White Blood Cell Classification with Data-Efficient Vision Transformer: A Multi-Dataset Evaluation on Raabin-WBC and PBC Repositories

Traditional microscopic assessment of leukocyte morphology remains a labor-intensive workflow that is inherently vulnerable to inter-observer variability and subjective assessment risks. Automated classification systems are increasingly used as diagnostic aids for the identification and characterization of hematological samples. This study proposes a data-efficient transformer-based framework for white blood cell (WBC) classification under limited training data and computational constraints. The proposed system employs a DeiT-Tiny backbone with a transfer learning strategy. Contrast-Limited Adaptive Histogram Equalization (CLAHE) was implemented to enhance cellular features. Model robustness and generalization capability were evaluated using a 5-fold cross-validation scheme on the Raabin-WBC and PBC datasets. The proposed framework achieved mean classification accuracies of 99.39% ± 0.20% and 99.63% ± 0.19% across 5-fold cross-validation on the Raabin-WBC and PBC datasets, respectively. For the Raabin-WBC dataset, the proposed model achieved macro-average precision, recall, and F1-score values of 98.10% ± 0.83%, 99.33% ± 0.36%, and 98.69% ± 0.48%, respectively. Similarly, for the PBC dataset, the corresponding macro-average precision, recall, and F1-score values were 99.54% ± 0.26%, 99.52% ± 0.31%, and 99.53% ± 0.28%, respectively. The results indicate that the model delivers consistent and high-level performance across all leukocyte classes. We also show additional views of the heatmaps generated by the Gradient-weighted Class Activation Mapping (Grad-CAM) technique, pointing out regions of the deep learning model’s feature maps that the network is looking at when classifying cells, and which correspond to biologically relevant parts of the cell, including the nucleus and cytoplasm. The experimental findings indicate that the proposed DeiT-Tiny framework provides highly accurate and computationally efficient leukocyte classification while reducing model complexity and maintaining stable performance across multiple microscopy datasets.