LWF-YOLO: a lightweight framework based YOLO for blood cell detection

Accurate detection of blood cells-red blood cells, white blood cells, and platelets-is essential for diagnosing hematological disorders such as anemia and leukemia. However, traditional approaches face critical limitations: manual microscopy is labor-intensive and subjective, automated analyzers fail to resolve complex morphologies in dense cellular scenes, and most deep learning models prioritize accuracy at the expense of computational efficiency, rendering them impractical for deployment on resource-constrained devices. Moreover, existing lightweight frameworks often sacrifice the fine-grained spatial details essential for precise boundary delineation. To overcome these challenges, we propose LWF-YOLO, a lightweight and high mean average precision (mAP) object detection framework built upon YOLOv11, tailored for blood cell detection. LWF-YOLO introduces the multi-scale edge-aware feature enhancement (MSEFE) module, which leverages Sobel operators and multi-scale fusion to sharpen cell boundary delineation, directly addressing overlapping cell separation. Complementing MSEFE, the dynamic channel-mixing gated former replaces conventional convolutional modules with a novel architecture that employs identity mapping, dynamic channel splitting, and adaptive gating. This design preserves fine-grained spatial details while optimizing feature modulation through depthwise convolutions and localized dependency aggregation, significantly reducing computational redundancy for lightweight deployment. Additionally, the dynamic deformable convolution network integrates adaptive sampling with multi-dimensional attention to improve robust localization across diverse cell scales. Extensive experiments on blood cell datasets (BCCD, complete blood count, LISC) and cross-domain validation demonstrate LWF-YOLO's exceptional performance, achieving 92.50% mAP@50 on the BCCD benchmark (+2.30pp over YOLOv11n baseline) with only 2.89 M parameters and 9.8 gigaflops, while demonstrating competitive or superior performance compared to state-of-the-art models like YOLOv11 and real-time detection transformer-R18. By prioritizing both efficiency and accuracy, LWF-YOLO enables real-time diagnostics on resource-constrained devices, advancing automated hematological analysis. The source code and pretrained weights are publicly available athttps://github.com/Luoyoooo/LWF-YOLO.