CNN and Transformer-based deep learning models for automated white blood cell detection

The classification and detection of white blood cells (WBCs) are of great significance in clinical diagnostics. With the rapid advancement of deep learning, it has become increasingly important in various domains, particularly medical imaging. Deep learning-based object detection techniques can accurately and rapidly identify and localize various WBC types in images. In this study, we first construct a subtype-stained WBC detection dataset, which contains approximately 5000 images, with an equal distribution across 1000 images per WBC subtype. Leveraging object detection techniques combined with two different networks, Transformer and Convolutional Neural Network (CNN), we implement the Single Shot Multibox Detector (SSD) and Faster Region-based CNN (R-CNN) object detection networks using ResNet-50, ConvNeXt-Tiny, and Swin-TransformerTiny as backbones, in addition to the You Only Look Once (YOLO) v5 network. These models are trained, validated, and tested on the constructed WBC dataset, achieving high performance while also exhibiting some differences in detection accuracy and real-time efficiency. Specifically, the mean average precision (mAP)@0.5 on the test dataset exceeds 98% for all models, with mAP@0.5:0.95 scores of 81.6%, 82.1%, 79.7%, 84.3%, 85.2%, 84.4%, and 88.4%, respectively. The detection speeds reach 112.58, 131.99, 90.85, 38.12, 41.94, 34.60, and 99.01 frames per second (FPS), respectively. These findings provide useful insights for selecting appropriate models in clinical applications, promoting more efficient and reliable diagnostics. The dataset is publicly available at https://github.com/LZF5411/dataset/tree/master.