A Hybrid CNN-Transformer Framework for Normal Blood Cell Classification: Towards Automated Hematological Analysis

Background: Accurate classification of normal blood cells is a critical foundation for automated hematological analysis, including the detection of pathological conditions like leukemia. While convolutional neural networks (CNNs) excel in local feature extraction, their ability to capture global contextual relationships in complex cellular morphologies is limited. This study introduces a hybrid CNN-Transformer framework to enhance normal blood cell classification, laying the groundwork for future leukemia diagnostics. Methods: The proposed architecture integrates pre-trained CNNs (ResNet50, EfficientNetB3, InceptionV3, CustomCNN) with Vision Transformer (ViT) layers to combine local and global feature modeling. Four hybrid models were evaluated on the publicly available Blood Cell Images dataset from Kaggle, comprising 17,092 annotated normal blood cell images across eight classes. The models were trained using transfer learning, fine-tuning, and computational optimizations, including cross-model parameter sharing to reduce redundancy by reusing weights across CNN backbones and attention-guided layer pruning to eliminate low-contribution layers based on attention scores, improving efficiency without sacrificing accuracy. Results: The InceptionV3-ViT model achieved a weighted accuracy of 97.66% (accounting for class imbalance by weighting each class’s contribution), a macro F1-score of 0.98, and a ROC-AUC of 0.998. The framework excelled in distinguishing morphologically similar cell types demonstrating robustness and reliable calibration (ECE of 0.019). The framework addresses generalization challenges, including class imbalance and morphological similarities, ensuring robust performance across diverse cell types. Conclusion: The hybrid CNN-Transformer framework significantly improves normal blood cell classification by capturing multi-scale features and long-range dependencies. Its high accuracy, efficiency, and generalization position it as a strong baseline for automated hematological analysis, with potential for extension to leukemia subtype classification through future validation on pathological samples.