Scholarlink

<p>Accurate and efficient blood cell detection in microscopic images remains challenging due to dense object distribution, frequent occlusion, and the small size of platelets, which often lead to missed detections and redundant computations in existing detectors. To address these issues, we propose YOLO-ECL, a lightweight blood cell detection framework built upon YOLOv11n. YOLO-ECL introduces three targeted improvements: (i) C3k2_EMBC, which replaces the original C3k2 backbone module by integrating MBConv with Squeeze-and-Excitation (SE) attention to enhance fine-grained feature extraction for minute cells; (ii) ContextGuidedBlock_Down in the neck to jointly model local cues and surrounding context, improving robustness in dense and overlapping regions; and (iii) a Lightweight Shared Convolutional Detection head (LSCD) with a scale layer to reduce parameter redundancy while preserving multi-scale representation. Extensive experiments on the BCCD dataset (1344 images; RBC/WBC/platelet) show that YOLO-ECL achieves 95.8% mAP@0.5 with only 2.50&#xa0;M parameters and 6.1 GFLOPs, outperforming representative lightweight and mainstream YOLO variants under comparable settings. Moreover, cross-domain evaluation on the LISC dataset demonstrates improved generalization, reaching 86.2% mAP@0.5.Deployment results on embedded platforms (Jetson Nano, RK3588, Snapdragon 8 Gen 2) confirm real-time inference (≥ 25 FPS) with low energy consumption (&lt; 0.25&#xa0;J/image) and limited memory footprint (&lt; 600&#xa0;MB), supporting practical integration into portable intelligent medical testing equipment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

YOLOECL a lightweight model for blood cell detection in intelligent medical testing equipment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月29日

查看全文
  • Enqiang Liang,
  • Can Zhong,
  • Huimin Lv,
  • Haiying Sun,
  • Shengtao Li,
  • Minghui Zheng

摘要

Accurate and efficient blood cell detection in microscopic images remains challenging due to dense object distribution, frequent occlusion, and the small size of platelets, which often lead to missed detections and redundant computations in existing detectors. To address these issues, we propose YOLO-ECL, a lightweight blood cell detection framework built upon YOLOv11n. YOLO-ECL introduces three targeted improvements: (i) C3k2_EMBC, which replaces the original C3k2 backbone module by integrating MBConv with Squeeze-and-Excitation (SE) attention to enhance fine-grained feature extraction for minute cells; (ii) ContextGuidedBlock_Down in the neck to jointly model local cues and surrounding context, improving robustness in dense and overlapping regions; and (iii) a Lightweight Shared Convolutional Detection head (LSCD) with a scale layer to reduce parameter redundancy while preserving multi-scale representation. Extensive experiments on the BCCD dataset (1344 images; RBC/WBC/platelet) show that YOLO-ECL achieves 95.8% mAP@0.5 with only 2.50 M parameters and 6.1 GFLOPs, outperforming representative lightweight and mainstream YOLO variants under comparable settings. Moreover, cross-domain evaluation on the LISC dataset demonstrates improved generalization, reaching 86.2% mAP@0.5.Deployment results on embedded platforms (Jetson Nano, RK3588, Snapdragon 8 Gen 2) confirm real-time inference (≥ 25 FPS) with low energy consumption (< 0.25 J/image) and limited memory footprint (< 600 MB), supporting practical integration into portable intelligent medical testing equipment.