Wayne Matthew A. Dayata, Sabrinah Yonell C. Yap, Christine D. Bandalan
Acute Lymphoblastic Leukemia (ALL), caused by the continuous multiplication of malignant overproduction and immature white blood cells (WBCs), is one of the most common types of leukemia among children and adults with a high mortality rate mainly because of its late detection and diagnosis. This thereby drives the need to create systems that aid pathologists in the morphological analysis and detection of ALL blast cells, reducing error rates and increasing the likelihood of survival among ALL patients as a result. Five Convolutional Neural Network (CNN) models: ConvNeXtTiny, MobileNetV2, EfficientNetV2B3, InceptionV3, and DenseNet121 were integrated into a stacking ensemble to accurately distinguish ALL cells from healthy cells. Training and testing of the model were done using the C-NMC-2019 ALL dataset of 12,528 individual WBCs segmented from the bone marrow smear images of 101 subjects. The proposed approach has obtained a maximum accuracy score of 95.843%, precision score of 95.329%, and a weighted F1-score of 95.801% from a test set of 4,522 cells from 40 subjects in the said dataset. The results show that the proposed ensemble was able to better support clinical decisions to detect ALL in patients than the individual CNN models do. © 2023 IEEE.
University of San Carlos, Department of Computer, Information Sciences and Mathematics, Cebu City, Philippines
