CT纹理特征联合机器学习对发生骨质疏松性压缩骨折的预测价值

Predictive value of CT texture features combined with machine learning for the osteoporotic vertebral compression fractures

目的:探究CT纹理特征联合机器学习对发生骨质疏松性压缩骨折(OVCFs)的预测价值。方法:回顾性分析45例新鲜的腰椎OVCFs患者的CT图像(实验组,选择相邻的无骨折的椎体共152个)和同期45例无OVCFs患者的腰椎椎体的CT图像(对照组,选择所有腰椎椎体共225个)。在CT图像上勾画ROI,并提取纹理特征参数。将所有患者按8∶2比例分为训练集和验证集,用费希尔算法(Fisher)、分类错误率+平均相关系数算法(POE+ACC)、交互信息算法(MI)3种方法对提取的特征参数进行降维筛选,再经t检验/秩和检验、Spearman相关分析做进一步筛选,采用筛选出的参数建立人工神经网络分类模型,结合5倍交叉验证法评估其效能。结果:经筛选,获得S(0,3)SumAverg、WavEnLL_s-4、Perc.50%3个CT纹理特征参数,其中S(0,3)SumAverg的曲线下面积(AUC)最高(0.817)。建立人工神经网络分类模型后,训练集和验证集的AUC分别为0.906和0.867。结论:CT纹理特征结合人工神经网络分类模型对可能发生OVCFs的患者的预测效果较好。

Objective:To investigate the predictive value of CT texture features combined with machine learning for the osteoporotic vertebral compression fractures(OVCFs).Methods:CT images of 45 patients with OVCFs in lumbar spine(case group,a total of 152 adjacent vertebrae without fractures)and 45 patients without OVCFs in lumbar spine(control group,a to-tal of 225 lumbar vertebrae)were retrospectively studied.ROIs were outlined on their CT images,and texture feature parame-ters were extracted.All patients were divided into training and testing sets in the ratio of 8∶2.The extracted feature parame-ters were filtered by three methods:Fisher’s algorithm(Fisher),classification error rate+average correlation coefficient algorithm(POE+ACC),and interaction information algorithm(MI)for dimensionality reduction,and then further filtered by t-test/rank sum test and Spearman correlation analyses.The filtered parameters were used to establish the performance of artificial neural net-work classification model,which was evaluated by 5-fold cross-validation.Results:After screening,three CT texture feature parameters,S(0,3)SumAverg,WavEnLL_s-4,and Perc.50%,were obtained,among which S(0,3)SumAverg had the highest AUC value(0.817).After building an artificial neural network classification model,the AUCs of the training and testing sets were 0.906 and 0.867,respectively.Conclusion:CT texture features combined with artificial neural networks are effective in predicting patients with possible OVCFs.