Exhaustive white blood cell extraction and profiling across whole-slide images of blood and bone marrow smears
- 1Institute of AI for Health, Helmholtz Munich
Abstract
Blood and bone marrow smears are central to diagnosing hematological disorders. While whole-slide scanning enables automated single-cell profiling at scale, the computational burden of gigapixel images has confined automated analysis to manually pre-selected regions of interest. Here, we present HemAutomaton, a framework built on neural cellular automata that extracts white blood cells across entire whole-slide images without manual region selection and with minimal annotation requirements. HemAutomaton matches the in-domain detection accuracy of Faster R-CNN while surpassing it under domain shifts, operating at a fraction of the computational cost and with over 10,000-fold fewer parameters. Across three blood cancer cohorts, we extract over 100 million single-cell images and achieve annotation-efficient differential counting via few-shot classification using a pretrained hematology foundation model. Slide-wide spatial analysis uncovers consistent blast-frequency gradients across smears and demonstrates that automated blast counts match expert evaluations within inter-expert variability. Scaling automated profiling from hundreds to tens of thousands of cells per smear overcomes conventional sampling limits and establishes the foundation for more sensitive AI supported hematological assessments.
Related
- Matched marrow and blood profiling reveals compartment- and feature-specific liquid-biopsy signals in multiple myeloma
- Single-Cell and Multi-Omics Profiling of the Multiple Myeloma Immune Niche: From Bone Marrow Architecture and Metabolic Reprogramming to Immunotherapy Response.
- Successful bone marrow transplantation across major histocompatibility barriers in rhesus monkeys
- Phytochemical Diversity, Extraction Methods, Analytical Profiling, and Quality Control of Rheum tanguticum (Maxim. ex Regel) Balf.: An Evidence-Stratified Narrative Review
- Multivariate analysis and ADME profiling of Rosmarinus officinalis L. leaf vs. stem extracts: Extraction method-driven bioactivity and pharmacokinetic implications
BibTeX
@misc{deutges2026exhaustive,
title = {{Exhaustive white blood cell extraction and profiling across whole-slide images of blood and bone marrow smears}},
author = {M. Deutges and T. Ghete and A. Namsaraeva and S. Sembill and M. Metzler and C. Marr and M. Buser},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.23.26366335},
url = {https://www.medrxiv.org/content/10.64898/2026.10.23.26366335v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Exhaustive white blood cell extraction and profiling across whole-slide images of blood and bone marrow smears
AU - M. Deutges
AU - T. Ghete
AU - A. Namsaraeva
AU - S. Sembill
AU - M. Metzler
AU - C. Marr
AU - M. Buser
PY - 2026
DA - 2026/10/23
PB - medRxiv
DO - 10.64898/2026.10.23.26366335
UR - https://www.medrxiv.org/content/10.64898/2026.10.23.26366335v1
AB - Blood and bone marrow smears are central to diagnosing hematological disorders. While whole-slide scanning enables automated single-cell profiling at scale, the computational burden of gigapixel images has confined automated analysis to manually pre-selected regions of interest. Here, we present HemAutomaton, a framework built on neural cellular automata that extracts white blood cells across entire whole-slide images without manual region selection and with minimal annotation requirements. HemAutomaton matches the in-domain detection accuracy of Faster R-CNN while surpassing it under domain shifts, operating at a fraction of the computational cost and with over 10,000-fold fewer parameters. Across three blood cancer cohorts, we extract over 100 million single-cell images and achieve annotation-efficient differential counting via few-shot classification using a pretrained hematology foundation model. Slide-wide spatial analysis uncovers consistent blast-frequency gradients across smears and demonstrates that automated blast counts match expert evaluations within inter-expert variability. Scaling automated profiling from hundreds to tens of thousands of cells per smear overcomes conventional sampling limits and establishes the foundation for more sensitive AI supported hematological assessments.
ER -