基于卷积神经网络的计算机自动学习系统判断全膝关节置换手术指征的初步研究

Explore the judgmental of the indications of total knee arthroplasty using deep convolutional neural network

目的探讨基于卷积神经网络（deep convolutional neural network，DCNN）的计算机自动学习系统判断全膝关节置换手术指征和预后的可能性。方法构建基于CIFAR-10的DCNN模型，从数据库中经筛选复核得到400例膝关节骨关节炎病例用于分析，包括手术治疗病例300例、非手术治疗病例100例。选取200例术前X线片及其扩增50倍后的10000例图像对DCNN进行训练，再采用经过训练的DCNN对另外200例术前X线片扩增图像的手术指征加以判断。临床标准经过三次复核一致并经过临床疗效验证，作为金标准。采用Pearson相关分析评估DCNN判断结果与临床标准的相关性，DCNN内建精度分析用作参考。结果用于DCNN机器学习的200例与用于判断的200例患者一般情况和治疗方法的差异无统计学意义。经200例学习后，DCNN机器判断结果与临床标准没有相关性（r=0．000，F=0．001，P=0．970）。判断的假阳性率为16．8％（1681／10000），假阴性率为33．0％（3296／10000）；扩增10000例后，DCNN机器判断结果与临床标准之间有相关性（r=0．727，止11228．735，P=0．000，R^2=0．529）。软件内建评估的最高精度为0．860。结论基于卷积神经网络的计算机自动学习系统有望用于判断全膝关节置换手术指征，大样本数据学习后判断准确率将显著提升。

Objective To explore the feasibility of the deep convolutional neural network （DCNN） judging the indica- tions and prognosis of the total knee arthroplasty based on the trained DCNN computer learning system. Methods CIFAR- 10 DCNN model based on TensorFlow （an open source system, Google, USA） optimized by Alex Krizhevsky were constructed. There were 400 cases with knee osteoarthritis from different databases used for analysis. Three hundred patients underwent total knee ar- throplasty, while 100 did not. X-ray of 200 preoperative cases from the 400 cases and their enlarged image （50 times） were applied for training DCNN, while the enlarged images from other 200 cases were used to test the DCNN. The correlation and the regression between judgment of the DCNN and clinical truth were analyzed. The clinical truths were rechecked three times and were con- finned by treatment results. Pearson correlation and linear regression analysis were used. The relation test of the software was only used as a reference. Results There was no significant difference between the baseline of cases for learning and test. After learning 200 cases, the DCNN judged the 10 000 cases enlarged from remaining 200 cases. The correlation between the DCNN judgment and the clinical truth was not significant （r=0.000, F=0.001, P=0.970）. False positive was observed in 1 681 cases, false negative in 3 296. After enlarged to 10 000 images, the correlation between the two judgments was significant （F=11 228.735, P= 0.000, r=0.727 and R2=0.529）. The software detection precision was 0.860. Conclusion DCNN can be applied in judging the indications of the total knee arthroplasty. Large sample size can improve the accuracy of the judgment significantly.