Title: {\lambda}Split: Self-Supervised Content-Aware Spectral Unmixing for Fluorescence Microscopy

Abstract:

In the field of fluorescence microscopy, spectral unmixing is the process of isolating individual fluorophore concentrations from images where emissions from multiple fluorophores are mixed. Traditional techniques, which typically perform pixel-wise operations using least-squares fitting, tend to struggle as emission spectra overlap more significantly and noise levels rise. This limitation suggests that adopting a data-driven strategy, capable of leveraging structural priors, could yield superior outcomes. While learning-based solutions for spectral imaging are available, they are often either ill-suited for microscopy-specific data or designed for narrow applications that do not translate to fluorescence microscopy contexts.

To overcome these challenges, we introduce {\lambda}Split, a physics-informed deep generative model. This approach utilizes a hierarchical Variational Autoencoder to learn a conditional distribution across concentration maps. A fully differentiable Spectral Mixer ensures alignment with the physical image formation process, while the embedded structural priors facilitate state-of-the-art unmixing performance alongside implicit noise reduction.

We validated {\lambda}Split using three real-world datasets, which were synthetically transformed into 66 rigorous spectral unmixing benchmarks. Our evaluation included comparisons against ten baseline methods, encompassing both classical algorithms and various learning-based techniques. The results demonstrate that {\lambda}Split consistently achieves competitive performance and exhibits enhanced robustness, particularly in scenarios involving high noise, significant spectral overlap, or reduced spectral dimensionality. Consequently, {\lambda}Split establishes a new state-of-the-art for spectral unmixing in fluorescence microscopy data. Notably, the method is compatible with spectral data generated by standard confocal microscopes, allowing for immediate implementation without the need for specialized hardware modifications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC