J Pathol Inform. 2026 Nov;23
100710
Background: Automated classification of hematopoietic cells presents unique challenges due to morphological overlap across maturation stages and interobserver variability. Whereas deep learning models have demonstrated promising performance in digital pathology, limited attention has been given to systematic interpretability analysis in hematological cytomorphology.
Objective: To develop and rigorously evaluate an interpretability-centered deep learning framework for multiclass classification of 21 bone marrow cytomorphological cell types, incorporating cross-validation, formal statistical comparison of architectures, and multimodal visualization to assess both predictive performance and biological plausibility.
Methods: We developed a multiclass deep learning framework using EfficientNet-B3 to classify 171,373 bone marrow single-cell images spanning 21 morphological categories. Model performance was evaluated using top-1 accuracy, top-5 accuracy, macro-F1, and weighted-F1 scores. To assess robustness, 5-fold stratified cross-validation was performed. EfficientNet-B3 was compared against ResNet50 and DenseNet121, with statistical significance assessed using McNemar's test and bootstrap confidence intervals. Interpretability was evaluated through Grad-CAM visualization of confusion pairs and SHAP-based feature attribution. Latent feature structure was examined using PCA, UMAP, and t-SNE projections.
Results: On the held-out validation set, EfficientNet-B3 achieved top-1 accuracy of 87.6%, whereas cross-validated performance averaged 76.3% ± 0.27. Performance was statistically superior to both ResNet50 and DenseNet121, although the magnitude of improvement over DenseNet121 was modest. Most misclassifications occurred between morphologically adjacent classes, consistent with biological lineage continuity. Grad-CAM analysis demonstrated biologically plausible attention patterns in nuclear and cytoplasmic regions. Embedding projections revealed partial class clustering with expected overlap among transitional cell types.
Conclusion: This study reframes deep learning-based hematopoietic classification as an interpretability-centered problem. By integrating cross-validation, statistical model comparison, and multimodal visualization, we provide a comprehensive framework for understanding both performance and failure modes in multiclass bone marrow cytomorphology classification. These findings support the role of explainable artificial intelligence as a decision-support tool in hematopathology rather than a standalone diagnostic system.
Keywords: Bone marrow cytomorphology; Cross-validation; Deep learning; Explainable artificial intelligence; Grad-CAM; Hematopathology; Multiclass classification; SHAP