Optimizing Acute Myeloid Leukemia Subtype Classification Through Data Balancing Techniques

Acute Myeloid Leukemia is a deadly blood cancer with several subtypes, each requiring accurate classification for effective treatment. Traditional detection methods depend on the manual examination of blood smears or bone marrow samples, which can be labor-intensive and inconsistent, emphasizing the need for automatic classification. This study uses the ResNet50 architecture to classify three AML subtypes: AML without maturation, Acute monocytic leukemia, and Acute erythroid leukemia. Various data balancing techniques were assessed to address class imbalance, including oversampling, undersampling, and their combinations with augmentation. The results revealed that a combination of oversampling and augmentation achieved the highest accuracy of 95.29%, while undersampling led to the lowest accuracy of 34.11%. These findings suggest that oversampling effectively balances class distribution while undersampling risks information loss. Additionally, data augmentation increases variability in the dataset, making the combined approach an effective strategy for enhancing model performance.