21 Jan 2025
Deep learning image transformation bridges the limitations of different modalities.
A team at South Korea's Pohang University of Science and Technology (POSTECH) has developed a way to process low-resolution label-free photoacoustic images into high-resolution virtually stained images resembling those generated by confocal fluorescence methods.Confocal fluorescence microscopy (CFM) is widely used to make high-resolution cellular images but requires fluorescent staining, which poses risks of photobleaching or phototoxicity, potentially damaging the cells under study.
Conversely, mid-infrared photoacoustic microscopy (MIR-PAM) allows for label-free imaging preserving cell integrity. Its reliance on longer wavelengths limits spatial resolution, however, making it difficult to visualize fine cellular structures with precision.
"MIR-PAM provides low-resolution protein-selective imaging in unlabeled cells using a monochromatic wavelength, whereas CFM provides high-resolution multiplexed imaging in immunofluorescent stained cells using multiple wavelengths," noted the team in its paper.
To bridge the gap, POSTECH has developed an imaging method based on explainable deep learning (XDL), a computational approach designed to let its algorithms be more easily monitored and followed than closed "black box" methods. XDL "offers enhanced transparency by visualizing the features that contribute most to the outcome, ensuring both reliability and accuracy," said POSTECH.
As described in Nature Communications, the project applied an XDL approach via a two-stage imaging process. First a Resolution Enhancement phase processes MIR-PAM images to distinguish intricate cellular structures such as nuclei and filamentous actin proteins.
Then a Virtual Staining phase produces virtually stained images without fluorescent dyes, eliminating the risks associated with staining while maintaining CFM-quality imaging.
New possibilities for cellular imaging
In trials, POSTECH applied its system to photoacoustic images of human cardiac fibroblasts, and compared the results to corresponding CFM images. The results were "qualitatively and quantitatively similar to the CFM ones," noted the project, with future advances in XDL models likely to improve the performance still further.
One hurdle noted in the project's paper will be the inherently lower performance of the approach when applied to living cells rather than fixed cells, due to the lower contrast provided by photoacoustic images in those scenarios. The project intends to explore a number of optimizations of the platform to address this and improve its suitability for live cell imaging.
But POSTECH already believes its XDL approach to be a cross-domain image transformation technology bridging the physical limitations of different imaging modalities, and offering complementary benefits. The research has also helped to significantly enhance the stability and reliability of unsupervised learning operations, commented the project.
"This research unlocks new possibilities for multiplexed, high-resolution cellular imaging without labeling," said POSTECH's Jinah Jang. "It holds immense potential for applications in live-cell analysis and disease model studies."
© 2025 SPIE Europe |
|