Skip to content

TIGRE 3.2.0

Latest

Choose a tag to compare

@github-actions github-actions released this 11 Nov 18:29
· 139 commits to master since this release
a7d84b5

Changed

  • Python 3.11 or newer is required; Python 3.10 is no longer supported (#825).
  • TIGRE is no longer published to PyPI. The tigre name there belongs to an unrelated
    package, so pip install tigre does not install TIGRE. Install from source or with
    pip install git+https://github.com/CERN/TIGRE.git; built wheels are attached to each
    GitHub Release (#819).
  • FDK/FBP filtering in Python runs in cache-sized blocks on several threads (#790).
  • FISTA chooses its step size automatically by default (power iteration) (#769).
  • MATLAB SART, SIRT and PCSD are now thin wrappers over shared code (#786).
  • Parker weighting detects near-360° scans more robustly (#743).
  • FDK no longer applies Wang detector-offset weights on short scans (#778).

Added

  • Parallel-beam Siddon projector supports full Euler angles (alpha, theta, psi) (#792).
  • MATLAB plotSinogram (#771).
  • Continuous integration: the Python test suite runs on every push and pull request (#816).
  • GPU test suite for the projectors and every reconstruction algorithm, run locally with
    pytest -m gpu Python/tests (#821).
  • CITATION.cff, a rewritten contributing guide and a pull request template (#818, #817).

Fixed

  • minTV/AwminTV (used by ASD-POCS, PCSD, PICCS and the TV variants) could
    intermittently return an all-zero volume: a race between cudaMemset and stream
    work. The same pattern is fixed in the projectors and backprojectors (#823).
  • Windows source builds failed with setuptools 84 or newer (CUDA files were sent to
    the C++ compiler) (#822).
  • Multi-GPU TV halo exchange: synchronization and offsets (#804).
  • Parallel-beam projections had stripe artifacts from wrong intersection bounds (#752).
  • Siddon projector could hang on rays that miss the volume (#779).
  • TV/PICCS minimizers destroyed the caller's CUDA context on success, and leaked
    GPU memory; IRN-TV-CGLS failed under NumPy 2 (#770, #800).
  • PICCS: could hang computing W, and other Python PICCS fixes (#784, #794).
  • POCS methods (ASD-POCS, PICCS, PCSD): lambda_red never reached the data step (#766).
  • Krylov methods: float64 norm accumulation, working restarts, results no longer
    discarded (#774).
  • Iterative algorithms: per-angle geometry indexing (MATLAB) and block size 1
    (Python) (#801, #742).
  • Parker weights took the fan angle from the wrong axis (#767).
  • im3DNORM returned None for supported norm orders (#805).
  • Measure_Quality UQI now uses the N-1 variance, matching MATLAB (#796).
  • FBP filter parameter was not propagated (#760).
  • Geometry: single-angle input under NumPy 2, and cast_to_single turning integers
    into floats (#749, #750).
  • im3ddenoise rejects 2D input with a clear error (#748).
  • Varian loader rotation direction and XIM parsing (#762).
  • gpuids=None now means all GPUs everywhere; helical demo offset axis; computeCOR
    port (#773).
  • Importing a compiled module (e.g. _minTV) before tigre failed with a circular
    import; removed a deprecated numpy.core import (#824).
  • Plotting: one-dimensional and ZYZ angle arrays, empty savegif, MATLAB quality
    histories (#812, #746, #813).
  • Clearer error when the bundled head phantom is missing (#811).
  • Demos: duplicate 2π projections, quality demo inputs, geometry demo header (#759, #810, #761).
  • Build: compute capability table and detection, NumPy build pin for manylinux2014,
    documented GCC version (#732, #738, #798, #799).

Removed

  • Read the Docs configuration (the docs it pointed to no longer exist) and other stale
    files (#817).

Full changelog: changelog.md