aluk4/iabi-ct-slices

IABI CT Slice Pack

Clean ground-truth chest CT slices for a university course on inverse problems in biomedical imaging. Derived from the ground-truth images of LoDoPaB-CT.

This pack contains images only — no sinograms. Low-dose measurements are simulated from these images at training time (Beer-Lambert attenuation, Poisson photon counting, log transform), which makes the projection-angle count and the dose free experimental knobs.

Contents

Split	Slices	Shape	dtype
train.npz	2000	128×128	float16
val.npz	250	128×128	float16
test.npz	250	128×128	float16

Each file holds one array under the key images.

import numpy as np
from huggingface_hub import hf_hub_download
path = hf_hub_download("aluk4/iabi-ct-slices", "train.npz", repo_type="dataset")
images = np.load(path)["images"]   # (2000, 128, 128) float16 in [0, 1]

How it was built

Slices sampled across the LoDoPaB-CT ground-truth archives, centre-cropped to square, resized to 128×128 by area-averaging, and normalised per slice to [0, 1]. Only the ground-truth archives were used; the observation archives are not needed, since measurements are simulated.

Splits come from LoDoPaB's own train/validation/test archives, which are patient-disjoint upstream. They are never mixed, so no patient appears in more than one split.

Empty slices are removed. LoDoPaB ships entirely zero ground-truth images as padding at the ends of a scan. Slices with a standard deviation below 0.02 are dropped — they score a perfect PSNR on nothing, and after per-slice normalisation every constant slice becomes the same all-zero image, so one landing in two splits looks exactly like a data leak.

Images are masked to a circular field of view. The course's parallel-beam projector rotates the image and sums columns, so at oblique angles the corners of a square leave the frame and are never measured — corner content loses 75% of its mass at the worst angle. Keeping it would charge every reconstruction method for failing to recover something no method could see. This also matches how a clinical CT console displays a reconstruction.

Limitations

The images are downsampled to 128×128 from LoDoPaB's 362×362. Results on this pack are not comparable to published LoDoPaB-CT benchmark numbers.

Licence and attribution

ODC-By 1.0, inherited from LoDoPaB-CT.

LoDoPaB-CT was created by Leuschner, Schmidt, Baguer and Maass, and its ground-truth images are themselves derived from the LIDC-IDRI chest CT collection. Please credit both in any work using this pack:

Leuschner, J., Schmidt, M., Baguer, D.O., Maass, P. LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruction. Sci Data 8, 109 (2021).

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, 915–931 (2011).