Analysis of automatically generated embedding guides for cell classification

Abstract

Automated cell classification in human bone marrow microscopy images could lead to faster acquisition and, therefore, to a considerably larger number of cells for the statistical cell count analysis. As basis for the diagnosis of hematopoietic dis-eases such as leukemia, this would be a significant improvement of clinical workflows. The classification of such cells, however, is challenging, partially due to dependencies between different cell types. In 2021, guided representation learning was introduced as an approach to include this domain knowledge into training by providing “embedding guides” as an optimization target for individual cell types. In this work, we propose improvements to guided repre-sentation learning by automatically generating guides based on graph optimization algorithms. We incorporate information about the visual similarity and the impact on diagnosis of mis-classifications. We show that this reduces critical false predictions and improves the overall classification F-score by up to 2.5 percentage points.