A unified biologically informed framework for bone marrow cell classification under ambiguity and imbalance

Bone marrow cytomorphology underpins diagnosis of myelodysplastic syndromes and other myeloid neoplasms, but automated classification is limited by strong inter-class similarity, long-tailed class distributions, and annotation ambiguity near developmental boundaries. We propose a biologically aware deep learning framework that encodes hematopoietic structure into both training and inference through: (i) a hierarchical coarse-to-fine classifier, (ii) a neighbor-structured soft auxiliary loss that relaxes penalties for morphologically adjacent classes, (iii) confidence-adaptive soft relabeling that preserves ambiguous samples with blended targets, and (iv) embedding-space k-nearest-neighbor retrieval-augmented inference with validation-only strategy selection. We evaluate under five-fold cross-validation on MK-11 (7,204 megakaryocyte images; 70 patients; official patient-/WSI-wise folds) and a curated 23-class subset of BM Cell MDS (24,811 images; stratified image-level folds). On MK-11, the framework achieves 77.36 ± 6.00% accuracy, 0.7311 ± 0.0355 macro-F1, and 91.65 ± 4.85% top-3 accuracy. On BM Cell MDS, it achieves 81.65 ± 0.27% accuracy, 0.7159 ± 0.0151 macro-F1, and 95.15 ± 1.87% top-3 accuracy. These results suggest that modeling biological locality and ambiguity can yield consistent and clinically interpretable gains without increasing backbone scale and with only a small retrieval-memory and lookup-latency overhead.