Exhaustive white blood cell extraction and profiling across whole-slide images of blood and bone marrow smears

Deutges, Ghete, Namsaraeva and colleagues propose the HemAutomaton framework based on neural cellular automata, which can extract white blood cells from whole-slide images of blood and bone marrow smears without the need for manual region selection and with minimal annotation requirements. The study reports that the framework is comparable to Faster R-CNN in in-domain detection accuracy, outperforms the latter under domain shift, and incurs only a small fraction of the latter's computational cost, with more than 10,000-fold fewer parameters. The authors extracted more than 100 million single-cell images across three blood cancer cohorts and, using a pretrained hematology foundation model, achieved annotation-efficient classification and counting through few-shot classification. Whole-slide spatial analysis revealed a consistent blast frequency gradient within smears, and the discrepancy between automated blast counts and expert assessment fell within the range of inter-expert variability. The authors state that scaling automated analysis from hundreds of cells per smear to tens of thousands of cells can overcome the limitations of traditional sampling.