the official TCIA collection page.
Subject ID	string	De-identified patient identifier (e.g., LIDC-IDRI-0001).
Study UID	string	Unique DICOM identifier for the patient examination study.
Study Description	string	Descriptive study notes (if provided in DICOM header).
Study Date	string	Normalized anonymized study date (01-01-2000).
Series Description	string	Series description tag from scanner header.
Manufacturer	string	Medical imaging equipment manufacturer (e.g. GE MEDICAL SYSTEMS, SIEMENS).
Modality	string	Imaging modality: CT (Computed Tomography), DX (Digital X-Ray), CR (Computed Radiography).
SOP Class Name	string	Full DICOM SOP Class name (e.g. CT Image Storage).
SOP Class UID	string	DICOM unique identifier for SOP class.
Number of Images	int64	Total number of slice files in the series.
File Size	string	Human-readable aggregate size of the series (e.g. 70.33 MB).
File Location	string	Relative path to the folder containing the slice files.
Download Timestamp	string	ISO timestamp of the source acquisition download.

Modalities and Imaging Details

• CT (Computed Tomography): High-resolution volumetric chest scans. Slice thickness varies across cases from 0.6 mm to 5.0 mm. Pixel data requires rescale transformation (HU = pixel * RescaleSlope + RescaleIntercept) for calibrated radiological tissue analysis.
• DX (Digital Radiography) & CR (Computed Radiography): Planar chest X-ray radiographs accompanying the CT scans for multi-view or comparative diagnostic tasks.

Dataset Creation and Provenance

Source Data

• Upstream Collection: The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI)
• Host Institution: The Cancer Imaging Archive (TCIA) / National Cancer Institute (NCI)
• DOI: 10.7937/K9/TCIA.2015.LO9QL9SX
• Official TCIA Page: TCIA LIDC-IDRI Overview

Anonymization and Privacy

All patient cases in the LIDC-IDRI dataset have been rigorously de-identified according to the Health Insurance Portability and Accountability Act (HIPAA) Safe Harbor standard before release by TCIA. All Protected Health Information (PHI) such as patient names, real dates of birth, medical record numbers, and clinical institution identifiers have been stripped and replaced with deterministic anonymous codes (LIDC-IDRI-xxxx). Dates have been re-anchored to nominal reference points (e.g. 01-01-2000).

Considerations for Using the Data

Research Use Only

This dataset is provided strictly for academic, research, and technical evaluation purposes. It must not be used as a sole basis for clinical diagnosis, patient management, or clinical decision support without certified regulatory approvals and clinical validation.

Equipment and Protocol Variations

The dataset spans multiple CT scanner manufacturers, detector generations, reconstruction kernels, slice thicknesses, and tube current protocols. When training machine learning models:

• Account for voxel anisotropy (differing SliceThickness and PixelSpacing).
• Apply standardized resampling (e.g., isotropic 1.0 mm³ spacing) and intensity normalization (e.g. HU clipping to [-1000, 400] for lung/soft tissue, or [150, 1500] for bone extraction).

Additional Information

Licensing Information

The LIDC-IDRI collection is published under the Creative Commons Attribution 3.0 Unported (CC BY 3.0) License: https://creativecommons.org/licenses/by/3.0/

Users are free to share and adapt the material for any purpose, even commercially, under the terms of providing appropriate credit and citing the original publications.

Citation Information

When utilizing this dataset in research or derivative works, please cite the foundational TCIA publications:

@article{armato2011lung,
  title={The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A completed reference database of lung nodules on CT scans},
  author={Armato III, Samuel G and McLennan, Geoffrey and Bidaut, Luc and McNitt-Gray, Michael F and Meyer, Charles R and Reeves, Anthony P and Zhao, Binsheng and Aberle, Denise R and Henschke, Claudia I and Hoffman, Eric A and others},
  journal={Medical Physics},
  volume={38},
  number={2},
  pages={915--931},
  year={2011},
  publisher={Wiley Online Library},
  doi={10.1118/1.3528204}
}

@article{clark2013cancer,
  title={The Cancer Imaging Archive (TCIA): Maintaining and operating a public information repository},
  author={Clark, Kenneth and Vendt, Bruce and Smith, Kirk and Freymann, John and Kirby, Justin and Koppel, Paul and Moore, Stephen and Phillips, Shawn and Maffitt, David and Pringle, Michael and others},
  journal={Journal of Digital Imaging},
  volume={26},
  number={6},
  pages={1045--1057},
  year={2013},
  publisher={Springer},
  doi={10.1007/s10278-013-9622-7}
}

Repository Citation

If referencing this Hugging Face repository package:

Kotomiya07. (2026). LIDC-IDRI Image Collection on Hugging Face.
URL: https://huggingface.co/datasets/Kotomiya07/LIDC-IDRI