TinyNCFormer: A Tensor-Decomposed Dual-Stream Transformer with Diffusion-Guided Ensemble Knowledge Distillation for White Blood Cell Classification and Instance Segmentation

Abstract White blood cell (WBC) classification from peripheral blood smear images is central to diagnosing leukemia, infection, and other hematologic disorders, and remains one of the most frequently requested laboratory analyses in clinical practice. Automated classification is difficult because WBC subtypes are morphologically similar, clinical microscopy fields routinely contain multiple overlapping cells, and large, accurate models are impractical to deploy on point-of-care hardware. This trade-off is addressed with a knowledge-distillation framework in which a single teacher, formed by fusing five diverse pretrained backbones over a nuclear-cytoplasmic, stain-separated representation, supervises a compact dual-stream Transformer student, TinyNCFormer, through confidence-calibrated soft-target distillation, diffusion-based feature denoising, and an instance-segmentation head for multi-cell scenes. Evaluated on our newly introduced RV-PBS dataset and three additional blood cell corpora (CRV-PBS, PBC, and C-PBC), the resulting 16.5-million-parameter student exceeds every constituent backbone, reaching classification F1-scores of 96.32--98.32\% and mean average precisions of 0.865--0.883 for instance segmentation across the four datasets, outperforming prior baselines by an average of 8.7\% (range 2.7--15.3\%) in relative mAP. These results show that diverse ensemble supervision with uncertainty- and denoising-aware distillation yields a compact model more accurate than its teacher and substantially more efficient, a practical candidate for point-of-care hematological screening.