<![CDATA[Accurate segmentation and classification of white blood cell (WBC) are essential for clinical hematology, yet remain challenging due to staining variability, complex backgrounds, and class imbalance. This study introduces an explainable, saliency-enhanced deep learning framework designed to achieve stain-robust leukocyte analysis. The framework integrates a saliency-driven preprocessing module, a lightweight EfficientSwin hybrid backbone, and a ResNeXt-CC–inspired cross-layer feature fusion block to capture complementary fine-grained and global features. A multi-task head jointly performs WBC segmentation and subtype classification, while a saliency-alignment loss enforces consistency between learned attention and saliency priors, providing training-time interpretability rather than post-hoc visualization alone. SG-CLDFF was evaluated on three public datasets (BCCD, LISC, ALL-IDB) and further tested under cross-stain and cross-dataset conditions. The framework achieved 95.8% accuracy, 0.94 F1-score, and 0.82 IoU, improving over strong CNN and transformer baselines. Ablation studies confirmed that both saliency preprocessing and cross-layer fusion contribute independently to performance, with saliency alignment yielding ≥2 IoU improvement in cross-stain scenarios (p < 0.05). Qualitative results using saliency maps and Grad-CAM demonstrate focused attention on diagnostically meaningful regions. These findings validate SG-CLDFF as a robust, interpretable, and stain-resilient solution for automated WBC analysis, offering a practical foundation for deployment in digital hematology workflows.]]>
