A Novel Boundary-Aware Transformer Based Fire Hawk Algorithm for Leukemia Classification using Blood Smear Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hematology

Background:

• Leukemia is a leading cause of cancer death globally, with diagnosis often requiring invasive, costly, and time-consuming procedures.
• Acute lymphoblastic leukemia (ALL) is a common form of bone marrow leukemia.
• There is a critical need for non-invasive, accurate, and efficient methods for early leukemia detection.

Purpose of the Study:

• To propose a novel Boundary-Aware Transformer based Fire Hawk Algorithm (BAT-FIRE) model for the early multiclass classification of leukemia.
• To enhance the accuracy and efficiency of leukemia diagnosis using microscopic blood smear images.
• To reduce reliance on invasive diagnostic techniques.

Main Methods:

• Microscopic blood smear (MBS) images were denoised using a Relative Total Variation (RTV) Regularization filter.
• A Boundary-Aware Transformer (BAT) was utilized for precise cell boundary segmentation.
• The Fire Hawk Optimization (FHO) algorithm performed feature selection, and a fully connected layer (FCL) classified images into five classes (normal, ALL, AML, CLL, CML).

Main Results:

• The BAT-FIRE model achieved an overall accuracy of 99.17% in classifying leukemia.
• The model demonstrated significant accuracy improvements over traditional deep learning models like CNN, ALNet, VGG 16, and SVM.
• Compared to Attention ResNest, the proposed model showed superior performance gains.

Conclusions:

• The BAT-FIRE model offers a highly accurate (99.17%) and efficient method for early leukemia detection from MBS images.
• This AI-driven approach minimizes the need for invasive diagnostic procedures.
• Optimized feature selection and segmentation contribute to the superior classification performance of BAT-FIRE.