SNR enhancement in brillouin microspectroscopy using spectrum reconstruction

Abstract

Brillouin imaging suffers from intrinsically low signal-to-noise ratios (SNR). Such low SNRs can render common data analysis protocols unreliable, especially for SNRs below ∼ 10. In this work we exploit two denoising algorithms, namely maximum entropy reconstruction (MER) and wavelet analysis (WA), to improve the accuracy and precision in determination of Brillouin shifts and linewidth. Algorithm performance is quantified using Monte-Carlo simulations and benchmarked against the Cramér-Rao lower bound. Superior estimation results are demonstrated even at low SNRS (≥ 1). Denoising was furthermore applied to experimental Brillouin spectra of distilled water at room temperature, allowing the speed of sound in water to be extracted. Experimental and theoretical values were found to be consistent to within ±1% at unity SNR.

Background-free fibre optic Brillouin probe for remote mapping of micromechanics

“…The surface profile of the droplet can be visualised by simply displaying the raw spectra obtained along the direction of scanning, as heat maps in Figure 3B. Figure 3b are such heat maps showing the Anti-Stokes peaks of the spectra that are generated by first denoising the spectral ROI [27]. It can be seen that there is a clear contrast that identifies the hydrogel region even with the coarser scan.…”

Recent progress and current opinions in Brillouin microscopy for life science applications

“…The problem of assessing the frequency shift of a Brillouin peak amid different noise sources is conceptually the same as that encountered in localization fluorescence microscopy, and similar experimental and data analysis guidelines may thus be followed (e.g., mapping the peaks to correspond to 2-3 pixels on the detector array). In this regard, specific "denoising" approaches based on maximum entropy reconstruction and wavelet analysis have been shown to also be effective for analysis of poor signal-to-noise Brillouin scattering spectra (Xiang et al 2020), for which practical limits on the extractable information can be derived (Török and Foreman 2019). An alternative approach, which is yet to be demonstrated for BM but likely holds considerable potential, involves denoising by convolution of the spectrum with a Lorentzian function (Farahani et al 2011).…”

Multivariate analysis of Brillouin imaging data by supervised and unsupervised learning

“…In this work, minimal pre-processing was Note that the plot for the BF image is not directly comparable to the hyperspectral equivalents and is included for reference only applied to reflect the native performance of the various algorithms presented. As the accuracy of the information obtainable from Lorentzian fitting is sensitive to the amount of Rayleigh background 12 present, background removal was applied wherever possible to optimise the accuracy of the univariate approach. Along with an interferometric filter in the optical path, 54 a software filter was also applied on raw spectral data acquired from the phantom that suffered from a moderate level (unsaturated) of background.…”

“…Further improvements are possible by tuning this threshold and also by denoising the images before classification. 12 While the two cells did not necessarily have identical spectral parameters, the Brillouin properties of PBS were expected to stay constant in the sample. It was observed that with the errors considered, there are still small differences between values obtained from HCA and VCA for PBS.…”

“…This protocol, however, has been shown to rely heavily on a good signal-to-noise ratio (SNR) and even then may still suffer from bias due to noise and complex sample structures. 12,13 Moreover, using least-squares fitting to form a typical Brillouin image, which may easily comprise up to 10 4 pixels (spectra), quickly becomes a time-consuming procedure. The use of spectral phasor and moment, 14,15 have been proposed to improve the accuracy of fitting a single peak to a noisy, multi-component spectral feature, with the former being truly fitting-free and relying on faster computational processes.…”