nnU-Net级联网络在膝关节MRI图像自动分割中的应用  

Application of the Nnu-net Cascaded Network in Automatic Aegmentation of Knee Joint MRI Images

作　　者：程羽 韩晓奇[2] 李效凯 CHENG Yu;HAN Xiaoqi;LI Xiaokai(College of Sports and Health,Shanghai University of Sport,Shanghai 200438,China;Magnetic Resonance Department,The Fifth People's Hospital of Fuyang City,Anhui 236035,China)

机构地区：[1]上海体育大学运动健康学院,上海200438 [2]阜阳市第五人民医院磁共振科室,安徽236035

出　　处：《影像技术》2025年第2期64-69,共6页Image Technology

摘　　要：目的:本研究首次深入探讨基于nnU-Net级联网络的膝关节MRI图像自动分割模型,并通过多个指标对其性能进行全面评估。方法:采用创新的级联分割网络管道结合低分辨率和全分辨率的3D双网络结构,实现对膝关节MRI图像的高效粗分割和精细分割。结果:在SKI10数据集上的验证显示,该模型在膝关节股骨、股骨软骨、胫骨和胫骨软骨的分割上取得了优异的Dice相似系数(DSC),分别为0.98±0.01、0.80±0.05、0.98±0.02和0.78±0.05,且准确率(Acc)、精确率(Prec)、召回率(Rec)和交并比(IOU)均超过0.88。结论:相较于现有方法,本研究在维持高分割精度的同时实现了更高效的计算。该模型有望为临床膝关节疾病的精准诊断和治疗提供新的技术支持。Objective:This study provides the first in-depth exploration of an automatic knee joint MRI image segmentation model based on the nnU-Net cascaded network,and comprehensively evaluates its performance using multiple metrics.Methods:We employ an innovative cascaded segmentation network pipeline integrating low-resolution and full-resolution 3D dual-network structures to achieve efficient coarse segmentation and fine segmentation of knee joint MRI images.Results:Validation on the SKI10 dataset demonstrates outstanding Dice similarity coefficients(DSC)for segmentation of the femur,femoral cartilage,tibia,and tibial cartilage,with scores of 0.98±0.01,0.80±0.05,0.98±0.02,and 0.78±0.05,respectively.Additionally,accuracy(Acc),precision(Prec),recall(Rec),and intersection over union(IOU)all surpass 0.88.Conclusion:Compared to existing methods,this study achieves greater computational efficiency while maintaining high segmentation accuracy.This model holds promise for providing new technical support for precise diagnosis and treatment of clinical knee joint diseases.

关 键 词：膝关节 nnU-Net级联网络 MRI 自动分割 

分 类 号：R684.3[医药卫生—骨科学]