Ensemble of CNNs for Enhanced Leukocyte Classification in Acute Myeloid Leukemia Diagnosis

• Leonardo P. Sousa UFPI
• Romuere R. V. Silva UFPI
• Maíla L. Claro IFPI
• Flávio H. D. Araújo UFPI
• Rodrigo N. Borges UFPI
• Vinicius P. Machado UFPI
• Rodrigo M. S. Veras UFPI

Resumo

Acute Myeloid Leukemia (AML) is one of the most lethal and aggressive forms of hematological cancer, characterized by the rapid proliferation of immature leukocytes. This disease is diagnosed by highly trained specialists who meticulously analyze microscopic images of blood smears. This work explores the feasibility and effectiveness of using eight Convolutional Neural Network (CNN) architectures to form specialized ensembles capable of accurately differentiating between mature and immature leukocytes. We used voting ensemble techniques and the bagging method to integrate CNNs that achieved the best individual performances. The bagging strategy, explicitly using the EfficientNet B3 CNN, stood out by achieving an accuracy of 96.62%, a precision of 98.11%, and a Kappa index of 92.27% on a dataset of 48,100 blood cell images. This performance enhancement highlights the superior diagnostic capabilities of this approach compared to the individual architectures of CNNs in identifying cell types in the context of AML diagnosis.