MS-YOLO: A Multi-Scale Model for Accurate and Efficient Blood Cell Detection

2025-06-04

AI Summary

Addressing challenges in detecting overlapping and multi-scale objects—particularly tiny platelets—in blood cell microscopic images, including low accuracy, high deployment costs, and poor generalization of existing methods, this paper proposes a lightweight and efficient object detection framework. Methodologically, it integrates dilated convolutions, a dynamic feature pyramid, and spatially weighted downsampling into the YOLOv11 architecture. Key innovations include a Multi-Scale Dilated Residual Module (MS-DRM), a Dynamic Cross-Path Feature Enhancement Module (DCFEM), and a Lightweight Adaptive Weighted Downsampling Module (LADS), collectively enhancing multi-scale discriminability, cross-layer feature fusion quality, and computational efficiency. Evaluated on the CBC benchmark, the method achieves 97.4% mAP@50—significantly surpassing state-of-the-art approaches. It also demonstrates strong generalization on the WBCDD dataset and enables real-time inference, fulfilling clinical deployment requirements.

Technology Category

• Computer Vision: Object Detection & Categorization
• Machine Learning: Hardware-aware ML
• Search and Optimization: Learning to Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Complete blood cell detection holds significant value in clinical diagnostics. Conventional manual microscopy methods suffer from time inefficiency and diagnostic inaccuracies. Existing automated detection approaches remain constrained by high deployment costs and suboptimal accuracy. While deep learning has introduced powerful paradigms to this field, persistent challenges in detecting overlapping cells and multi-scale objects hinder practical deployment. This study proposes the multi-scale YOLO (MS-YOLO), a blood cell detection model based on the YOLOv11 framework, incorporating three key architectural innovations to enhance detection performance. Specifically, the multi-scale dilated residual module (MS-DRM) replaces the original C3K2 modules to improve multi-scale discriminability; the dynamic cross-path feature enhancement module (DCFEM) enables the fusion of hierarchical features from the backbone with aggregated features from the neck to enhance feature representations; and the light adaptive-weight downsampling module (LADS) improves feature downsampling through adaptive spatial weighting while reducing computational complexity. Experimental results on the CBC benchmark demonstrate that MS-YOLO achieves precise detection of overlapping cells and multi-scale objects, particularly small targets such as platelets, achieving an mAP@50 of 97.4% that outperforms existing models. Further validation on the supplementary WBCDD dataset confirms its robust generalization capability. Additionally, with a lightweight architecture and real-time inference efficiency, MS-YOLO meets clinical deployment requirements, providing reliable technical support for standardized blood pathology assessment.

Problem

• Improves detection of overlapping blood cells
• Enhances accuracy for multi-scale cell detection
• Reduces computational costs for clinical deployment

Innovation

• Multi-scale dilated residual module enhances discriminability
• Dynamic cross-path feature fusion improves representation
• Light adaptive-weight downsampling reduces complexity