Wu Boyue, Feng Shilun, Jiang Shuyue, Luo Shaobo, Zhao Xi, Zhao Jianlong
State Key Laboratory of Transducer Technology, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai, 200050, China; School of Graduate Study, University of Chinese Academy of Sciences, Beijing, 100049, China.
State Key Laboratory of Transducer Technology, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai, 200050, China.
Comput Biol Med. 2025 Jun;192(Pt A):110288. doi: 10.1016/j.compbiomed.2025.110288. Epub 2025 Apr 30.
Blood cell detection is an important part of medical diagnosis. Object detection is trending for blood cell analysis, with research focusing on high-precision neural network models. However, these models have complex architectures and high computational costs. They cannot achieve rapid detection on low-end devices. Although lightweight models can greatly enhance the detection speed and achieve the real-time detection on low-end devices, their accuracy is poor in complex tasks. The development of efficient and highly accurate blood cell detectors for environments with limited computational resources is of great practical value. This study proposes an Efficient Blood Cell Detector based on YOLO (EB-YOLO) for blood cell detection. The model uses ShuffleNet as the backbone network for feature extraction to reduce the number of parameters and computational load. It incorporates the Convolutional Block Attention Module (CBAM) to enhance feature representation. In the neck network, Adaptive Spatial Feature Fusion (ASFF) is used for feature integration to improve multi-scale target feature extraction. Depth-wise separable convolution replaces standard convolution to reduce parameters while maintaining performance. Experimental results on the BCCD dataset show that the proposed model achieves 92.1 % mAP@50 %, the computational complexity is only 0.9 GFLOPs, and the number of parameters is 0.289M. The comparison results of the inference speed on Raspberry PI 5 show that the detection speed of the model is better than the classic YOLO algorithm model. The proposed method successfully balances lightweight design and high accuracy, which shows promise for deployment on low-end embedded systems.
血细胞检测是医学诊断的重要组成部分。目标检测在血细胞分析中呈发展趋势，研究主要集中在高精度神经网络模型上。然而，这些模型架构复杂且计算成本高。它们无法在低端设备上实现快速检测。尽管轻量级模型可以大大提高检测速度并在低端设备上实现实时检测，但在复杂任务中其准确性较差。开发适用于计算资源有限环境的高效且高精度的血细胞检测器具有重要的实际价值。本研究提出了一种基于YOLO的高效血细胞检测器（EB-YOLO）用于血细胞检测。该模型使用ShuffleNet作为骨干网络进行特征提取，以减少参数数量和计算量。它融入了卷积块注意力模块（CBAM）来增强特征表示。在颈部网络中，采用自适应空间特征融合（ASFF）进行特征整合，以改进多尺度目标特征提取。深度可分离卷积取代标准卷积以减少参数同时保持性能。在BCCD数据集上的实验结果表明，所提出的模型实现了92.1%的mAP@50%，计算复杂度仅为0.9 GFLOPs，参数数量为0.289M。在Raspberry PI 5上的推理速度比较结果表明，该模型的检测速度优于经典的YOLO算法模型。所提出的方法成功地平衡了轻量级设计和高精度，在低端嵌入式系统上部署具有前景。
