A Comparative Study of Convolutional Neural Network in Detecting Blast Cells for Diagnose Acute Myeloid Leukemia

Understanding blood plays a crucial role in obtaining information for monitoring health conditions and diagnosis of hematologic diseases such as acute myeloid leukemia. It is characterized by irregular expansion of immature white blood cells called blast cells in the blood and bone marrow. To diagnose acute myeloid leukemia, a sample of bone marrow is necessary to be examined under a microscope through bone marrow examination. As for minimizing human subjectivity and automating medical screening, this study performed image classification for detecting blast cells in leukocytes from microscopic images. The main objective of this study is to examine an established convolution neural network structure in detecting blast cells. We compared well-known architecture such as ResNet, ResNeXt, and EfficientNetV2. The model’s performance assessment was done by two evaluation levels which are at a macro level and per class level. The experiment results show ResNet architecture with 18 layers (ResNet 18) outperforms the remaining models at both levels. Furthermore, as the architecture utilizes residual learning, ResNet and ResNeXt models converge faster than EfficientNetV2 at the training phase. In addition, ResNet architecture with 50 layers (ResNet 50) outperforms the remaining models specifically at blast cell identification in case of medical screening. However, EfficientNetV2 shows a promising potential at a macro level to classify leukocytes in general while maintaining a competitive performance to ResNet and ResNeXt in the same numbers of parameter. Therefore, this study concludes that residual learning shows an outstanding performance in a few numbers of iteration. In addition, a model with shallow layer is the best model for classifying leukocyte in general and a model with deeper layer is the best model for detecting blast cells in leukocyte specifically.