A six-stage hierarchical deep-learning framework has achieved an overall accuracy of 96.3 percent in classifying seven categories of bone marrow cells, according to a research study published in the International Journal of Computational Intelligence Systems. Developed to support the morphological assessment of hematological disorders, the computational architecture decomposes seven-class cellular recognition into a structured sequence of clinically interpretable binary decisions. In an end-to-end evaluation conducted on a locked 700-image test set, the framework achieved micro-F1 and macro-F1 scores of 0.963, substantially outperforming flat seven-class baseline models that recorded accuracies of only 30.0 to 37.0 percent under the same experimental settings.
The investigation addresses longstanding operational challenges in diagnostic hematology, where manual microscopic examination of bone marrow smears remains labor-intensive and subject to observer-dependent variability. To address these diagnostic challenges, the researchers structured their framework to divide the complex task into a hierarchy of binary decisions that evaluate cell identifiability, lineage, maturity, and specific cellular subtype, establishing an interpretable classification sequence.
To power each stage of the decision hierarchy, the researchers evaluated three distinct transfer-learning architectures as stage-specific backbones: ResNet-101, DenseNet-201, and EfficientNet-B0. The authors utilized a Taguchi-guided experimental design, applying an L9 orthogonal array to systematically screen and optimize stage-specific training hyperparameters. This structured screening process evaluated different optimization solvers, initial learning rates, and mini-batch sizes, identifying tailored configurations for each neural network backbone across the six classification stages.
The framework was trained and evaluated using a curated balanced dataset containing 7,000 annotated bone marrow cell images. In node-level evaluations across the six individual binary decision points, balanced test accuracies consistently ranged from 95 to 98 percent, with corresponding node-level F1-scores between 0.95 and 0.98. These individual node results demonstrated that breaking the complex multi-class problem into isolated binary stages enabled the deep-learning backbones to maintain high classification fidelity at each distinct branch of the decision tree.
To demonstrate its clinical workflow potential, the researchers implemented the complete hierarchical framework as a standalone proof-of-concept user interface in MATLAB. The experimental findings indicated that hierarchical decomposition was particularly beneficial for this dataset, enabling the architecture to avoid the steep performance drops observed in standard flat classification baselines, which achieved only 30.0 to 37.0 percent accuracy under the recorded conditions.
However, the authors highlighted several methodological limitations that temper the clinical generalizability of the current system. The study utilized image-level data partitioning without external dataset validation, and the sequential nature of hierarchical classification creates a risk where errors made at early routing nodes can propagate into downstream decisions. Because of these factors, the authors emphasized that prospective evaluations with practicing hematopathologists, along with external patient-level and slide-level validation, are required before the framework can be considered for real-world clinical use.
Subject of Research: Hierarchical deep learning for bone marrow cell classification
Article Title: Hierarchical Deep Learning for Reliable Bone Marrow Cell Classification in Digital Healthcare
Article References: Deepak, G. D., H, G., Hiremath, P., & Bhat, S. K. (2026). Hierarchical Deep Learning for Reliable Bone Marrow Cell Classification in Digital Healthcare. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01651-y
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01651-y
Keywords: bone marrow cell classification, hierarchical deep learning, hematological disorders, morphological assessment, transfer learning, ResNet-101, DenseNet-201, EfficientNet-B0, Taguchi method, orthogonal array, binary decision tree, digital healthcare, MATLAB interface, image-level partitioning, F1-score, hematopathology, cell lineage, cell maturity
News Source: Blake Davidson. (November 18, 2026). Six-Stage Hierarchical Deep Learning Achieves 96.3 Percent Bone Marrow Cell Accuracy. Scienmag.



