Detecting Nucleated Cells in Bone Marrow Smears using Deep Learning

Abstract: Being able to detect nucleated cells in human blood is a very important part of health care. CellaVision has machines which can automatically detect cells in blood samples from peripheral blood instead of using manual microscopy. Sometimes it is not enough to only investigate the peripheral blood but also samples from the bone marrow. The bone marrow is different from the peripheral blood and it is generally more difficult to detect cells in. The aim of this master's thesis is to produce a machine learning model that can detect nucleated cells in bone marrow smears. We investigated several different models to see which model would yield the best performance while being sufficiently fast. After having found the type of model that worked best, we investigated different improvement techniques such as hyperparameter optimization, active learning and pseudo labeling to see if this could help improve the network's performance. Our results show that we found a fast enough model that could detect nucleated cells in bone marrow smears with a good performance and that the performance improved after application of the improvement techniques.