摘要
针对儿童腕部X光图像中细微骨折检测精度较低的问题,提出一种基于YOLOv11n改进的MSEF-YOLO(multi-scale efficient fusion network-YOLO)目标检测算法。首先,将空间和通道重构卷积(SCConv)模块与C3k2模块融合,通过空间重构单元(SRU)和通道重构单元(CRU)并行处理空间与通道的冗余,增强对小目标的感知能力;其次,引入多尺度扩张注意力(MSDA)机制提高特征提取能力,进而提高模型检测精度与泛化性,有效减少漏检和误检;最后,优化尺度序列特征融合(SSFF)模块并设计SSFF-X模块,通过3D卷积增强多尺度特征融合能力,进一步提升对细微骨折的检测效果。实验结果表明,相较于原YOLOv11n算法,MSEF-YOLO算法的精确率、召回率、mAP@0.5和mAP@0.5~0.95分别提高了3.1%、3.8%、3.0%和3.5%。MSEF-YOLO算法能够有效协助放射科医生检测儿童腕部骨折,为医学图像的诊断提供技术支持。
To address the issue of low detection accuracy for subtle fractures in pediatric wrist X-ray images,an improved YOLOv11n-based object detection algorithm called MSEF-YOLO(multi-scale efficient fusion network-YOLO)is proposed.First,the spatial and channel reconstruction con-volution(SCConv)module is integrated with the C3k2 module,utilizing the spatial reconstruction u-nit(SRU)and channel reconstruction unit(CRU)to process spatial and channel redundancy in par-allel,thereby enhancing the perception of small objects.Second,the multi-scale dilated attention(MSDA)mechanism is introduced to improve feature extraction capability,thereby enhancing de-tection accuracy and generalization ability while effectively reducing missed and false detections.Fi-nally,the scale sequence feature fusion(SSFF)module is optimized and the SSFF-X module is de-signed,leveraging 3D convolution to enhance multi-scale feature fusion,further improving the de-tection performance for subtle fractures.Experimental results demonstrate that,compared to the o-riginal YOLOv11n algorithm,the MSEF-YOLO algorithm improves precision,recall,mAP@0.5,and mAP@0.5~0.95 by 3.1%,3.8%,3.0%,and 3.5%,respectively.The MSEF-YOLO algo-rithm effectively assists radiologists in detecting pediatric wrist fractures,providing technical support for medical image diagnosis.
作者
宫硕
蒋强
李婷雪
GONG Shuo;JIANG Qiang;LI Tingxue(Shenyang Ligong University,Shenyang 110159,China)
出处
《沈阳理工大学学报》
2025年第5期29-36,共8页
Journal of Shenyang Ligong University
基金
辽宁省教育厅高等学校基本科研项目(LJ21241014452)。
关键词
儿童腕部骨折
空间和通道重构卷积
多尺度扩张注意力
特征融合
pediatric wrist fracture
spatial and channel reconstruction convolution
multi-scale di-lated attention
feature fusion
