Second-harmonic generation (SHG) microscopy has advantages in biomedical research, such as low phototoxicity along with high resolution and deep penetration of tissue. Michael Zachariadis, who manages instruments in the Imaging Facility at the University of Bath, said their lab has been used for identification of collagen structure in swine or the tracking of cellular interactions in zebrafish.
The utility comes from the unique nature of the technique. In SHG, two photons sharing the same frequency interact with a material and are combined into one photon with double the energy and frequency. No energy is lost in the transfer. Zachariadis said that the lab at Bath provides a point-scanning system that integrates a ZEISS confocal microscope and a multiphoton laser from MKS.
Researchers from multiple disciplines have imaged biomolecules with non-centrosymmetric structures (where one side does not mirror the other perfectly) without staining. This effect has been used for the live imaging of contracting muscles and action potentials in neural membranes.
And machine learning algorithms can enhance the data that SHG microscopy provides, which Arash Aghigh and François Légaré describe in this issue’s cover story here. The duo from the Institut national de la recherche scientifique in Québec City write about SHG and polarization-resolved SHG (P-SHG), which reveal insights about the alignment of fibrous structures in the extracellular matrix — a key indicator of some conditions such as fibrosis and cancer.
Trade-offs in these approaches include factors such as image quality, imaging time, and the required laser power. High laser power could damage a sample while low laser power could produce blurry images. And the interpretation of images has typically been performed manually by researchers, which is time-consuming.
By integrating machine learning into the workflow, processing and analysis of images — coupled with effective denoising — can be more efficient. Traditional denoising involves the use of filters, which can inadvertently remove key details, whereas machine learning can provide content-aware image restoration. The authors describe two approaches, including CARE 2D, which pairs noisy and high-quality reference images for training, and Noise2Void, which is trained based on noise patterns in the images themselves.
And to accelerate imaging over large tissue areas, they point to the advancement of enhanced superresolution generative adversarial networks (ESRGANs), which use two neural networks that convert low-resolution images to high-resolution images and then distinguishes between them. These combined capabilities have the potential to rapidly provide researchers with a treasure trove of information, which has not been lost on the scientific community.
Enjoy the issue!
Douglas J. Farmer
