Hierarchical Deep Learning for Reliable Bone Marrow Cell Classification in Digital Healthcare

Abstract Accurate bone marrow cell classification can support the morphological assessment of hematological disorders, but manual examination is labor-intensive and observer-dependent. This study presents a Taguchi-guided, six-stage hierarchical deep-learning framework that decomposes seven-class bone marrow cell classification into clinically interpretable binary decisions involving identifiability, lineage, maturity, and subtype. ResNet-101, DenseNet-201, and EfficientNet-B0 were evaluated as stage-specific transfer-learning backbones, with solver, learning rate, and mini-batch size screened using an $${L}_{9}$$ orthogonal array. The framework was evaluated on a curated balanced dataset of 7,000 annotated images. Balanced node-level test accuracies ranged from 95 to 98%, with F1-scores from 0.95 to 0.98. End-to-end evaluation on the locked 700-image test set produced an overall accuracy of 96.3%, Micro-F1 of 0.963, and Macro-F1 of 0.963. Flat seven-class baselines achieved 30.0–37.0% accuracy under the recorded experimental settings. These results suggest that hierarchical decomposition was beneficial for this dataset, although image-level partitioning, the absence of external validation, and possible propagation of early routing errors limit generalizability. The system was implemented as a standalone MATLAB proof-of-concept interface. Prospective evaluation with hematopathologists and external patient- or slide-level validation are required before clinical use.