De-MSI: A Deep Learning-Based Data Denoising Method to Enhance Mass Spectrometry Imaging by Leveraging the Chemical Prior Knowledge

Mass spectrometry imaging (MSI) is a label-free technique that enables the visualization of the spatial distribution of thousands of ions within biosamples. Data denoising is the computational strategy aimed at enhancing the MSI data quality, providing an effective alternative to experimental methods. However, due to the complex noise pattern inherent in MSI data and the difficulty in obtaining ground truth from noise-free data, achieving reliable denoised images remains challenging. In this study, we introduce De-MSI, a novel deep learning-based method specifically developed for denoising MSI data without ground truth. The core concept of De-MSI involves constructing the reliable training data set by leveraging prior knowledge of mass spectrometry from the noisy MSI data, followed by training a deep neural network to improve the data quality by removing the noise from the original images. De-MSI has demonstrated superior performance in improving data quality over the commonly used methods when applied to MALDI-acquired mouse fetus data sets on visual inspection. Quantitative evaluations further confirm its superiority, with De-MSI achieving a mean PSNR of 18.93 and a mean SSIM of 0.74 across all ion images. The ability of De-MSI to enhance data quality in high-resolution MSI data sets is confirmed using the mouse brain data set at a pixel size of 5 μm. Additionally, its application to denoise rat brain data sets using the DESI technique showcases its adaptability across different ionization methods. The proposed model holds significant promise as a vital tool for the efficient analysis and interpretation of MSI data.