LeukocyteNet: An explainable transfer-transformer fusion learning model for Leukocyte classification

Abstract

White Blood Cells (WBCs), or leukocytes, are essential components of the immune system that protect the body against infections and malignant disorders. Even minor fluctuations in leukocyte count can indicate serious pathological conditions, including life-threatening malignancies such as leukemia, lymphoma, and myelodysplastic syndromes. Conventional diagnosis through manual microscopic examination is time-consuming, subjective, and heavily dependent on the pathologist’s expertise. To overcome these challenges, this study introduces LeukocyteNet, a transfer–transformer fusion model designed for the automated classification of ten malignant leukocyte categories. The model integrates convolutional feature extraction from VGG19 with the Swin Transformer’s global attention mechanism, enabling robust representations of both local morphology and global spatial dependencies. The LeukocyteNet model was trained on three publicly available datasets “ALL-IDB, the American Society of Hematology Image Bank, and Tehran Taleqani Hospital” and achieved an overall accuracy of 97.34%, a macro-averaged F1-score of 0.95, and a recall of 0.93, outperforming all evaluated baseline models. Furthermore, the inclusion of explainable AI techniques Grad-CAM, LIME, and Saliency Map enhances explainability by visualizing class-specific decision regions, thereby increasing clinical transparency and reliability. These findings demonstrate that LeukocyteNet not only achieves state-of-the-art predictive performance but also provides interpretable insights critical for trustworthy medical diagnostics.