Dimensionality reduction of mass spectrometry imaging data using autoencoders

Abstract

The use of mass spectrometry imaging (MSI) techniques has become a powerful tool in the fields of biology, pharmacology and healthcare. Next generation experimental techniques are able to generate 100s of gigabytes of data from a single image acquisition and thus require advanced algorithms in order to analyse these data. At present, analytical work-flows begin with pre-processing of the data to reduce its size. However, the pre-processed data is also high in dimensionality and requires reduction techniques in order to analyse the data. At present, mostly linear dimensionality reduction techniques are used for hyper-spectral data. Here we successfully apply an autoencoder to MSI data with over 165,000 pixels and more than 7,000 spectral channels reducing it into a few core features. Our unsupervised method provides the MSI community with an effective non-linear dimensionality reduction technique which includes the mapping to and from the reduced dimensional space. This method has added benefits over methods such as PCA by removing the need to select meaningful features from the entire list of components, reducing subjectivity and significant human interaction from the analysis.