Feasibility Study for Autovalidation of Blood Cells in a Digital Morphology System

Abstract: This thesis was carried out at CellaVision, a company within digital blood analysis, with the aim of investigating the possibility of autovalidation of blood cells using neural networks. The thesis started with preparing a dataset for training and validation, including cell features and binary labels indicating if the cell is easy or difficult to classify. This dataset consisted of 122 227 cell images. The binary labels were generated through several different methods such as using neural networks, statistics from CellaVision’s system’s classification result and manual classification by a morphology expert. Three kinds of neural networks were tested with the aim of separating easy cells from difficult ones: a binary Artificial Neural Network (ANN), an autoencoder for anomaly detection and a Self-Organizing Feature Map (SOFM) visualizing the position of difficult cells in clusters. The methods were compared and the performance were evaluated. It was found that the ANN was not useful for this task, while the autoencoder could be used for successfully autovalidating 74% of the cells. With better labeling techniques for the dataset, the performance could potentially be improved. The SOFM was not used for anomaly detection in this study, but for visual analysis of the clusters and labels. However, it may become relevant in the future to look further into this method’s possibility of anomaly detection, as patterns in the clusters were apparent.