Optimizing Leukemia Classification with Fine-Tuned SENet Variants: A Comparative Study on Binary and Multi-Class Tasks

Humaira, Nisar and Kye-Fung, Lee and Siti Atiyah, Ali (2025) Optimizing Leukemia Classification with Fine-Tuned SENet Variants: A Comparative Study on Binary and Multi-Class Tasks. In: 2025 4th International Conference on Smart Cities, Automation & Intelligent Computing Systems (ICON-SONICS), 14-17 October 2025, Malacca, Malaysia.

PDF Optimizing Leukemia.pdf Download (660kB)

Abstract

Leukemia, a blood cancer diagnosed primarily through microscopic examination of blood smears, poses challenges due to the cost, time, and labor-intensive nature of the process. This paper proposes an automated deep learningbased approach for leukemia classification, addressing two tasks: (i) binary classification between acute lymphocytic leukemia (ALL) and healthy cells (HTL), and (ii) five-class classification involving ALL, acute myelogenous leukemia (AML), chronic lymphocytic leukemia (CLL), chronic myelogenous leukemia (CML), and HTL. Microscopic images were obtained from the ALL Image Database (ALL-IDB1) and the American Society of Hematology (ASH) Image Bank. Transfer learning was applied using three state-of-the-art models: Inception-V3, ResNeXt, and SENet. Results demonstrate that SENet consistently outperforms the other models. To further enhance performance, three SENet variants were developed by integrating support vector machine (SVM) classifiers, additional hidden layers with diverse feature vectors, and dropout regularization. The best-performing variant (SENet-D), combining feature selection with dropout regularization, achieved testing accuracies of 99.84% for binary classification and 84.48% for five-class classification, establishing its robustness and effectiveness for automated leukemia detection.

Actions (For repository members only: login required)

View Item