Synthetic Degradation of Blood Cell Images for Acute Myeloid Leukemia Classification

Degradation study of blood cell images, specifically for Acute Myeloid Leukemia (AML) classification, entails determining how image quality concerns affect the accuracy of automated systems. This study is critical because microscopic blood smear pictures, which are the primary source for AML diagnosis, can suffer from a variety of degradation such as blur, noise, and abnormalities, preventing correct cell identification and classification. Furthermore, CNN networks that can perform and classify well using degraded images are preferred. The objective of this research is: i) to study several type of blood cell image degradation using synthetic filtering technique, ii) to study several combination of synthetic filtering to make the cIa ssification of AML more cha lleng ing. These degradation images is feed as an input for a CNN model for AML classification. U sing a CNN classifier that includes AlexNet, SqueezeNet, MobileNetV2, ResNet18, and VGG-16, normal white blood cell (WBC) and AML cells are predicted and classified from microscopic peripheral blood cell images. ROC curves, confusion matrices, and evaluation metrics are used to evaluate the effectiveness of each classifier. The result shows that the most challenging combination of filter is the salt and pepper as prefiltering, gaussian as filtering and salt and pepper as post filtering. VGG-16 outperformed the other models, identifying normal WBC and AML cells with an accuracy of 96.07%. There is still room for improvement, with an Fl-score of around 0.96. Further research is needed to improve classification accuracy and minimize misclassifications. This study demonstrates CNNs' potential in AML diagnosis, including accurate classification and improved patient care.