Blood smears stop at 200 cells: a neural cellular automaton scans the whole slide and holds F1 0.79 across centers
Synopsis
Automated blood and bone marrow smear analysis had been confined to manually selected regions of interest because gigapixel whole-slide scanning is computationally heavy. HemAutomaton uses a neural cellular automaton to locate white blood cells across entire slides without manual region selection, matching Faster R-CNN in-domain detection and reaching a mean cross-domain F1 of 0.79 versus 0.65, with 3.3k trainable parameters against 41.4M. Across three blood cancer cohorts it extracted 108 million single-cell images and performed automated differential counting through few-shot classification with a pretrained hematology foundation model.
Interpretation
White blood cell localization moves from manually drawn regions of interest to entire whole-slide images, with no region chosen in advance. Earlier automated pipelines mirrored the manual reading order, preselecting regions of interest, assessing quality, then detecting and segmenting single cells at high resolution only inside those regions. In a three-domain cross-validation over three independent cohorts, the neural cellular automaton model reached a mean cross-domain F1 of 0.79±0.11 across n=6 cross-domain settings, against 0.65±0.18 for Faster R-CNN and 0.51±0.18 for a color-and-intensity blob baseline.
Detection holds up with very little annotation, needing a single training patch containing 100 white blood cells to move to a new dataset. Faster R-CNN improved steadily with more in-domain training data but degraded sharply in the 100-to-800 annotated-cell regime under domain shift. Trained on B-ALL bone marrow and evaluated on B-ALL peripheral blood, the Höfener et al. bone marrow set and the de Almeida et al. peripheral blood set, the neural cellular automaton model stayed robust in every setting, with the clearest advantage on the de Almeida set, which combines cross-modality and cross-institution shift.
Whole-slide processing is tractable, averaging about 9.4 minutes per slide, of which detection itself takes about 0.6 minutes. Parameter count drops from 41.4M for Faster R-CNN to 3.3k, FLOPs per tile from 126.1G to 10.0G, latency from 32.4 ms to 17.7 ms, and throughput rises from 30.9 to 56.3 tiles per second. Measured on the same inference pipeline with 512×512 tiles at roughly 1 µm per pixel, where tile reading and high-resolution cell cropping dominate runtime.
Automated blast-frequency estimates disagree with experts no more than the experts disagree with each other. Two experts estimating the same 200 cells per slide differed by a mean absolute deviation of 11.9%, while model-versus-expert cross-model deviation was 11.4%. On 20 B-ALL slides, one bone marrow and one peripheral blood slide from each of 10 patients, with patient-level 10-fold cross-validation, the non-inferiority test gave a mean difference of −0.54% (one-sided Wilcoxon signed-rank, p=0.0032) against a predefined 5-percentage-point margin.
Perspective
The demonstrated scope is whole-slide images of bone marrow and peripheral blood smears from three independent cohorts, covering five diseases plus normal controls on four scanning systems, with analysis restricted to white blood cells. The authors state that clinical translation requires broader testing across diverse scanners and disease profiles.
The authors note that macro-spatial analyses rely directly on model predictions, and high cell density in crowded regions can disproportionately hinder detection of larger cells such as blasts, potentially confounding the observed gradients; preprocessing uses brightness-based heuristics for tile filtering that may need manual tuning for variable slide backgrounds; and the current implementation covers white blood cells only, leaving red blood cell morphology, platelet clumping and smudge cells out of scope.