We present two deep learning models capable of simultaneous denoising and correction of Raman spectra, significantly accelerating high-quality data acquisition. Raman spectroscopy is a non-destructive analytical technique that reveals molecular vibrations, enabling precise identification of chemical compounds and material properties. Its spatial resolution and compatibility with microscopic imaging allow for high-resolution chemical mapping of heterogeneous samples. However, spectral artifacts such as baseline drift, cosmic rays, and instrumental noise complicate data interpretation, necessitating correction. Our models reduce exposure time by 90% while preserving signal integrity, as demonstrated on noisy spectra from brain tumor samples. These models are versatile and can be readily applied to novel Raman datasets, offering an order-of-magnitude improvement in acquisition efficiency. This work advances Raman spectroscopy as a faster, more reliable tool for chemical analysis, with broad applications in materials science, biomedical research, and beyond.
The spectra for the Live cell sample and the glioma tumor sample are available at seafile: https://seafile.utu.fi/d/dd325af0161b41bc989c/
Published version: https://doi.org/10.1002/adom.202500736