FocusAugMix: A data augmentation method for enhancing Acute Lymphoblastic Leukemia classification

• FocusAugMix introduces an innovative data augmentation method that combines Grad-CAM, Multi-Head Attention, and SaliencyMix to enhance ALL classification. • Utilizes superpixel and multi-head attention techniques, effectively identifying and preserving critical features in complex medical images. • Integrates advanced augmentation strategies to improve feature representation and address data limitations in medical image classification. • FocusAugMix enhances model robustness, reducing overfitting and ensuring strong performance with unseen data. • Applied across multiple leukemia datasets, FocusAugMix demonstrates superior classification performance compared to existing augmentation methods. The detection of various subtypes of Acute Lymphoblastic Leukemia (ALL) is crucial for precise medical identification, even though it is often hindered by the diverse appearance of leukemia cells and limited medical resources. Challenges arise from the subjectivity of evaluations and constraints in datasets, impacting the accuracy of classification. Existing methods face difficulties in achieving precise localization and building robust classification models due to the complexities in morphology and variations in subtypes, leading to challenges in accurate classification. This research proposes the FocusAugMix, a novel data augmentation method based on superpixels, which integrates Gradient-weighted Class Activation Mapping (Grad-CAM), Multi-Head Attention, and SaliencyMix to improve classification performance, especially in situations with limited datasets. The dynamic selection of superpixel contour images for each images allows this method to achieve a peak accuracy of 99.07 %, surpassing the previous method. Integrating Multi-Head Attention and Grad-CAM improves the accuracy and effectiveness of class representation in data augmentation methods for medical diagnosis.