eXplore CT 120 Reconstruction

Cone-beam CT reconstruction for the GE eXplore CT 120 micro-CT scanner. Input is a scan folder as the scanner writes it, output is an HU-calibrated .vff volume.

The families share scan loading, preprocessing, geometry and HU calibration and output format, so their volumes compare directly.

Install

pip install eXplore_CT_120_reconstruction_algorithms          # import name: explore_ct120_recon
pip install "eXplore_CT_120_reconstruction_algorithms[gui]"   # with the web page

Optional: astra-toolbox, TIGRE (pip install "git+https://github.com/CERN/TIGRE.git#subdirectory=Python"), wandb for run logging, and ./install_rasteriser.sh for the Gaussian algorithms. The rasteriser is r2_gaussian's (research licence, not on PyPI), built at a pinned commit that the Gaussian backend was calibrated against; the build is stamped so each run's log says which it used. Python 3.10+, PyTorch 2.0+.

Usage

ct120-fdk data/scans/Scan_1988
ct120-iterative data/scans/Scan_1988 --backend astra --algorithm SIRT3D_CUDA --iterations 100
ct120-iterative data/scans/Scan_1988 --backend tigre --algorithm ossart --tv-lambda 10
ct120-learned data/scans/Scan_1988 --downsample 3
ct120-learned data/scans/Scan_1988 --algorithm gaussian
ct120-learned data/scans/Scan_1510 --algorithm gaussian_dynamic
ct120-volume-report VOLUME.vff
ct120-bilateral-filter VOLUME.vff [SD]
ct120-projection-report data/scans/Scan_1988 --volume a.vff --volume b.vff
ct120-geometry-calibration data/scans/Scan_1988
ct120-noise-calibration data/scans/Scan_1988
ct120-simulate-scan VOLUME.vff --template data/scans/Scan_1988 --output data/scans/Sim_1988
ct120-motion-show RESULT_DIR
ct120-export-4d RESULT_DIR/recon_4d.nrrd --format nrrd-slicer
ct120-ventilation RESULT_DIR
ct120-breathing-phantom build REST.vff --template data/scans/Scan_1510 --output data/scans/Phantom_1510
ct120-gui

--help on any command lists its flags; python -m explore_ct120_recon.run_fdk_recon … is the same as the console script. A scan folder holds the projection .vff files, scan.xml and the bright.vff / dark.vff frames. A run writes a .vff volume next to the scan, a .json sidecar with the grid and the HU map, and a _plots/ folder.

Pipeline

raw projections → flat-field + log → air normalization + ring correction
  → detector rotation calibration (measured once per scan, cached)
  → FDK | ASTRA / TIGRE | learned
  → μ (mm⁻¹) → HU calibration → .vff + .json + plots

Defaults, each with an override flag:

• Field of view and ROI are auto: the grid is what the detector subtends, tightened to the object; the export region is the scanner's own SubVolumeCoordinates.xml box, so results line up with the vendor volume.
• The iterative and learned families reconstruct everything a ray crosses, bed included; otherwise the missing attenuation becomes a smooth HU bias. --roi only crops the output.
• HU calibration fits air at −1000 HU and bulk soft tissue at --tissue-hu (default +120; 0 for a water phantom) from the volume's own histogram.
• Preflight estimates VRAM and RAM first; --preflight-only prints it.
• Weights & Biases logging is on when --wandb-project or $WANDB_PROJECT names a project.

Learning-based reconstruction

Fits a volume model to the line integrals through a differentiable ray tracer: Adam with non-negativity, LR warmup, plateau reduction and a held-out projection for early stopping (--loss, --iterations, --rays-per-batch, --emulate-sart, --compile). All representations share one loss registry. The default wls weights each ray by the inverse variance of its counts, Var(p) = A/count + (sigma_r/ds)^2/count^2: the Poisson slope A is measured from consecutive views on every run, the read noise sigma_r from the scan when possible, else from the detector calibration (--wls-slope, --wls-read-noise pin either). --loss mse is SIRT's objective, for comparisons with the classical solvers.

To add a representation, subclass LearnedReconstructor (build_model, build_domain, export_volume) and register a LearnedAlgorithm; voxel/ is the short example, gaussian/ the full one. A descriptor can change the shared defaults through driver_defaults, and a representation in another package is loaded with --algorithm-module my_package.algorithms.

Dynamic Gaussian splatting (--algorithm gaussian_dynamic)

In a gated step-and-shoot scan of a breathing animal every projection sees a slightly different state. gaussian_dynamic keeps the static cloud as the mean anatomy and gives every projection its own state: a column for the gate plus a small learned code, decoded through a low-rank displacement basis on a 1 mm control grid. The displacement transports each primitive's shape by the deformation gradient and divides its amplitude by |det F|, so mass is conserved; divergence is allowed only in cells containing air; three motion parts with learned membership let the lung slide along the chest wall. All of it is estimated from the projections, with no image registration.

