Spectral Unmixing: A modular and reproducible Python package for directed and blind spectral unmixing in multidimensional microscopy stacks

Abstract

Background: Spectral bleed-through is a persistent source of bias in multichannel fluorescence microscopy, where signal from one fluorophore is recorded in another detection channel. Routine correction workflows often remain fragmented across manual graphical procedures, laboratory-specific scripts, or method-specific blind-unmixing implementations with limited provenance. Results: We present spectral-unmixing, an open-source Python package for reproducible directed and blind bleed-through correction in multidimensional microscopy stacks. The package combines directed two-channel correction with multiple coefficient-estimation strategies, optional bidirectional two-channel correction through explicit inversion of a two-by-two mixing model, and PICASSO-family blind unmixing for multichannel data. Input stacks from multiple microscopy file formats are normalized to a canonical axis order before processing, reducing the need for manual pre-formatting. Each processing run also writes a machine-readable sidecar file that records the effective configuration and estimated coefficients. Synthetic and real-data-derived benchmarks demonstrated robust correction across directed, bidirectional, and blind-unmixing settings. In fixed-alpha two-channel simulations, directed correction reduced target-channel normalized root mean squared error from approximately 0.029 to about 0.003. In time-varying data, per-time-point estimation reduced mean absolute alpha error from approximately 0.099 to 0.003 compared with reference-time-point estimation, while bidirectional inverse-model correction reduced reciprocal-mixture channel errors from approximately 0.022 to 0.037 to about 0.004. Multichannel benchmarks further showed that blind-unmixing workflows can reduce residual inter-channel dependence while preserving fluorophore identity, and that source-sink priors provide a controllable alternative when plausible contamination pathways are known. Conclusions: Spectral-unmixing provides a modular, scriptable, and extensible platform that standardizes directed, bidirectional, and blind spectral-unmixing workflows, thereby lowering barriers to reproducible bleed-through correction in quantitative bioimage analysis.