Background Blood cell classification through microscopic image analysis is a critical yet challenging task in hematological diagnostics. Traditional manual examination is time-consuming, prone to human error, and often requires extensive expertise. While existing automated solutions using conventional deep learning approaches show promise, they frequently struggle with capturing both local cellular details and global contextual features necessary for accurate classification. Aim This study proposes a novel approach to automate blood cell type classification by developing a hybrid deep learning architecture that effectively combines the strengths of both convolutional neural networks (CNNs) and transformer models to enhance diagnostic accuracy and reliability. Methodology We present a customized CoAtNet architecture that hierarchically integrates convolutional and transformer layers. The model comprises five progressive stages: two initial CNN stages with MBConv blocks for local feature extraction, followed by three transformer stages for global context modeling. The architecture employs relative attention mechanisms and features an adaptive channel scaling strategy 64→768 with multi-head attention (heads: 1→8). The model was trained using the AdamW optimizer with decoupled weight decay (1e−5) and a carefully crafted regularization strategy combining dropout and batch normalization. We utilized a publicly available Kaggle blood cell image dataset comprising 9,957 images across four distinct blood cell classes. The dataset was split into 6,770 images for training, 1,693 for validation, and 1,494 for testing. Results The proposed model achieved remarkable performance with 99% classification accuracy on the test dataset, significantly outperforming traditional CNN-only and transformer-only architectures. Conclusion The model demonstrated robust performance across different blood cell subtypes, with high precision and recall values. The exceptional accuracy of 99% demonstrates the effectiveness of combining local and global feature processing through the integration of CNNs and transformers.
Alshehri et al. (Wed,) studied this question.