Hybrid Vision Transformer and MambaVision Network for Leukemia Subtype Classification

For the quick diagnosis and effective treatment plan, peripheral blood smear images can be used for leukemia subtype classification. We propose a Hybrid Vision Transformer and MambaVision Network for automated classification of four major leukemia subtypes: Acute Lymphoblastic Leukemia (ALL), Acute Myeloid Leukemia (AML), Chronic Lymphocytic Leukemia (CLL) and Chronic Myeloid Leukemia (CML). The proposed framework first processes the input images to make the cellular features more visible. The improved images are further passed to parallel branches: Vision Transformer (ViT) and MambaVision to extract complementary global contextual features and efficient sequential representation. The ViT branch then uses a cross-attention module to utilize the informative features extracted by the MambaVision, concatenates them for feature fusion, and classifies the features using a multilayer perceptron (MLP). The experimental results show that the proposed hybrid architecture has a classification accuracy of 95.20% over all the four types of leukemia, while also outperforming the individual backbone models and provides a powerful and efficient approach towards computer-aided diagnosis of leukemia.