Segmenting Overlapping Red Blood Cells With Classical Image Processing and Deep Learning

Nils Brünggel, Pascal Vallotton, Patrick Conway

2022

Abstract

In hematology the ability to count and analyze red blood cells (RBCs) is of major importance. Roche’s proprietary Bloodhound ® technology allows the automated printing and staining of slides to generate a monolayer of blood cells. While the RBCs are spread evenly, overlaps cannot be avoided completely. In the presence of such overlaps several tasks become problematic such as counting cells, quantifying the mean cellular volume or measuring cell shapes, critical for particular conditions such as anisocytosis (RBCs that are unequal in size) or rouleaux (clumps of RBCs that look like stacked coins). Modern deep learning models such as U-Net make it possible to accurately segment images given the appropriate training data (images and segmentation masks). The U-Net paper highlights the ability to train a model with only few images by applying data augmentation. We apply the learnings from their work and show that the mask creation can largely be automated: We collected images of free-standing RBCs, automatically segmented these using traditional image processing algorithms and combined these to generate artificial overlaps. We then used these images to train a model and show that it generalizes to real overlaps.

Paper Citation

in Harvard Style

Brünggel N., Vallotton P. and Conway P. (2022). Segmenting Overlapping Red Blood Cells With Classical Image Processing and Deep Learning. In Proceedings of the 1st Workshop on Scarce Data in Artificial Intelligence for Healthcare - Volume 1: SDAIH, ISBN 978-989-758-629-3, SciTePress, pages 47-52. DOI: 10.5220/0011537000003523

in Bibtex Style

@conference{sdaih22,
author={Nils Brünggel and Pascal Vallotton and Patrick Conway},
title={Segmenting Overlapping Red Blood Cells With Classical Image Processing and Deep Learning},
booktitle={Proceedings of the 1st Workshop on Scarce Data in Artificial Intelligence for Healthcare - Volume 1: SDAIH,},
year={2022},
pages={47-52},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0011537000003523},
isbn={978-989-758-629-3},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st Workshop on Scarce Data in Artificial Intelligence for Healthcare - Volume 1: SDAIH,
TI - Segmenting Overlapping Red Blood Cells With Classical Image Processing and Deep Learning
SN - 978-989-758-629-3
AU - Brünggel N.
AU - Vallotton P.
AU - Conway P.
PY - 2022
SP - 47
EP - 52
DO - 10.5220/0011537000003523
PB - SciTePress