Surrounding Cell Suppression For Unsupervised Representation Learning In Hematological Cell Classification.

Analysis of hematopoietic cells in bone marrow images is a newly emerging field in computer vision. Deep neural networks provide promising approaches for detection and classification tasks in this field. However, labelling a sufficiently large amount of images by medical experts is infeasible in practice. This can be resolved by semi-supervised methods that use image reconstruction as a way to incorporate images without labelled cells. However, this inevitably leads to an inclusion of surrounding cells into the learned representation. We propose and analyze several techniques for reducing their influence and show that this improves classification results of unsupervisedly learned cell representations.