基于Transformer与Vector Loss模块的椎骨Cobb角点定位网络

Vertebral Cobb corner localization using neural network with Transformer and Vector Loss modules

目的:通过定位脊椎骨角点计算Cobb角度数。方法:使用神经网络的方法进行脊椎骨角点坐标的定位,通过嵌入Transformer与Vector Loss模块解决了在清晰度不高、拍摄角度不同的X光图像上计算Cobb角度数精确不高的问题。结果:在MICCAI 2019公开脊椎挑战赛数据集中,本文方法的平均对称百分比(SMAPE)高达9.01。相较于最新的方法,该方法在SMAPE值上提高了1.80。结论:本文所提出的算法嵌入Transformer与Vector Loss模块,与现有的诸多算法相比,具有较高的准确率和稳健性,可以辅助临床医生选择适合患者的治疗方案。

Objective To calculate the Cobb angle by locating the vertebral corners.Methods Neural network was used to locate the coordinates of the vertebral corners.By embedding Transformer and Vector Loss modules,the problem of the poor accuracy in calculating the Cobb angle in X-ray images with low definition and different shooting angles was solved.Results In the MICCAI 2019 Open Spine Challenge data set,the proposed method achieved a symmetric mean absolute precentage error(SMAPE)as high as 9.01.Compared with the latest methods,the proposed method improved the SMAPE by 1.80.Conclusion The proposed algorithm with Transformer and Vector Loss modules is superior to the existing algorithms in accuracy and robustness,and can assist clinicians in selecting the treatment schemes suitable for patients.