Chan Zuckerberg Initiative
A differentiable wave-optical framework for label-agnostic computational microscopy of biomolecular density and orientation across diverse imaging modalities.
WaveOrder is a computational microscopy framework developed by the Computational Microscopy Platform (Mehta Lab) at CZ Biohub San Francisco. It provides a unified, physics-informed, PyTorch-based library for designing and solving inverse problems in biological light microscopy — spanning quantitative phase imaging (QPI), fluorescence deconvolution, polarization-resolved imaging, and label-free modalities — within a single coherent mathematical framework. The paper describing the framework was posted to arXiv in December 2024.
The core challenge WaveOrder addresses is that modern biological microscopy operates across a wide range of physical principles: widefield fluorescence, confocal sectioning, light-sheet illumination, oblique illumination, and polarization-resolved imaging all require distinct physical models and reconstruction pipelines. WaveOrder unifies these under a differentiable wave-optical formalism, representing all linear optical properties with a shared mathematical language that allows gradient-based optimization to be applied across modalities. Rather than relying on large labeled training datasets, the framework uses physics-informed machine learning to auto-tune model parameters — including those governing spatially varying aberrations — and to solve blind shift-variant deconvolution problems without supervision.
WaveOrder has been validated across scales ranging from subcellular organelles to whole adult zebrafish, demonstrating its versatility in biological imaging contexts. It is deeply integrated into the CZ Biohub computational microscopy stack through the companion recOrder napari plugin, and both tools share OME-Zarr as their native data format. A model card and quickstart notebook are available on the CZ Virtual Cells Platform, positioning WaveOrder as a reproducible computational microscopy building block for the broader CZI research community.
recOrder) enables interactive parameter exploration and visualization, while a configuration-file-driven CLI supports batch processing on HPC clusters, with both reading and writing OME-Zarr.WaveOrder is implemented in Python using PyTorch, enabling GPU acceleration and autodifferentiation through all forward model components. The framework represents the optical system through a linear transfer function model that encodes the pupil function, aberrations, polarization state, and illumination geometry. This representation is differentiable by design, so physical parameters can be optimized jointly with reconstruction weights via backpropagation. For fluorescence deconvolution, the forward model generates a spatially varying PSF from a parametric aberration expansion, and the corresponding inverse problem is regularized with standard priors (total variation, positivity). For label-free phase and polarization reconstruction, a Jones calculus-based model maps specimen birefringence and phase to raw intensity measurements across polarization states.
The library supports input data in OME-Zarr format, a community standard for multi-dimensional biological imaging that supports chunked, cloud-accessible storage. Benchmarks in the preprint demonstrate that the physics-informed approach recovers biomolecular structure beyond the limits of prior phase and polarization reconstruction methods, with improvements validated through both synthetic data with known ground truth and biological specimens including fixed cells, live cells, and whole zebrafish. The codebase is modular: core wave optics (waveorder), label-free imaging (recOrder), and the napari plugin can be installed separately or together, with all components publicly available on PyPI and GitHub.
WaveOrder is used by microscopists and computational biologists seeking quantitative structural information from label-free or multi-contrast imaging experiments. In the context of the CZ Virtual Cells Platform, WaveOrder provides the phase and fluorescence deconvolution preprocessing that feeds downstream cell modeling pipelines such as Cytoland. Researchers in cell biology use the quantitative phase channel to measure dry mass distributions and track organelle morphology without phototoxic dyes. The polarization reconstruction capability supports studies of cytoskeletal architecture, chromatin organization, and fibrillar extracellular matrix, where molecular alignment is the biologically relevant quantity. Core facilities and imaging centers can use the CLI for large-scale batch deconvolution of multi-well plate datasets or light-sheet acquisitions, taking advantage of GPU acceleration for throughput.
WaveOrder consolidates what has historically been a fragmented ecosystem of modality-specific reconstruction tools into a single well-maintained open-source package under the CZ Biohub computational microscopy umbrella. By grounding reconstruction in differentiable physics rather than supervised deep learning, it sidesteps the data requirements and generalization limitations that hamper purely data-driven approaches in low-throughput biological imaging contexts. As of early 2026, the framework remains a preprint awaiting peer review, and independent benchmarking of the physics-informed parameter estimation against dedicated calibration workflows has not yet been published. The tight integration with the CZ Virtual Cells Platform and napari ecosystem positions WaveOrder to become a standard preprocessing component for quantitative cell imaging pipelines as the platform matures.