Multi-View Feature Fusion with Cross-Attention for Robust Leukemia Detection in Microscopic Imaging

Automated leukemia detection from microscopic blood cell images plays a crucial role in hematological diagnostics, enabling early and accurate disease identification. Traditional deep learning (DL) models primarily rely on single-view feature extraction, making them susceptible to variations in staining intensity, illumination, and imaging artifacts. These limitations hinder classification robustness and generalization, posing challenges in real-world clinical applications. To address these issues, this work introduces a multi-view feature fusion framework that integrates raw images, optical density (OD) transformed images, and anisotropically diffused images to enhance feature representation. A cross-attention mechanism is incorporated to adaptively refine salient features while suppressing redundant information across views. Experimental results on a high-resolution leukemia dataset demonstrate that the proposed model achieves a test accuracy of 97.90%, significantly outperforming single-view approaches. Furthermore, an ablation study confirms the impact of OD transformation and anisotropic diffusion in improving classification robustness. The confusion matrix and ROC analysis validate the model’s effectiveness, while LIME-based visual explanations enhance interpretability. These findings highlight the potential of multi-view learning in hematopathology, paving the way for more reliable AI-assisted leukemia diagnostics.