Enhancing classification of mass spectrometry imaging data with deep neural networks

Mass spectrometry imaging (MSI) determines the spatial distribution of thousands of molecules and chemical species simultaneously and has emerged as a powerful suite of tools in pathology, pharmaceuticals, healthcare and material science applications. MSI experiments typically produce between 10
⁴
and 10
⁶
pixels, each of which contain a full mass spectrum. The use of MSI in pathological applications may involve classification of tissues based on the spectral information, though to date there have been no systematic studies of classification algorithms, comparison of performance when using the full versus reduced dimensionality of the data, or investigation into multi-class classification problems. Here we evaluate a number of algorithms for classifying regions in a MSI dataset before and after unsupervised non-linear dimensionality reduction with a deep neural network. We evaluate the performance of each algorithm with eight metrics in both the high (1,601) and low (3) dimensional feature space. Our results show that in the high dimensional space, only a Softmax classifier and support vector machine (SVM) with a linear kernel, perform to a satisfactory level, with all other algorithms overfitting the training data and performing poorly on the testing data. We also observe that that multi-class classification performance is drastically improved by the non-linear reduction with the deep neural network, improving scores and reducing variation between classes compared with the original high dimensional data. In the low dimensional space, the best performance observed is from a Decision Tree, though KNN, SVM (Gaussian kernel) and Softmax classifiers also perform well, scoring over 0.93 across all metrics and classes.