Application of deep learning for linking microscope image features to ionic signal distribution

Introduction
Mass spectrometry imaging (MSI) is a powerful technique to identify and localize metabolites on biological tissue sections through their m/z values revealing molecular heterogeneity in various samples. Understanding the link between molecular signals and microscopic features is essential to the comprehension of biochemical mechanisms. However, it is not an easy task as increasing spatial resolution in MSI is associated with a loss of signal. In the opposite, microscopic images offer a high spatial resolution significantly faster than MSI but without or with few chemical insides. In this work, we map spatial intensity distribution of ions in MSI on the higher spatial resolution of microscopic images by an adaptation of the deep convolutional neural network U-Net with class activation map (CAM).
Methods
Microscopic images are taken from a sample deposit on glass before and after the MSI acquisition. An MSI is acquired with a MALDI source on the same glass. An image registration pipeline attribute to each mass spectrum its corresponding microscopic image according to the position of the ablation spot in the matrix. Then, a U-Net redesigned for regression is trained by taking as input a microscopic image and output m/z intensities of preselected ions. When trained, CAM from the last convolutional layer of the network generates a heatmap corresponding to regions of the microscopic image used by the network for predicting ion intensities giving inside of microscopic spatial features link to particular ions.
Preliminary Data
Mammalian cells, WI38 and MCF7, are growth and placed on glasses. The 2 cell types are used to generate 3 samples, one where the 2 cell types are mixed, and two others where cell types are isolated. Bright field microscopy images are taken from the glasses and MSI of the samples are acquired by MALDI FT-ICR-MS.
Then, we associate to each mass spectrum its corresponding part of the microscopic image of dimension 100x100 µm needed to cover the spot of the LASER ablation. From the MSI signal, matrix ions are manually annotated and lipids are annotated by METASPACE. Phosphatidylcholine (PC) and Sphingomyelin (SM) lipids, in particular, show differences in their expressions when compared between cell types in isolated samples. Indeed, SM lipids seem to be over expressed in WI38 cells and PC lipids in MCF7. However, this effect is not clearly visible when cell types are mixed. This is a consequence of the MSI resolution as the LASER ablation is larger than the cells which is often the case in real application. Then, we train our network on the mix samples by predicting PC, SM and matrix signals from associated microscope images. The trained network is applied to mix and isolated cells for evaluation. The CAM heatmap is retrieved for each prediction, showing the microscopic image spatial features used for each ion prediction. Interestingly, the CAM heatmaps are more intense at the location of WI38 cells for SM prediction, more intense at the location of MCF7 for PC prediction, more intense in regions with no cell for the matrix signal prediction. Those results show that the network successfully link image spatial features to lipid signal variations, showing the potential of using deep convolutional neural network to increase artificially MSI spatial resolution. Further work will focus on tissue sections.
Novel Aspect
Use of a deep learning network for artificially increasing spatial resolution of MSI by linking microscopic features and ions intensity.