White Blood Cell (WBC) detection is crucial in medical image processing. A higher accuracy result is better doctor assistance in disease diagnosis. Several methods in object detection have achieved some notable results. However, there are some limitations when applying those models in WBC detection from blood cell smear images because the level of similarity among blood cell types is high, which is a challenge for methods to separate them in analysis. This paper aims to evaluate promising models in WBC detection on a practical blood cell smear image dataset from Vietnamese hospitals. The modularized neck architecture, which allows flexible configuration of the backbone, neck, and detector modules, was utilized to speed up the implementation of the models. Four built models based on YOLOv3, Faster R-CNN, EfficientNet, and HRNet were evaluated on the dataset and achieved promising results with mAP(s) at 61.70%, 71.00%, 57.20%, 50.30% respectively. The result is a premise for further research of an automatic WBC detection system that assists doctors in blood cell examinations.