Generative adversarial networks‐based super‐resolution algorithm enables high signal‐to‐noise ratio spatial heterodyne Raman spectra

• Published in: Journal of Raman spectroscopy Vol. 54; no. 12; pp. 1490 - 1501
• Main Authors: Hu, Lilei, Shen, Jingxuan, Chen, Zhu, Zhang, Yichen, Chen, Chang
• Format: Journal Article
• Language: English
• Published: Bognor Regis Wiley Subscription Services, Inc 01.12.2023
• Subjects: Acetaminophen; Algorithms; Blood levels; deep learning; generative adversarial network; Generative adversarial networks; Image enhancement; image processing; Image reconstruction; Interference; Raman spectra; Raman spectroscopy; spatial heterodyne Raman spectroscopy; Spectrometers; Spectrum analysis; super‐resolution; Training
• ISSN: 0377-0486; 1097-4555
• DOI: 10.1002/jrs.6598

High‐resolution interference pattern images are vital for spatial heterodyne Raman spectroscopy to produce quality Raman spectra with a good signal‐to‐noise ratio. A sought‐after super‐resolution algorithm can enhance Raman interference pattern images, reducing reliance on expensive imaging sensors. This enables miniaturization and portability for point‐of‐care testing. This study proposes a generative adversarial network (GAN)‐based super‐resolution reconstruction algorithm explicitly designed for Raman interference patterns. Employing GAN adversarial training, the algorithm effectively reconstructs interference pattern images at higher resolution, improving Raman spectra signal‐to‐noise ratio. Considering Raman spectra analysis mainly focuses on characteristic peaks, a Raman characteristic peak‐focused network training scheme is used. For instance, acetaminophen is studied with two selected Raman characteristic peaks centered at 388 and 858 cm−1. These peaks are observed in Fourier transformed Raman spectra from corresponding low‐resolution interference pattern images obtained by down‐sampling high‐resolution ones by twofold and fourfold. The proposed GAN‐based algorithm successfully reconstructs low‐resolution interference patterns into high‐resolution ones, achieving high R‐square values (96.05% for twofold and 91.1% for fourfold). This innovation holds potential for point‐of‐care applications, like noninvasive blood glucose concentration measurements, enabling cost‐effective, portable Raman spectrometers with improved capabilities. As for point‐of‐care testing, a generative adversarial network (GAN)‐based super‐resolution reconstruction algorithm is designed for spatial heterodyne Raman interference patterns. The algorithm effectively reconstructs interference pattern images at higher resolution, improving Raman spectra signal‐to‐noise ratio. Using a proposed Raman characteristic peak‐focused network training scheme, acetaminophen is studied with two selected Raman characteristic peaks centered at 388 and 858 cm−1. The proposed GAN‐based algorithm can successfully reconstruct low‐resolution interference patterns into high‐resolution ones, achieving a high R‐square value of 96.05%.