Scholarlink

<p>Accurate detection of blood cells in peripheral blood smears is important for clinical analysis, and real-time processing is desirable for practical microscopy workflows. However, microscopic images contain strong scale variation, dense cell adhesion, boundary truncation, and staining artifacts. We propose MR-YOLO, a lightweight YOLOv8n-based detector for blood-cell detection. MR-YOLO integrates MB-D2CM for efficient multi-scale representation, RMSPF for contextual aggregation, MSFFM for adaptive local–global feature fusion, and CSB-QAL for class- and scale-aware optimization. On BCCD, MR-YOLO achieves 94.7% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text {mAP}@0.5\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>mAP</mtext> <mo>@</mo> <mn>0.5</mn> </mrow> </math></EquationSource> </InlineEquation> and 66.0% <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\text {mAP}@0.5:0.95\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>mAP</mtext> <mo>@</mo> <mn>0.5</mn> <mo>:</mo> <mn>0.95</mn> </mrow> </math></EquationSource> </InlineEquation> with 2.9M parameters and 8.3 GFLOPs, while running at approximately 130 FPS on an RTX 4060Ti GPU with batch size 1. Compared with YOLOv8n, it improves <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\text {mAP}@0.5\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>mAP</mtext> <mo>@</mo> <mn>0.5</mn> </mrow> </math></EquationSource> </InlineEquation> by 3.4 points with a throughput reduction from 149 to 130 FPS. Because BCCD is small, its results are treated as controlled within-dataset evidence, and target-domain generalization is further evaluated on TXL-PBC. The results suggest that MR-YOLO provides a compact and GPU real-time solution for microscopic blood-cell detection.</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

MR-YOLO: lightweight multi-scale YOLO for blood-cell detection in microscopic images

  • 期刊论文

  • 发表日期：

    2026年7月26日

查看全文
  • Shiwei Zhang,
  • Xiongwen Zhong,
  • Yuxiang Wu,
  • Huaye Chen

摘要

Accurate detection of blood cells in peripheral blood smears is important for clinical analysis, and real-time processing is desirable for practical microscopy workflows. However, microscopic images contain strong scale variation, dense cell adhesion, boundary truncation, and staining artifacts. We propose MR-YOLO, a lightweight YOLOv8n-based detector for blood-cell detection. MR-YOLO integrates MB-D2CM for efficient multi-scale representation, RMSPF for contextual aggregation, MSFFM for adaptive local–global feature fusion, and CSB-QAL for class- and scale-aware optimization. On BCCD, MR-YOLO achieves 94.7% \(\text {mAP}@0.5\) mAP @ 0.5 and 66.0% \(\text {mAP}@0.5:0.95\) mAP @ 0.5 : 0.95 with 2.9M parameters and 8.3 GFLOPs, while running at approximately 130 FPS on an RTX 4060Ti GPU with batch size 1. Compared with YOLOv8n, it improves \(\text {mAP}@0.5\) mAP @ 0.5 by 3.4 points with a throughput reduction from 149 to 130 FPS. Because BCCD is small, its results are treated as controlled within-dataset evidence, and target-domain generalization is further evaluated on TXL-PBC. The results suggest that MR-YOLO provides a compact and GPU real-time solution for microscopic blood-cell detection.