基于深度学习的负重位锥形束CT下肢骨骼自动分割及髋-膝-踝角测量的研究  

Deep learning-based automatic segmentation of lower limb bones and hip-knee-ankle angle measurement using weight-bearing cone-beam CT

作　　者：李函宇 梁泽军 何建容 荣繁壮 余晓成 夏春潮[2] 李真林[2] 唐静[2] Han-yu;LIANG Ze-jun;HE Jian-rong;RONG Fan-zhuang;YU Xiao-cheng;XIA Chun-chao;LI Zhen-lin;TANG Jing(Department of Equipment and Materials,West China Hospital,Sichuan University,Chengdu,Sichuan,610041,China)

机构地区：[1]四川大学华西医院设备物资部,成都610041 [2]四川大学华西医院放射科,成都610041 [3]深圳市安健科技股份有限公司研发中心,广东省518000

出　　处：《中国骨与关节杂志》2024年第11期902-907,共6页Chinese Journal of Bone and Joint

摘　　要：目的构建基于深度学习的负重位锥形束CT(cone beam computed tomography,CBCT)下肢骨骼自动分割模型,实现三维髋-膝-踝(hip-knee-ankle,HKA)角度的自动测量,并比较膝关节炎患者二维与三维HKA角的差异。方法本研究纳入65例负重位CBCT全下肢三维图像,包括训练集(60例)和测试集(5例),分别用于下肢各骨骼自动分割的卷积神经网络(CNN)模型的构建和Dice系数精度评估。另外,纳入56例(112个下肢)膝关节炎患者分别进行双侧全下肢的二维数字化X射线(digital radiography,DR)和负重位CBCT扫描,DR图像用于手动测量二维HKA角,CBCT图像用于自动测量三维HKA角,通过配对t检验分析二维与三维HKA角的差异。结果在测试集中,自动分割模型对下肢各骨骼的分割Dice系数分别为:股骨0.950,髌骨0.928,胫骨0.946,腓骨0.947,距骨0.946,该分割模型的精度确保了三维HKA角测量的准确性。进一步分析显示,二维与三维HKA角存在显著差异[MD±SD:(-1.23±3.07),MAD±SD:(2.05±2.60),P<0.001]。结论二维与三维HKA角存在显著差异;且本研究构建的基于负重位CBCT图像的下肢骨骼自动分割模型具有较高的分割准确性,为下肢骨骼三维参数的自动测量提供了技术基础。Objective To develop a deep learning-based automatic segmentation model for the lower limb bones using weight-bearing cone-beam CT(CBCT)and achieve automatic measurement of three-dimensionalhip-knee-ankle(HKA)angles.Additionally,to compare and analyze the differences in HKA angles measured by two-dimensional digital X-ray(DR)and three-dimensional CBCT in patients with knee osteoarthritis.Methods65 patients underwent full lower extremity 3D imaging with CBCT in weight-bearing position.The data were divided into a training set(60 cases)and a test set(5 cases),which were used for the construction of the convolutional neural network(CNN)model for automatic segmentation of bones of the lower limbs and the Dice coefficient accuracy evaluation,respectively.Additionally,56 patients with knee osteoarthritis underwent both bilateral full lower limb DR and weight-bearing CBCT scans.DR images were used for manual measurement of two-dimensional HKA angles,while CBCT images were used for automatic measurement of three-dimensional HKA angles.The differences between the two-dimensional and three-dimensional HKA angles were analyzed for a total of 112 lower limbs using paired t-test.Results In the test set,the Dice coefficients for the automatic segmentation model of each lower limb bone were:femur 0.950,patella 0.928,tibia 0.946,fibula 0.947,and talus 0.946.The accuracy of this segmentation model ensured the precision of the three-dimensional HKA angle measurements.Further analysis showed significant differences between the two-dimensional and three-dimensional HKA angles(MD±SD:-1.23±3.07,MAD±SD:2.05±2.60,P<0.001).Conclusions There are significant differences between two-dimensional and three-dimensional HKA angles.Additionally,the lower limb bone automatic segmentation model based on weight-bearing CBCT images developed in this study has high segmentation accuracy,providing technical foundation for the automatic measurement of three-dimensional lower limb bone parameters.

关 键 词：锥形束计算机断层扫描 深度学习 负重位 膝骨关节炎 髋-膝-踝角 

分 类 号：R682[医药卫生—骨科学]