Spatial Maturity Regression for the Classification of Hematopoietic Cells

In contrast to peripheral blood, cells in bone marrow microscopy images are not only characterized by the cell lineage but also a maturity stage within the lineage. As maturation is a continuous process, the differentiation between various stages falls into the category of (ordinal) regression. In this work, we propose Spatial Maturity Regression - a technique that regularizes the learning process to enforce a sensible positioning of maturity stages in the embedding space. To this end, we propose and evaluate several curve models, target definitions and loss function that incorporate this domain knowledge. We show that the classification F-scores improve up to 2.4 percentage points when enforcing regression targets along learnable curves in the embedding space. This technique further allows visualization of individual predictions by providing the projected position along the learnt curve.