The run loads all gated phases and writes, besides the gate-0 state in recon.vff: recon_4d.nrrd (nine states from gate 0 to gate 1; --dyn-export-4d-format nrrd|nrrd-slicer|nifti, ct120-export-4d converts), recon_ventilation.nrrd with its sidecar (regional specific ventilation at 0.5 mm; ct120-ventilation re-exports), and motion/* panels (ct120-motion-show redraws them from the checkpoint). The --dyn-* flags are in --help. ct120-breathing-phantom builds a gated scan of a breathing digital phantom with a known displacement field and scores a reconstruction against it.

Web page

ct120-gui              # starts the server in the background if needed and opens the page
ct120-gui shortcut     # an icon in the application menu (Start menu, ~/Applications); --desktop, --remove
ct120-gui serve --host 0.0.0.0 --scans-root data/scans --wandb-project my-project   # a lab server, foreground

Closing the browser leaves the server running; the "stop server" button on the page or ct120-gui stop ends it. Jobs never stop with the server: local ones run in their own sessions, remote ones on their machines, and the next server picks them up. The page has no login, so serve it on the lab network only.

New job. One row per flag, generated from the driver's own argparse parser, with the ones changed from the defaults marked; only those are passed, so the recorded command is what a person would have typed. As soon as an input is named, the page checks it on the machine the job will run on: scan.xml, projection count, bright and dark fields, detector serial, cached calibrations; for a volume its header, size and sidecar; for the output whether the folder is writable and the name free. Every path field has a folder button that browses that machine's file system. The form can be pre-filled from the address bar (/new?driver=learned&algorithm=gaussian&input=…&all=1; the viewer takes /view?machine=&path=&window=).

The same form aimed at a Slurm cluster. The #SBATCH directives appear as chips: the teal ones are per-job overrides and can be edited in place, the dashed ones are left to the scheduler; "script" shows the rendered sbatch file. A Slurm job is the machine's template with {job_name}, {log}, {workdir}, {env} and {command} filled in; module loads and the account belong in the machine's own copy.

Jobs. Every run, filterable by status, machine, driver and text.

One job. Command line, machine, GPU or node, output, W&B link, and the log tailed from wherever it runs. A running job can be cancelled; a finished one deleted with or without the files it wrote, which are listed by the run's own naming rule and shown with their sizes first.

Volume. Any .vff on any connected machine, reached from a job's output or by path; here the scanner's own reconstruction of a mouse, shown in the vendor's axis convention. Only the slices on view are read, where the volume lives, so a 2 GB reconstruction on the cluster opens at once; the window is applied in the browser (lung, soft tissue, bone presets or free), with the value under the cursor and a crosshair linking the three planes.

Machines. The machine serving the page, plus any GPU workstation over SSH or Slurm cluster added on the page (SSH destination, Python interpreter with this package, working directory, scans folder, Slurm template), each with a health check. For a cluster the user's whole queue is shown; rows submitted from the page link to their job, and any row can be cancelled. SSH uses the user's own setup, one multiplexed connection per machine, never storing passwords or keys; when a host asks for a second factor, Connect opens the login in a terminal panel on the page. Log out drops the connection.

Job records, logs and machines.toml live in the per-user data directory (~/.local/share/explore_ct120_recon/gui on Linux) or --data-dir; relative paths resolve against --cwd. The page is a React application in gui-frontend/, built into src/explore_ct120_recon/gui/static/ (committed, so pip install ships it); rebuild with npm install && npm run build.

Report tools

ct120-volume-report draws slices, the HU histogram and the fitted anchors of a finished volume, vendor volumes included (--vendor applies GE's axis conventions). ct120-bilateral-filter ports the lab's MATLAB filter_vff.m (imbilatfilt per axial slice, degreeOfSmoothing = 2*SD^2, SD measured from the patch im(20:40,160:180,100) when not given) and writes <volume>_imbilatfilt.vff. ct120-projection-report forward-projects one or more volumes and scores them against the measured projections. ct120-geometry-calibration measures the detector rotation ahead of time. ct120-noise-calibration measures the Poisson slope and read noise; run it once per detector on a dense phantom, since a mouse scan cannot separate the two terms, and the wls loss picks the value up from there.

Everything is importable:

from explore_ct120_recon.ct_core.vff_io import read_vff, write_vff
from explore_ct120_recon.learning_based_iterative import LearnedReconstructor

Layout

src/explore_ct120_recon/
  run_*.py                      drivers, one per family plus the report tools
  ct_core/                      scan loading, preprocessing, geometry and HU
                                calibration, noise model, preflight, VFF I/O
  fdk/                          filtered backprojection
  iterative/{astra,tigre}/      toolbox wrappers
  simulate/                     forward projection of a volume into a scan folder
  learning_based_iterative/     trainer, ray sampler, renderer, losses
    voxel/                      dense grid
    gaussian/                   Gaussian splatting
    dynamic_gaussian/           Gaussian splatting with a deformation field
  gui/                          the web page: API, job store, executors; static/ is the built page
gui-frontend/                   React sources of the page

Calibrations measured from a scan are cached in data/calibration/ when running from a checkout, otherwise in the per-user data directory or $CT_CALIBRATION_DIR.

License

MIT, see LICENSE.