Abstract
Hyperspectral imaging is a powerful technique to simultaneously study multiple fluorophore labels with overlapping emissions. Here we present a computational hyperspectral imaging method, which uses the sample spatial fluorescence information as a reconstruction constraint. Our method addresses both the under-sampling issue of compressive hyperspectral imaging and the low throughput issue of scanning hyperspectral imaging. With simulated and experimental data, we have demonstrated the superior reconstruction precision of our method in two and three-color imaging. We have experimentally validated this method in differentiating cellular structures labeled with two red-colored fluorescent proteins, tdTomato and mCherry, which have highly overlapping emission spectra. Our method has the advantage of totally free wavelength choice and can also be combined with conventional filter-based sequential multi-color imaging to further expand the choices of probes.
Introduction
In modern fluorescence microscopy, the ability to visualize multiple cellular components simultaneously is crucial for understanding complex biological interactions. However, researchers are frequently hindered by "spectral overlap." Most common fluorophores possess broad emission spectra; when more than four colors are utilized, their signals inevitably bleed into one another, creating crosstalk that obscures biological detail.
Traditional solutions present significant trade-offs. Scanning hyperspectral imaging offers high accuracy but is limited by slow acquisition speeds. Conversely, compressive hyperspectral imaging allows for snapshot acquisition but often suffers from low spatial resolution and artifacts due to the "under-sampling" of information onto a 2D sensor. This paper introduces a novel middle ground: Space-Constrained Computational Hyperspectral Imaging, designed for high-precision multicolor separation without the speed bottlenecks of scanning systems.
Methods: The Dual-View Architecture
The core of this method is a "Dual-View" detection scheme that splits the microscope's emission light into two distinct paths.
As shown in the figure above, Path 1 (Spatial) captures an undispersed image, providing a high-fidelity map of the sample's structure. Path 2 (Spectral) passes the light through a wedge prism, which induces a lateral shift proportional to the wavelength.
The "secret sauce" lies in the reconstruction algorithm. Rather than treating the dispersed spectral image as an isolated puzzle, the system utilizes the spatial image from Path 1 as a mathematical constraint. By employing the Two-step Iterative Shrinkage/Thresholding (TwIST) algorithm with Total Variation (TV) regularization, the system forces the spectral reconstruction to align with the known physical structures of the sample. This approach effectively anchors the spectral data to spatial coordinates, significantly improving the accuracy of fluorophore assignment.
For more complex scenarios involving three or more colors, the researchers integrated a Digital Micromirror Device (DMD) into the excitation path. The DMD generates random illumination patterns, providing the algorithm with additional independent "equations" to solve for overlapping fluorophore distributions—a technique derived from compressive sensing.
Key Results and Validation
The researchers validated their approach through rigorous simulation and experimental testing.
In simulations using "ground truth" images of microtubules and mitochondria labeled with EGFP and EYFP (peaks separated by only 20nm), the method demonstrated remarkable robustness. As illustrated in the RMSE plots, the algorithm maintained a relative Root Mean Square Error of less than 0.1 even at extremely low signal levels (100 photons) and high background noise.
The system's real-world utility was proven by imaging BSC-1 cells where Clathrin was labeled with tdTomato and Vimentin with mCherry. These two red fluorescent proteins are notoriously difficult to distinguish because their emission peaks are separated by only 19nm. While standard filter-based imaging would result in an inseparable red blur, this space-constrained method successfully resolved the distinct punctate structures of clathrin from the filamentous network of vimentin.
Conclusion
By leveraging spatial information as a reconstruction anchor, this method overcomes the traditional limitations of hyperspectral imaging. It provides a high-fidelity way to separate fluorophores with highly overlapping spectra, offering researchers total freedom in wavelength choice. While the method performs best on samples with distinct structural features, its ability to be combined with conventional filter-based imaging suggests a powerful path forward for expanding the palette of probes available for high-speed, multicolor biological imaging.
Summary updated