Cosmic spike removal based on multi-feature fusion for long-time integrated Raman spectral detection

Raman spectral detection is commonly performed with a relatively long integration time to suppress the dominant interference of random detecting noise for better signal-to-noise ratio, while inevitably increasing the occurrence probability of intense cosmic spikes that hamper the spectral analysis for substance detection. Moreover, there exist several variants of cosmic spikes other than the typical ultra-high and ultra-narrow peaks, which greatly challenges the discrimination between cosmic spikes and Raman scattering peaks. Thus, there is a great need to improve the spike removal in terms of identification accuracy.

Cosmic spike removal based on multi-feature fusion has been explored to accurately identify and correct the typical cosmic spikes and three main variants, namely narrow but lower spikes, high but wider spikes and overlapping spikes, through a pipeline of four main steps, namely spike rough screening, spike boundary localization, spike fine discrimination and spike interpolating correction. Given the distinctive steepness and narrowness of cosmic spikes, a set of Widths in Pixels was determined as multiple features and then combined with weighting to generate a comprehensive Score for discriminating cosmic spikes from Raman peaks. When tested on Raman spectra of breast tissues and several solid pure substances that are quite different in terms of Raman peak shapes and fluorescence baselines, spike removal of over 3300 cosmic spikes can be accomplished with a high removal rate of 98.8% by using the same weights and thresholds, indicating the good accuracy and versatility.

Based on multi-feature fusion, cosmic spike removal can be efficiently realized through single-spectrum spike identification and correction without compromising any Raman peaks, providing high-quality Raman spectra for high-efficiency identification and quantification of different substances. Moreover, large Raman spectral datasets can be automatically preprocessed for the emerging deep learning-based feature learning and intelligent detection.