Detection of Acute Myeloid Leukemia Using Deep Learning Models Based Systems

Abstract

The aim of this research is to develop a leukemia cell detection system based on deep learning models. The dataset was selected from The Cancer Imaging Archive, which had 3390 leukemia cells and 14,975 normal white blood cell images. Three deep learning models, CNN, YOLOv4, and YOLOv8 were built for this purpose. The CNN model was built from scratch and multiple parameters (epochs, network architecture, learning rate) were optimized. The overall detection accuracy, precision, recall, and F1 score achieved more than 90% after optimization. For the YOLOv4 model, two parameters (learning rate and randomization) were optimized and the overall accuracy, precision, recall, and F1 score achieved at or above 95%. In the YOLOv8 model, three parameters (epochs, learning rate, and architecture) were optimized, and the overall accuracy, precision, recall, F1 score, and mAP achieved above 95%. All three models have been successfully developed to detect leukemia cells and normal white blood cells with high performance. This deep learning system can significantly improve the efficiency of leukemia detection. The YOLOv8 and YOLOv4 models had similar performance but the YOLOv8 model training time was much shorter than the YOLOv4 model. In addition, the CNN model was more intuitive to build whereas the YOLO models did not need data augmentation. This is the first report to compare the YOLOv4 and YOLOv8 models with CNN model for detection of acute myeloid leukemia. The clinical recommendation will be given.