Nilay Vora

What is Two-Photon Excited Fluorescence Microscopy?

Two-Photon excited fluorescence microscopy in a nonlinear optical imaging technique. Nonlinear optical (NLO) imaging exploits nonlinear light matter interactions that generally occur only at the focal plane, providing intrinsic depth sectioning^4-6. Multiphoton microscopy (MPM) is an NLO imaging technique that leverages the simultaneous absorption of two or more low energy photons that arrive “simultaneously” at the fluorophore⁷ . Owing to the “simultaneous” absorption of low energy photons, the fluorophore achieves a greater energy state^5,7. After a short period of time, the molecule returns to its ground state; however, due to the higher energy state from multiphoton excitation, the emitted fluorescence is of a lower wavelength compared to the excitation wavelength^5,7.

To generate nonlinear optical events, high-powered, ultrafast pulsed lasers (on the order of femtoseconds) are used⁸. Multiple nonlinear optical processes can be generated using these lasers; however, here, we focus specifically on Two-Photon excited fluorescence (TPEF).

TPEF is a process in which two photons are absorbed by atoms within a time window of an attosecond (10-18 s), leading to the molecule reaching an excited electronic state^9,10. Quantum mechanically, a single photon is absorbed, bringing the atoms to a virtual intermediate state before eventually reaching the final excited state by absorbing a second photon⁹. Comparatively, in single-photon fluorescence (SPF), a single photon is absorbed, causing atoms to reach their final excited state (Figure 1)⁹. After reaching the final excited state, vibrational relaxation is observed in both SPF and TPEF before the atoms return to their ground state, emitting a photon (fluorescence).

Another feature of TPEF is the highly confined three-dimensional excitation volume. In SPF, excitation occurs along the entire axial plane constantly⁹. However, in TPEF, nonlinear excitation needs high photon flux (1020-1030 photons/cm²·s) which is only satisfied near the focus (Figure 1)^9,10. This feature of TPEF enables intrinsic depth sectioning and minimized photobleaching in out-of-focus regions.

The final advantage of TPEF compared to SPF is that TPEF typically uses NIR wavelength laser light sources compared to visible wavelength lasers used in SPF. The absorption and scattering of light by tissue are greater in the visible wavelength regime, minimizing the potential penetration depth of light into bulk tissue. NIR light enables deeper penetration of light by reducing the number of scattering events and absorption of the low energy photons⁹. Simultaneously, TPEF mimics behaviors of higher energy SPF allowing for the excitation of endogenous fluorophores that typically require high-energy photon excitation. Further, since a NIR illumination source is used to excite a fluorophore with emission in the visible region, greater separation in excitation and emission bands exist, leading to a higher SNR signal⁹.

Sources of Endogenous Contrast for Nonlinear Optical Imaging

A variety of endogenous biomolecules can be visualized using TPEF in biological samples. These fluorophores include Keratin, Retinol, LipDH, NAD(P)H, FAD, Melanin, Lipofuscin, and electron-transferring flavoprotein (ETF)⁴. Free, unbound FAD is a small portion of all FAD autofluorescence; the majority of FAD signal is found within other flavoproteins like LipDH⁴. In the work described in this project, FAD is used interchangeably to describe autofluorescence detected from all flavin-associated proteins.

The most common sources of intracellular signal in TPEF are NAD(P)H and FAD^4,6. Extracellularly, collagen and elastin fibers contribute significant signals in TPEF due to crosslinking structures within these proteins^11,12. For all analyses of TPEF images in this project, NAD(P)H and FAD signals are used, with collagen and elastin signals being masked. NAD(P)H is commonly excited in TPEF using a 755 nm laser, and emission is collected at 460 ± 25 nm^4,6. FAD is excited in TPEF using a 860 nm laser, and emission is collected at 525 ± 25 nm^4,6. These intrinsic contrast sources have enabled the rapid translation of TPEF from the benchtop to clinical settings^13-16.

To learn more about how these sources of endogenous contrast can be used for cancer diagnostics, look at the link below.

Application of Deep Learning to NLO Images

Many ML and DL networks have been proposed for image restoration of biomedical images. However, most proposed networks have been used to restore image quality in fluorescently stained samples. Exogenously stained sample images feature higher SNR than label-free sample images owing to enhanced contrast from the fluorescence stain. The same networks are not expected to have comparable performance on label-free images which have lower SNR.

DL-enabled label-free image restoration was first explored by Fast et al. (2020) to enhance image contrast and extend the scanning area of a custom in vivo MPM exoscope imaging system²¹. Fast et al. (2020) demonstrated using the content aware image restoration and enhancement (CARE) network sub-micron lateral resolution with a large field of view in a fraction of the time compared to current commercial MPM imaging systems (DermaInspect)²¹. However, no metrics on image quality or model performance were provided. More recently, Shen et al. (2022) presented alternative methods for DL-enabled label-free image restoration of TPEF images²². Shen et al. (2022) demonstrated using a modified generative adversarial network (GAN) ~4.5 dB improvement in peak SNR (PSNR) and 79% improvement in structural similarity index measure (SSIM), two metrics used to assess the quality of a noisy image compared to a clean/ground truth image²². However, tissue images were collected from thin optical sections (OS), featuring consistent SNR between each image compared to bulk tissue, where SNR is expected to decay as a function of depth.

Shen et al. (2022) and Fast et al. (2020) are the only studies to examine DL-enabled restoration of label-free TPEF images^21,22. Despite the observed improvement in image quality after restoration, there is a lack of understanding regarding the impact denoising has on biomedical metrics extracted from these images. Morphofunctional metrics of metabolic activity show immense potential for sensitive (93.3%) and specific (83.3%) detection of cervical precancer¹⁸. It is therefore imperative for DL algorithms to demonstrate recovery of not only image quality but also biomedical metrics. In this project, we cover some of our early efforts to understand the impact of DL-based image restoration on metabolic metric recovery compared to standard image quality metrics and explore how these models could be used for clinical diagnosis of cervical pre-cancer. To read our published work click on the link below.