Usefulness of Data Simulation for Training Deep Learning Denoising Algorithms in Infrared Spectral Histology

This study investigates the use of simulated data to train deep learning models for denoising infrared spectral images of paraffin-embedded tissue sections in clinical applications. Noise in Fourier-transform infrared spectroscopy poses significant challenges, particularly when reducing the number of scans per pixel to save acquisition time. To address this, we propose an approach based on a simulated linear generative model that incorporates different band shapes (Voigt, Gaussian, and Lorentzian) and noise types (additive and multiplicative Gaussian as well as Poisson noise) to enhance the diversity of training data and improve model generalization. Several data sets were designed to mimic different characteristics of infrared spectra acquired from paraffin-embedded tissues. Our findings reveal that the specific configurations of simulated spectra significantly influence model performance, with varying degrees of success depending on the properties of the simulated data. Furthermore, results show that a ResUNet-1D-CNN architecture trained on simulated data achieves performance comparable to that of models trained on real data, with a slight advantage when using a combination of real and simulated data. This approach demonstrates that training on simulated data not only enables robust reconstructions but also reduces the need for extensive real data acquisition. These findings highlight the effectiveness of data simulation as a powerful tool for the rapid and efficient development of denoising techniques in spectral imaging, paving the way for practical clinical deployment.