Autoencoder in Time-Series Analysis for Unsupervised Tissues Characterisation in a Large Unlabelled Medical Image Dataset

The topic of deep-learning has recently received considerable attention in the machine learning research community, having great potential to liberate computer scientists from hand-engineering training datasets, because the method can learn the desired features automatically. This is particularly beneficial in medical research applications of machine learning, where getting good hand labelling of data is especially expensive. We propose application of a single-layer sparse-auto encoder to dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for fully automatic classification of tissue types in a large unlabelled dataset with minimal human interference -- in a manner similar to data-mining. DCE-MRI analysis, looking at the change of the MR contrast-agent concentration over successively acquired images, is time-series analysis. We analyse the change of brightness (which is related to the contrast-agent concentration) of the DCE-MRI images over time to classify different tissue types in the images. Therefore our system is an application of an auto encoder to time-series analysis while the demonstrated result and further possible successive application areas are in computer vision. We discuss the important factors affecting performance of the system in applying the auto encoder to the time-series analysis of DCE-MRI medical image data.