Hyperbolic neural networks for biologically informed classification of white blood cells and hematopoietic cell states in peripheral blood smears

Abstract Automated white blood cell (WBC) classification in peripheral blood smears supports morphological assessment of hematological disorders, but most deep learning classifiers treat cell categories as independent labels and do not account for hematopoietic maturation. We propose a hyperbolic classification framework that combines a ResNet-18 feature extractor with a prototype-based classifier in the Poincaré ball, allowing biologically related cell states to be represented in a geometry suited to hierarchical structure. The model was evaluated on a curated dataset of 129,174 annotated single-cell images mapped to 14 blood cell classes, including leukocyte subtypes, hematopoietic precursor states, and multiple stages of granulocytic maturation. To reduce acquisition-level information leakage, we used a grouped evaluation protocol that keeps linked multi-focus acquisitions within the same subset. Across the six top-ranked grouped-split configurations, the hyperbolic model improved macro-averaged F1-score from 0.812 ± 0.005 to 0.819 ± 0.004 relative to the Euclidean baseline, while overall accuracy remained slightly lower (0.826 ± 0.002 vs 0.830 ± 0.006). Recall gains were most pronounced for promyelocytes, immature WBCs, and metamyelocytes. These results suggest that hyperbolic representation learning can improve class-balanced classification of WBCs and related hematopoietic cell states while maintaining competitive overall accuracy.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-9930528/v1
Primary Topic
Digital Imaging for Blood Diseases
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Hyperbolic neural networks for biologically informed classification of white blood cells and hematopoietic cell states in peripheral blood smears

Markéta Vašinková, Petr Gajdoš, Lukáš Jochymek
1 links
Research Square
Digital Imaging for Blood Diseases
preprint

Hyperbolic neural networks for biologically informed classification of white blood cells and hematopoietic cell states in peripheral blood smears

Markéta Vašinková, Petr Gajdoš, Lukáš Jochymek
preprint en

Abstract

Abstract Automated white blood cell (WBC) classification in peripheral blood smears supports morphological assessment of hematological disorders, but most deep learning classifiers treat cell categories as independent labels and do not account for hematopoietic maturation. We propose a hyperbolic classification framework that combines a ResNet-18 feature extractor with a prototype-based classifier in the Poincaré ball, allowing biologically related cell states to be represented in a geometry suited to hierarchical structure. The model was evaluated on a curated dataset of 129,174 annotated single-cell images mapped to 14 blood cell classes, including leukocyte subtypes, hematopoietic precursor states, and multiple stages of granulocytic maturation. To reduce acquisition-level information leakage, we used a grouped evaluation protocol that keeps linked multi-focus acquisitions within the same subset. Across the six top-ranked grouped-split configurations, the hyperbolic model improved macro-averaged F1-score from 0.812 ± 0.005 to 0.819 ± 0.004 relative to the Euclidean baseline, while overall accuracy remained slightly lower (0.826 ± 0.002 vs 0.830 ± 0.006). Recall gains were most pronounced for promyelocytes, immature WBCs, and metamyelocytes. These results suggest that hyperbolic representation learning can improve class-balanced classification of WBCs and related hematopoietic cell states while maintaining competitive overall accuracy.

Research Square
VSB - Technical University of Ostrava (CZ)
Good health and well-being
Digital Imaging for Blood Diseases
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon