Purpose/Objective: Accurate classification and instance segmentation of white blood cells (WBCs) in peripheral blood smears underpin the differential count, a core diagnostic test in hematology, yet manual microscopic examination is time consuming, operator-dependent, and difficult to scale across heterogeneous staining protocols and imaging devices. Existing deep learning methods are predominantly trained on single-cell crops from a single dataset, and the few that address cross-dataset domain shift rely on a single, global feature alignment that does not model the multiple WBC instances co-occurring in a realistic multi-cell smear. This dissertation aims to develop a WBC classification and instance segmentation framework that transfers cell-level, discriminative knowledge across heterogeneous datasets without extensive target-domain annotation. Methodology: A multi-teacher knowledge distillation framework is proposed, in which the number of teacher models corresponds to the number of WBC instances per image, enabling cell-level rather than image-level knowledge transfer. The framework combines bidirectional cross-modality translation, contrastive feature alignment, momentum-based self-distillation, and structure-aware hierarchical source/sparse-target attention within a Mask R-CNN-based teacher-student architecture, with ResNet, InceptionV3, VGG16, and NASNetLarge independently evaluated as teacher backbones. The framework is evaluated on four datasets, RV-PBS, CRV-PBS, PBC, and C-PBC, using precision, recall, F1-score, mAP, and mIoU, together with an ablation study and comparison against state-of-the art methods. Findings/Results: The distilled student model achieves a precision of 98.25%, recall of 98.70%, and F1-score of 98.47% on RV-PBS, outperforming every individual teacher backbone, and attains mAPscoresof0.863, 0.850, 0.856, and0.845 on RV-PBS,CRV-PBS, PBC, and C-PBC respectively, exceeding the closest prior baseline while using substantially fewer parameters(16.5 M versus 18–26M).Ablation results confirm that each proposed component contributes non-redundant performance gains. Conclusion/Implications: Cell-level, multi-teacher distillation provides a robust, annotation-efficient, and computationally practical route to generalizable WBC classification and segmentation across heterogeneous clinical imaging conditions, with future work directed toward larger multi-center datasets and prospective clinical validation.
