Synthetic paired PA/LAT chest radiographs with projected nodule masks from LIDC-IDRI CT

GEOMETRIC_LUNG_RX_PAIRS_LIDC v1.0.0

GEOMETRIC_LUNG_RX_PAIRS_LIDC v1.0.0 is a synthetic radiographic dataset derived from selected LIDC-IDRI thoracic CT cases. It contains paired posteroanterior (PA) and lateral (LAT) chest-radiograph-like TIFF images generated from CT volumes, together with corresponding projected 2D nodule masks.

Dataset contents

The archive contains 683 generated PA/LAT image pairs. For each generated case, the dataset provides:

• data/images/<case_id>/PA.tif
• data/images/<case_id>/LAT.tif
• data/masks/<case_id>/PA_nodule_mask.tif
• data/masks/<case_id>/LAT_nodule_mask.tif
• Metadata files under metadata/, including pairs.csv and lesions.csv

Image and mask format

Images are stored as uint16 TIFF files. Nodule masks are stored as uint8 TIFF files, with:

• 0 = background
• 255 = nodule

The PA and LAT masks are projected from the same 3D CT nodule segmentation into the corresponding radiographic view.

Nodule annotation processing

The source segmentation files were binary masks. Some contained tiny disconnected voxel fragments. For consistency, the released nodule annotation retains the largest connected component of each binary segmentation as the single nodule.

Image intensity and display polarity

The final TIFF images use conventional attenuation display polarity: larger projected attenuation is brighter, while lower attenuation and background are darker.

The TIFF photometric interpretation is MINISBLACK, meaning that:

• pixel value 0 should be displayed as black
• pixel value 65535 should be displayed as white

Generation methodology

CT volumes were first canonicalized to RAS orientation. Hounsfield Units (HU) were clipped and converted to linear attenuation coefficients using a monoenergetic 60 keV water coefficient.

Parallel-beam PA and LAT projections were then generated by line integration, with voxel spacing converted from millimeters to centimeters:

A = sum(mu_cm_inv * ds_mm / 10)

Final display images were generated using the selected V5 CXR-like deterministic postprocessing profile and SAMPLING_B detector sampling, while preserving the physical aspect ratio.

The released archive therefore contains the generated radiographic image pairs, their corresponding projected nodule masks, and the associated metadata required to interpret the dataset.

This dataset is derived from LIDC-IDRI data obtained from The Cancer Imaging Archive (TCIA). The original LIDC-IDRI collection is distributed by TCIA under Creative Commons Attribution 3.0 and requires data citation. Users of this derived dataset should also cite the original LIDC-IDRI TCIA dataset.