Skip to content

Fusion

Fusion combines the pre-transformed views into a single output image. Built-in methods are selected with fusion.fuse(..., fusion_func=..., weights_func=...). Custom fusion and weight functions are described in the Extension API section.

Fusion methods

weighted_average_fusion (default)

  • Uses blending weights with optional extra fusion_weights.
  • Good general-purpose choice for smooth transitions across overlaps.

simple_average_fusion

  • Averages valid pixels (ignores NaNs), without blending weights.
  • Useful as a baseline when weights are not desired.

max_fusion

  • Pixel-wise maximum across views.
  • Useful for sparse or bright features where max intensity is preferred.

multi_view_deconvolution

Implements the efficient Bayesian multi-view deconvolution method from:

Preibisch et al., "Efficient Bayesian-based multiview deconvolution", Nature Methods 11, 645–648 (2014). https://doi.org/10.1038/nmeth.2929

Iteratively refines a single fused estimate ψ by applying, for each view in turn, a forward blur, a ratio correction, and a back-projected multiplicative update. Blending weights gate updates so that only views with genuine coverage contribute at each pixel.

Key parameters (passed via fusion_func_kwargs):

Parameter Default Description
psfs None List of per-view PSF arrays (one per view). When None, a Gaussian PSF is estimated from output_spacing, na, and wavelength_um.
psf_type "EFFICIENT_BAYESIAN" Compound back-projection kernel variant: "EFFICIENT_BAYESIAN" (most accurate), "OPTIMIZATION_I", "OPTIMIZATION_II", or "INDEPENDENT" (standard RL, no cross-view coupling).
n_iterations 10 Number of deconvolution iterations.
lambda_reg 0.0 Tikhonov regularisation strength; 0 disables it. Values in 1e-4–1e-2 give mild smoothing.
output_spacing None Physical pixel spacing (µm) keyed by dimension, used for automatic PSF estimation.
na 0.8 Numerical aperture for PSF estimation.
wavelength_um 0.5 Emission wavelength in µm for PSF estimation.
sample_boundary_erosion_px 0 Pixels to erode the union coverage mask before zeroing output; removes the bright-ring artefact at the outer sample boundary.

GPU support: When the input arrays are cupy arrays, convolutions run on the GPU via cupyx.scipy.ndimage. Use backend="cupy" in fusion.fuse to enable this automatically.

Chunk overlap: multi_view_deconvolution declares a required_overlap equal to the PSF half-width, ensuring chunked fusion does not introduce seam artefacts at block boundaries.

Minimal usage example:

from multiview_stitcher import fusion
from multiview_stitcher.fusion import multi_view_deconvolution

fused = fusion.fuse(
    images=sims,
    transform_key="affine_registered",
    fusion_func=multi_view_deconvolution,
    fusion_func_kwargs=dict(
        n_iterations=10,
        output_spacing={"z": 0.5, "y": 0.13, "x": 0.13},
        na=0.8,
        wavelength_um=0.52,
    ),
    backend="cupy",  # optional – omit for CPU-only
)

Blending weights

fusion.fuse computes smooth blending weights using weights.get_blending_weights. The falloff is controlled by blending_widths (physical units per dimension) and normalized across views.

Fusion weights

Additional per-view weights can be supplied via weights_func when the fusion method accepts fusion_weights.

  • weights.content_based: Content-based weights (Preibisch et al.) computed from local contrast; configure with sigma_1 and sigma_2.