CRViT-YOLO: A method for multi-morphological blood cell detection using convolution-restructured vision transformer

Complete blood cell counting plays a critical role in medical diagnostics; however, conventional manual examination is time-consuming and prone to errors due to variations in data sources, image quality, cell morphology, and staining characteristics. Deep learning has emerged as a promising solution to enhance both the accuracy and efficiency of blood cell detection. In this study, we present CRViT-YOLO, a novel detection framework built upon the YOLOv9 architecture. The proposed framework incorporates a Convolutional-Reconstructed Vision Transformer (CRViT) module to improve feature extraction by effectively capturing both local and global contextual information. Furthermore, a Feature Enhancement Module (FEM) is introduced to refine local feature representations, while the integration of the EIoU loss function enhances localization accuracy, particularly for densely packed or overlapping cells across diverse scales and types, and demonstrates robust performance in detecting polymorphic, healthy, and pathological cells. Extensive experiments conducted on four publicly available datasets-BCCD, BCDD, LISC, and BBBC041-validate the effectiveness and generalizability of the proposed approach, achieving mean average precision (mAP@50) scores of 93.9 %, 99.4 %, 98.8 %, and 76.0 %, respectively, in multi-class blood cell detection tasks.