Guided Representation Learning for the Classification of Hematopoietic Cells

IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), vol. 9, pp. 545–551

Abstract

Cell classification in human bone marrow microscopy images is a challenging image analysis task due to the number and inter-connection of cell types. While machine learning techniques have vastly higher throughput and could thus be more reliable, humans are intrinsically capable of understanding relations between cell types. In this paper, we propose methods to incorporate such intrinsic model knowledge based on representation learning. To this end, we construct a manually defined, two-dimensional reference embedding, coined embedding guide, which we use together with inverse dimensionality reduction, a distance-based loss and a growing embedding technique. Results show improved classification scores as well as a visually interpretable and clearly defined embedding space.

Authors 6

1. Philipp Gräbel Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   Institute of Imaging and Computer Vision, RWTH Aachen, Germany

2. Martina Margrit Crysandt Aachen Department of Hematology, UKA Aachen

   Universitätsklinikum Aachen

   Affiliation as printed

   Department of Hematology, UKA Aachen, Germany

3. Barbara M. Klinkhammer Aachen Institute of Pathology, UKA Aachen

   Universitätsklinikum Aachen

   Affiliation as printed

   Institute of Pathology, UKA Aachen, Germany

4. Peter Boor Aachen Institute of Pathology, UKA Aachen

   Universitätsklinikum Aachen

   Affiliation as printed

   Institute of Pathology, UKA Aachen, Germany

5. Tim H. Brümmendorf Aachen Department of Hematology, UKA Aachen

   Universitätsklinikum Aachen

   Affiliation as printed

   Dept. of Hematology, UKA Aachen, Germany

6. Dorit Merhof Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   Institute of Imaging and Computer Vision, RWTH Aachen, Germany