2DeteCT¶
A dataset of measured CT data for machine learning: fan-beam sinograms of 5,000 slices of 109 objects (fruit, nuts and other materials), each measured in three modes, with reference reconstructions and segmentations (Kiss et al., “2DeteCT - A large 2D expandable, trainable, experimental Computed Tomography dataset for machine learning”, Scientific Data, 2023). It is distributed on Zenodo.
Tube voltage |
90 kV |
90 kV |
60 kV |
Tube power |
3 W |
90 W |
60 W |
Filter |
Thoraeus |
Thoraeus |
none |
Meaning |
low dose |
high dose |
beam hardening |
Download and prepare¶
Downloading takes a few hours:
python LION/data_loaders/2deteCT/download.py
python LION/data_loaders/2deteCT/pre_process_2deteCT.py
The first writes LION_DATA_PATH/raw/2detect, the second
LION_DATA_PATH/processed/2detect; the raw folder can then be deleted.
In LION¶
The loader is LION.data_loaders.deteCT.deteCT. Its parameters
(deteCT.default_parameters()) choose:
the modes:
input_mode(the measurements) andtarget_mode(the reference reconstruction), mode2 for both by default;the task: “sino2recon” (sinogram, reconstruction) and others, e.g. for segmentation;
the split: samples in order, 80%, 10%, 10%;
corrections: flat- and dark-field correction and the log transform, on by default;
added noise (
add_noise,noise_params): Poisson noise and cross-talk on top of the measured sinograms, as Kiss et al. (2024) did.
deteCT.get_default_geometry() is the scan’s geometry: fan beam, 3600
angles over 360 degrees, a 956-pixel detector (2× subsampled), 1024×1024
images.
Note
The data were acquired with two detectors: samples 1–63 with one,
64–109 with the other, which gets about a third of the photons at the same
settings. Split in order, the validation and test sets are all from the
second detector and the training set mostly from the first. The data
record lists each sample’s detector, and noise_params.I0 can be one
value per detector (deteCT.noise_parameters()).
from LION.experiments import ct_experiments, deteCT_benchmark
# LION's experiments, on the measured data (only the angles the scan has)
experiment = ct_experiments.clinicalCTRecon(dataset="2DeteCT")
# the 2DeteCT benchmark paper's experiments
experiment = deteCT_benchmark.LimitedAngle90CTRecon()
training = experiment.get_dataset("train")
Two sets of experiments use it: LION’s (ct_experiments, with
dataset="2DeteCT": clinical dose is the measured mode2, extremely low dose
mode1; the low-dose experiments have no measured equivalent) and
those of the 2DeteCT benchmark paper (deteCT_benchmark: Kiss et al.,
“Benchmarking learned algorithms for computed tomography image reconstruction
tasks”, 2025), which reproduce it and are not changed.