引入注意力机制的后交叉韧带断裂的智能辅助诊断

Intelligent Diagnosis of Posterior Cruciate Ligament Deficiency with Self-Attention Module

目的:通过分析行走过程的足底压力数据,研究足底压力信息与后交叉韧带断裂的映射关系,从而实现借助足底压力对后交叉韧带断裂进行快速、准确的智能辅助诊断。方法:回顾性研究了北京大学第三医院2015至2017年收治的69名单纯左侧后交叉韧带断裂患者、69名单纯右侧后交叉韧带患者,以及63名健康志愿者在行走过程中的足底压力数据。通过归一化等方法对足底压力信息进行预处理,然后通过引入注意力机制,采用自注意力结构,建立深度神经网络模型,实现对足底压力信息中隐式信息的特征编码与提取,并基于完整步态过程中足底压力最大值特征,对膝关节后交叉韧带断裂做出智能辅助诊断。结果:在数据处理后,得到1208段单纯左侧后交叉韧带断裂、1096段单纯右侧后交叉韧带断裂,以及964段健康人的足底压力数据,随机选择90%作为训练集,10%作为测试集。经过训练的神经网络在测试集上,对单侧后交叉韧带断裂的预测精度达到了92.02%。在测试集上,模型的曲线下面积(AUC)值达到了0.9820,显著高于使用传统卷积神经网络(CNN)方法得到的88.50%的预测精度和0.9218的AUC值。在可解释性方面,通过对训练过程的足底压力分布梯度进行可视化操作,可以观察到所提出的神经网络准确地提取了足底压力的边缘特征和重点压力特征区域。结论:借助采用了注意力机制的深度神经网络,能够有效地提取行走过程中足底压力信息的空间和时序特征,最终实现对后交叉韧带断裂的辅助智能诊断。这种基于人工智能的诊断方法具备显著的临床应用与研究价值。

Objective To figure out the mapping relationship between plantar pressure and posterior cruciate ligament deficiency(PCLD) by analyzing pressure data during walking and to realize a rapid and accurate intelligent auxiliary diagnosis of posterior cruciate ligament deficiency with the help of the plantar pressure. Methods The plantar pressure data of 69 patients with left PCLD and 69 patients with right PCLD admitted to Peking University Third Hospital in years 2015,2016 and 2017,as well as 63 healthy volunteers were studied retrospectively. The data were pretreated using the normalization method. A deep neural network including a self-attention structure was proposed to encode and extract latent features. The maximum plantar pressure during the entire gait process was employed to make an accurate intelligent diagnosis of PCLD. Results After pretreatment,1208 pieces of right PCLD pressure data,1096 pieces of left PCLD pressure data,and 964 pieces of normal data were generated. Then90% and 10% of the entire dataset were selected randomly as the training set and test set respectively. We got a 92.02% prediction accuracy on the test set using the well-trained deep neural network and obtain a 0.9820 area under curve score,significantly higher than that of convolutional neural network models. Our model extracted the edge of plantar pressure precisely and correctly,and the critical area of the plantar pressure by visualizing the heat map of the gradient of plantar pressure data during the training process. Conclusion With the help of the deep neural network using the self-attention structure,our model can effectively extract the spatial and temporal characteristics of plantar pressure information during walking,and realize the intelligent diagnosis of posterior cruciate ligament exercise. This diagnosis method based on artificial intelligence is of great clinical application and research value.