Tomography

Dataset names: tomography_2d and tomography_3d

Both tomography datasets use DeepInv’s TomographyWithAstra physics, backed by the ASTRA Toolbox, but they represent different data and acquisition geometries. The 2D case generates a sinogram from a reference slice, whereas the 3D case uses measured cone-beam data and its supplied scanner trajectory.

2D Data and Inverse Problem

The 2D dataset downloads a Shepp–Logan phantom and resizes it to the requested square image size. It constructs evenly spaced projection angles over 180 degrees, creates an ASTRA parallel-beam operator, and generates noisy projections from the phantom.

When num_operators is greater than one, the angle sequence and sinogram are split into contiguous groups. Each group defines one physics operator and one measurement tensor. Together they still represent the complete acquisition.

3D Data and Inverse Problem

The 3D dataset downloads Walnut-CBCT_8.pt. It contains a dense reference volume, a cone-beam sinogram, and ASTRA geometry vectors describing the source and detector pose for every projection. The physical volume and detector geometry are fixed by the acquisition; resizing either would make the supplied geometry inconsistent. Consequently, image_size is accepted by the BenchOpt dataset interface but ignored when the Walnut data is loaded.

num_projections selects an approximately uniform subset of the stored views. The selected sinogram and matching geometry vectors are split across num_operators. An ASTRA cone-beam operator is then created for each split. The current implementation requires the stored sinogram: setting use_dataset_sinogram to false raises NotImplementedError because a new 3D sinogram is not generated with a forward pass.

What benchopt install Does

Neither tomography dataset declares an additional custom installer. The ASTRA, DeepInv, and PyTorch dependencies are normal project requirements. With the project already installed, benchopt install has no tomography-specific file to install; input data is handled by preparation.

What benchopt prepare Does

For tomography_2d, preparation downloads SheppLogan.png into BenchOpt’s shared tomography data directory. For tomography_3d, it downloads the Walnut .pt archive from the romainvo/ct_examples dataset on Hugging Face. Existing non-empty files are reused. Preparation is cached independently of the parameter grid because all configurations use the same source file.

Available 2D Parameters

image_size (default 512)

Square reconstruction size, or an explicit two-element spatial shape. The loaded phantom is resized accordingly; the final shape must be square.

batch_size (default 1)

Number of phantom images in the batch.

num_operators (default 1)

Number of groups into which the angles and measurements are split. It cannot exceed num_angles.

num_angles (default 100)

Number of evenly spaced projection angles used to generate the 2D sinogram.

num_projections (default 100)

Present in the common tomography interface but not used by the current 2D implementation; use num_angles to control 2D sampling.

noise_level (default 0.01)

Standard deviation of Gaussian noise applied by the 2D ASTRA physics.

seed (default 42)

Base seed for the per-operator 2D noise generators.

geometry_type_2d (default parallel)

ASTRA geometry used for the 2D acquisition.

detector_spacing_2d (default 1.0)

Spacing between detector elements.

pixel_spacing_2d (default 1.0)

Physical spacing of image pixels; a scalar or two-element spacing can be passed through the underlying physics interface.

Available 3D Parameters

image_size (default 512)

Accepted for a uniform dataset interface but ignored for the fixed Walnut volume.

batch_size (default 1)

Number of copies of the stored volume and sinogram.

num_operators (default 1)

Number of contiguous projection groups. It cannot exceed num_projections.

num_projections (default 100)

Number of stored Walnut views selected before splitting.

num_angles (default 100)

Present in the common interface but not used by the 3D path; sampling is controlled by num_projections.

geometry_type_3d (default conebeam)

ASTRA geometry type. The supplied geometry vectors describe the actual cone-beam trajectory.

use_dataset_sinogram (default true)

Use the stored Walnut sinogram. This must currently remain true.

noise_level (default 0.01) and seed (default 42)

Accepted by the dataset interface. They affect generated 2D measurements but are not applied to the stored 3D sinogram in the current implementation.

Example configurations are available under benchmark_inference/configs/examples/tomography_2d.yml and benchmark_inference/configs/examples/tomography_3d.yml.