VisTA: vision transformer-attention enhanced CNN ensemble for optimized classification of acute lymphoblastic leukemia benign and progressive malignant stages

作     者：Nunna, Hasmitha Krishna Altable, Ali Gundala, Pallavi Rangarajan, Prasanna Kumar 

作者机构：Department of Computer Science and Engineering Amrita School of Computing Amrita Vishwa Vidyapeetham Chennai India Department of Computer Engineering Prince Sattam bin Abdulaziz University Al-Kharj 11942 Saudi Arabia Faculty of Computing and Information Technology King Abdulaziz University Jeddah 21589 Saudi Arabia 

出 版 物：《International Journal of Information Technology (Singapore)》 (Int. J. Inf. Technol.)

年 卷 期：2024年

页      面：1-18页

主　　题：Acute lymphoblastic leukemia classification Attention mechanism Benign Blood smears CLAHE Ensemble model Leukemia detection Malignant Vision transformers 

摘      要：Acute lymphoblastic leukemia (ALL), a rare blood disorder marked by excessive immature lymphocyte proliferation in the bone marrow, primarily affects children but poses a more challenging prognosis in adults due to late detection. Pathologists employ bone marrow morphology and immunophenotyping to establish cell lineage based on antigens to diagnose acute lymphoblastic leukemia. These manual approaches are time-consuming and need specialized expertise. It is critical for the decision support system to train an accurate classification model to discriminate between mature and immature white blood cells along with the stratification of its phases which is a crucial component in diagnostic accuracy. Another major challenge in ALL classification arises from the morphological resemblance between leukemic B-lymphoblast cells and normal B-lymphoid progenitors. To improve the detection of leukemia, this work integrates an ensemble model of Vision transformer with attention-based pre-trained architectures that use high-resolution blood smear images, to accelerate the prompt detection of this insidious illness and its stages. The ensemble model, which combines ViT with attention-based CNNs, uses dual feature extraction methods from cell images to improve classification results. Our suggested model demonstrates promising performance with an accuracy of 99.96% on a test dataset with 1600 images. The suggested model provides a promising methodology that could be extremely useful in medical settings, allowing for the accurate classification and diagnosis of acute lymphoblastic leukemia. © Bharati Vidyapeeth s Institute of Computer Applications and Management 2024.