An explainable hybrid CNN-transformer framework for multi-representation leukaemia classification

Acute Lymphoblastic Leukaemia (ALL) is a severe form of hematologic cancer that requires precise and prompt diagnosis for improved survival rates of patients. The conventional technique of examining microscopic images of peripheral blood smears is quite subjective, laborious, and heavily reliant on expert opinion. However, recent developments in artificial intelligence (AI) and machine learning (ML), specifically deep learning, have made automated detection of leukaemia possible; but the available techniques are biased due to training datasets, lack generalizability and are poorly interpretable. This paper presents a novel Hybrid Deep Learning Approach for Multi-Class Classification of Acute Lymphoblastic Leukaemia using EfficientNet and ViT Models. The framework uses convolutional networks for feature extraction, transformer networks for global representation learning, and handcrafted features. We evaluate the method on two clinical datasets: ALL-IDB2 and the ALL Image Dataset from Kaggle, both of which include benign and malignant types (Benign, Early, Pre, Pro). The comparison with benchmark methods shows that the proposed approach outperforms others. To improve model interpretability, we used Grad-CAM visualisations and Vision Transformer attention maps to examine image regions associated with the model’s predictions, complemented by quantitative faithfulness analysis. The proposed model achieves strong accuracy, F1-scores, specificity, and Matthews Correlation Coefficient (MCC). Although the framework demonstrated strong performance on both datasets, patient-wise separation could not be independently verified because reliable patient identifiers were not consistently available; therefore, further patient-wise and multi-centre external validation is required. These findings demonstrate the potential of the proposed explainable hybrid framework for accurate and interpretable leukaemia classification.