Table 2 Baseline correction algorithms selected for the construction of the multichannel input from the raw spectra.
From: Classification of osteoarthritic and healthy cartilage using deep learning with Raman spectra
Category | Algorithm | Description |
|---|---|---|
Whittaker-smoothing and Spline | Penalized Spline Adaptive Smoothness Penalized Least Squares(PSPLINE asPLS)38,39 | Penalized spline version of asPLS to balance the fidelity and smoothness of the fitted baseline with an adaptive smoothing parameter based on the peak and non-peak regions. |
Morphological | Joint Baseline Correction and Denoising (JBCD)40 | Uses mathematical morphological operations along with regularised least-squares fitting for the removal of baseline distortion and the estimation of a smooth spectrum. |
Smoothing | Range Independent Algorithm (RIA)41 | A range independent background-subtraction algorithm that iteratively applies a Savitzky-Golay smoothing method (moving point average) on the spectra. This gradually eliminates the high frequency peaks, allowing the broad underlying baseline to be subtracted from the raw spectrum, thus yielding the true signal. |
Classification | Fully Automatic Baseline Correction (FABC)42 | It relies on the automatic recognition of signal-free regions to implement a Continuous Wavelet transform algorithm combined with the Whittaker smoothing algorithm for baseline modelling. It can automatically flatten the spectra with significant baseline distortion and is robust against spectra with low signal-to-noise ratios and varying widths. |
Optimizer | Adaptive MinMax43 | It selects the subtraction technique based on the fluorescence-to-signal ratio, effectively reducing RMS error while dealing with different fluorescence-to-signal ratio. |
Polynomial | Goldindec44 | An iterative algorithm that generates parameters automatically from raw data to fit the baseline without being affected by large peaks, peak number or wavenumber. |
Miscellaneous | Baseline Estimation And Denoising with Sparsity (BEADS)45,46 | It performs baseline correction and noise reduction by modelling the baseline as low-pass signal and the noise as high-pass contribution, while the peaks are considered as sparse with sparse derivatives. |
