Adaptive Reference Images for Blood Cells using Variational Autoencoders and Self-Organizing Maps

Abstract: CellaVision develops automated microscopy for blood analysis. Their products can pre-classify 19 different types of white blood cells and support the medical technologist performing the final classification. CellaVision provides reference cells. The reference cells are a fixed set of cells handpicked by a technologist, chosen to be typical for that specific class. The reference cells are static, i.e. the same for each cell image currently being classified and could be improved. We propose adaptive reference cells. Using a combination of machine learning techniques, we develop a pipeline consisting of an Xception classifier, a Variational Autoencoder (VAE) and a Self-Organizing Map (SOM). This pipeline is used to produce the adaptive reference cells which are specific for each cell image. The medical technologist classifying an image is thus presented with cells that are the most visually similar to that image. The adaptive reference cells are of particular use for cells that are hard to classify. The result consists of adaptive reference cells using both the output from the VAE and the SOM. The adaptive reference cells using the VAE are found superior. Also, cluster visualization using SOMs is presented together with proposed measurements for the robustness of the classifier and accuracy of the SOM.