改进2.5DU-Net的骨盆与股骨联合分割方法

打开文本图片集

关键词：深度学习；卷积神经网络；注意力机制；CT影像；髋关节分割DOI：10.15938/j. jhust.2025.04.012中图分类号：TP391.4 文献标志码：A 文章编号：1007-2683（2025）04-0111-12

Enhanced Method for Pelvis and Femur Joint Segmentation Using Improved 2. 5D U-Net

IU Zhaowen1， HE Bingtao¹， YANG Lei²， ZHANG Ruiqi，KONG Zhe'，GUO Ziyu 1 ， SONG Da 2 （20号 （1.College of Computer and Control Engineering，Northeastern Forestry University，Harbin 150040，China; 2.Department of Orthopedics，The First Hospital of Harbin Medical University，Harbin 15Oo01，China）

Abstract：Aimingattheproblemofpoorclarityof hipCTimagesandrelativelynarowjointspace，whichleads tothedifcultyof pelvisandfemurjointsegmentation，animproved.5DU-Netsegmentationmethodisproposed.Themethod processisdivided into three steps：（1）Adata normalization methodisproposedtotransformthe3Dhipsegmentationtask intoa2Dsegmentationtaskusing a2.5Dsegmentationmethod.（2）Using theU-Netnetworkmodelasthebackbonenetwork，theatentionmechanismisintroducedto enhancetefeatureextractionabilityofthenetwork，andthepointswithpoorsegmentationaccuracyareselectedfromtheclasification probability map for individual training，thus correcting the prediction results.（3） The proposed LMFD loss function is employed to solve thetrainigsampleclassifcationimbalanceproblem，whilemonitoringthenetworktrainingtoimprovethemodeljintsegmetation performance.Theresultsdemonstratethattheproposedmethodoutperformsthebaselinenetwork model（3DU-Net）withDice similarity coefcients （DSC） of 94. 75% and 95.74% ，and Hausdorff distances 95 （HD95） of 1.29 mm and 1.59 mm，respectively， for the task of joint segmentation of the pelvis and femur.

KeyWords：deep learning；convolutional neural network；attention mechanism；CT images；hip joint segmentation

0 引言

计算机辅助诊断已经成为骨科手术中必不可少的部分[1-2]。（剩余18173字）