Hybrid Vision Transformer and CNN for Leukemia Detection in Blood Smear Images

Leukemia, a life-threatening cancer of the blood and bone marrow, calls for rapid and precise diagnosis to improve patient survival outcomes. This study presents an innovative hybrid deep-learning model that collaborates with ViT and CNN for the automated detection of leukemia using microscopic blood smear images. Proposed ViT-CNN architecture takes advantage of ViT’s ability to capture global contextual features and CNN’s strength in extracting localized patterns, overcoming the limitations of traditional deep learning approaches. The methodology includes preprocessing and augmentation of blood smear images to enhance generalization. ViT segments the images into patches, embeds them into a high-dimensional space, and utilizes a transformer encoder to capture long-range dependencies. CNN layers further refine these embeddings to emphasize localized features critical for accurate classification, distinguishing between acute lymphoblastic leukemia, acute myeloid leukemia, and healthy cells. AdamW optimizer with weight decay ensures efficient training convergence and mitigates overfitting. Experimental validation in publicly available data sets (ALL-IDB1 and ALL-IDB2) demonstrates the superior performance of the model, achieving a classification accuracy of 97.4%, outperforming state-of-the-art architectures such as EfficientNet and ResNet. The hybrid model shows exceptional robustness to image variability, highlighting its potential for real-time clinical applications. By automating the diagnostic process, the proposed approach addresses the limitations of manual microscopy reviews, including labor intensity and human error. This research underscores the trans-formative potential of hybrid architectures in medical diagnostics, particularly for complex image-based diseases. Future work will focus on expanding datasets, optimizing computational efficiency, and exploring large-scale clinical integration to advance AI-driven healthcare solutions.