Interactive SELMA3D comparison#

The napari demo turns the flagship held-out microvessel study into an interactive comparison without redistributing source microscopy images. A curtain slider reveals SplineOps from the left while one selected method remains on the right.

Install and launch#

Install the optional demo dependencies and run the installed command:

python -m pip install 'splineops[selma3d-demo]'
splineops-selma3d-demo

The default installation compares SplineOps with SciPy, scikit-image, and SplineOps cubic interpolation without antialiasing. Add the optional PyTorch area comparison with:

python -m pip install 'splineops[selma3d-demo-all]'
splineops-selma3d-demo --comparisons torch_area scipy_gaussian skimage_resize

From a repository checkout, the compatibility launcher is also available:

python demos/selma3d_napari.py

The first run downloads only the selected WGA channel and expert vessel mask from BioImage Archive accession S-BIAD1196. It verifies pinned SHA-256 digests and caches the files under ~/.cache/splineops/selma3d. Use --help to select another held-out patch, restrict comparison methods, or provide an existing VessAP_vessel directory.

What the viewer shows#

  • Reveal SplineOps moves from comparison-only at 0% to SplineOps-only at 100%.

  • The method selector includes SciPy Gaussian+cubic, scikit-image cubic antialiasing, and SplineOps cubic without antialiasing. PyTorch area is available with the selma3d-demo-all installation.

  • The expert mask is composited on each method’s documented endpoint or half-pixel sampling grid.

  • Local single-patch timings and ROC AUC are displayed separately from the frozen 18-patch evidence.

  • Napari’s first dimension slider browses all 50 retained planes.

Patch 013 is the default because it gives clear visual separation. It is a post-study presentation choice, not a newly held-out selection. The 18-patch aggregate result remains the primary evidence.

Scope and data terms#

ROC AUC measures expert-labelled vessel/background voxel ranking; it is not a segmentation metric. The result covers one channel, one lateral twofold geometry, named methods, and the recorded machine. Read Claims and evidence and 3-D microvessel quality-at-speed validation before quoting the result.

The official source records contain conflicting CC BY 4.0 and CC BY-NC metadata. SplineOps follows the stricter CC BY-NC interpretation and redistributes no source images. The packaged demo contains only checksums and a non-image aggregate evidence summary.