An XAI-Driven Cross-Attentive Dual-Branch CNN (CADBCNN) for leukemia subtype classification from peripheral blood smear images

The growing need for accurate and prompt diagnosis of blood cancers, especially leukemia, has boosted the progression of deep learning-based classification systems. Conventional microscopic examination, although accurate, is a time-consuming process that requires expertise. Therefore, automated methods are being explored for this purpose. In this study, we propose a novel model, the Cross-Attentive Dual-Branch CNN (CADBCNN), for the classification of leukaemia blood cells. The proposed model combines two feature branches using the ResNet18 and ResNet34 architectures with a cross-attention fusion mechanism. The proposed CADBCNN model achieved 98.36% accuracy, 98.41% precision, 98.36% recall, and 98.38% F1-score using stratified splitting, augmentation, and weighted optimisation. The proposed model outperformed existing architectures, including ResNet34, ResNet50, EfficientNet-B0, and a dual-branch model. Grad-CAM visualisation demonstrated that our model focuses on important regions, such as the nuclei and cytoplasm. The t-SNE plots show that our model achieved clear class separation. Therefore, we can conclude that our proposed model has achieved high accuracy and has a high potential for being used as a tool for automated leukemia blood cell classification. • The proposed CADBCNN (Cross-Attentive Dual-Branch CNN) achieved a maximum validation accuracy of 98.90% and test accuracy of 98.36% across four leukemia stages. • Dual-branch architecture (ResNet18 + ResNet34) significantly improves feature representation and classification performance compared to single-branch networks. • The cross-attention mechanism enhances feature synergy between branches, leading to better discrimination of morphologically similar subtypes. • Data preprocessing and augmentation improve generalization, reduce overfitting, and enable robust performance across all classes. • Explainable AI (Grad-CAM) and t-SNE visualizations confirm that the model focuses on clinically relevant regions (WBC nuclei and cytoplasm) and learns highly discriminative latent representations. • Ablation studies indicate that removing any key component (dual-branch, cross-attention, or augmentation) reduces accuracy and feature separability, validating the necessity of each design choice.