Edge-prior and reliability-guided collaborative learning for white blood cell classification

Accurate white blood cell (WBC) classification is important for hematological screening and computer-aided blood smear image analysis. Manual microscopic examination is time-consuming and observer-dependent, particularly for subtypes with similar nuclear morphology, cytoplasmic texture, and staining appearance. Deep learning methods have improved automated WBC recognition, but fine-grained subtype classification still requires effective coordination between local morphological details and broader cellular context. We present an edge-prior and reliability-guided collaborative framework for WBC classification. A ConvNeXt branch extracts local morphology, and a Mamba-based branch models long-range context. The Structure-Aided Attention Fusion module uses multi-scale edge priors to align features around nuclear contours and cytoplasmic boundaries. Reliability-guided bilateral fusion adjusts branch contributions using predictive entropy, maximum class probability, and classification margin. Evaluations across three open-access WBC datasets (PBC, LDWBC, Raabin-WBC) return respective image-level classification accuracies at 99.32%, 98.18% and 99.03%. Further tests spanning ablation, transfer learning, robustness and interpretability dissect core performance contributors alongside model stability. The results support the feasibility of the framework for image-level WBC classification on public benchmarks. Full coding resources of our classification framework appear at https://github.com/si-yuan20/White-blood-cell-classification.