Abstract
Leukemia is a type of blood cancer affecting people of all ages and is the leading cause of death worldwide. The most common form of bone marrow leukemia is acute lymphoblastic leukemia (ALL). Diagnoses often require highly invasive diagnostic procedures, which are expensive and time-consuming. To overcome these issues, a novel Boundary-Aware Transformer based Fire Hawk Algorithm (BAT-FIRE) model has been proposed for multiclass classification of leukemia in its early stages. Initially, the input microscopic blood smear (MBS) images are denoised utilizing the Relative Total Variation (RTV) Regularization filter to reduce the noise and improve image quality. A Boundary-Aware Transformer (BAT) performs precise image segmentation to isolate cell boundaries. The Fire Hawk Optimization (FHO) algorithm is employed for selecting the most relevant features by eliminating the irrelevant features of the MBS images. Further, the classification is performed by a fully connected layer (FCL), which classifies the MBS images into five different classes: normal, ALL, Acute Myeloid Leukemia (AML), Chronic Lymphocytic Leukemia (CLL), and Chronic Myeloid Leukemia (CML). The proposed BAT-FIRE model attains an overall accuracy of 99.17% based on the gathered dataset. The proposed Attention ResNest model increases the accuracy by 7.90%, 5.40%, 4.08% and 2.01% better than AlexNet, DenseNet, ResNet and ResNest respectively. The proposed BAT-FIRE model enhances the overall ACC of 10.96%, 8.10%, 7.60%, and 4.63% better than CNN, ALNet, VGG 16, and SVM, respectively. BAT-FIRE model provides 99.17% accuracy in early leukemia detection, while reducing the need for invasive procedures. Compared to traditional deep learning models, its optimized feature selection and segmentation improve classification performance.