Mass spectral imaging of clinical samples using deep learning

Abstract

A better interpretation of tumour heterogeneity and variability is vital for the improvement of novel diagnostic techniques and personalized cancer treatments. Tumour tissue heterogeneity is characterized by biochemical heterogeneity, which can be investigated by unsupervised metabolomics. Mass Spectrometry Imaging (MSI) combined with Machine Learning techniques have generated increasing interest as analytical and diagnostic tools for the analysis of spatial molecular patterns in tissue samples. Considering the high complexity of data produced by the application of MSI, which can consist of many thousands of spectral peaks, statistical analysis and in particular machine learning and deep learning have been investigated as novel approaches to deduce the relationships between the measured molecular patterns and the local structural and biological properties of the tissues. Machine learning have historically been divided into two main categories: Supervised and Unsupervised learning. In MSI, supervised learning methods may be used to segment tissues into histologically relevant areas e.g. the classification of tissue regions in H&E (Haemotoxylin and Eosin) stained samples. Initial classification by an expert histopathologist, through visual inspection enables the development of univariate or multivariate models, based on tissue regions that have significantly up/down-regulated ions. However, complex data may result in underdetermined models, and alternative methods that can cope with high dimensionality and noisy data are required. Here, we describe, apply, and test a novel diagnostic procedure built using a combination of MSI and deep learning with the objective of delineating and identifying biochemical differences between cancerous and non-cancerous tissue in metastatic liver cancer and epithelial ovarian cancer. The workflow investigates the robustness of single (1D) to multidimensional (3D) tumour analyses and also highlights possible biomarkers which are not accessible from classical visual analysis of the H&E images. The identification of key molecular markers may provide a deeper understanding of tumour heterogeneity and potential targets for intervention.