Lightweight efficient channel attention network for cross dataset white blood cell classification

Automated white blood cell (WBC) classification is important for supporting early screening of haematological disorders, but many deep learning approaches depend on large backbones and are insufficiently evaluated under variations in staining protocols, microscope settings, and dataset curation. To address this limitation, this study investigates cross-dataset robustness using two widely used WBC datasets, PBC-Normal and Raabin-WBC, together with their combined distribution, where heterogeneous imaging conditions and class imbalance make five-class WBC recognition challenging. The aim is to develop a computationally efficient and reproducible classification framework while systematically examining the role of preprocessing and backbone design for Basophil, Eosinophil, Lymphocyte, Monocyte, and Neutrophil classification. We present ECA-CNN-Lite, a 2.8 M-parameter residual CNN that integrates the established Efficient Channel Attention mechanism for channel-wise feature recalibration, and evaluate it under four preprocessing conditions: not cropped and unprocessed, not cropped and processed, cropped and unprocessed, and cropped and processed. The model is trained on 224 × 224 RGB images using AdamW optimization and cosine learning-rate decay, and is benchmarked against CNN and CNN Transformer baselines, including ConcatNeXt, VGG16-ViT, ResNet50-ViT, and VGG19. On individual datasets, ECA-CNN-Lite achieves 0.99 accuracy on PBC-Normal and 0.97 accuracy on Raabin-WBC, outperforming the strongest baseline by up to 0.04 absolute accuracy. On the combined dataset, it obtains 0.97 accuracy under the cropped and processed setting, with class-wise F1-scores of at least 0.95 across all five WBC types. Training-curve and F1-heatmap analyses indicate faster convergence, reduced sensitivity to preprocessing variation, and more balanced performance across rare and frequent classes compared with the evaluated baselines. These findings suggest that ECA-CNN-Lite, when combined with controlled cropping and stain-normalisation, provides a practical lightweight reference framework for cross-dataset WBC classification under the evaluated benchmark conditions.