基于超声图像特征机器学习预测前列腺癌危险度的价值

The value of machine learning based on ultrasound image features in predicting the risk of prostate cancer

目的:建立联合前列腺经直肠超声图像特征及临床数据的决策树、K近邻、贝叶斯网络、Logistic回归、支持向量机5种机器学习模型,评价上述模型预测前列腺癌危险度的价值。资料与方法:回顾性分析我院接受前列腺经直肠超声检查并确诊为前列腺癌的198例患者。将经直肠前列腺超声图像特征、年龄、总前列腺特异抗原、游离前列腺特异抗原、游离前列腺特异抗原比值、体积、前列腺特异抗原密度录入相关属性值分析,通过单个属性和类别的相关性分析以分析特征。后将这些超声图像特征及临床数据录入5种机器学习模型进行训练及验证,通过F1值及ROC曲线评价机器学习模型的预测价值。结果:贝叶斯网络模型预测前列腺癌危险度ROC的曲线下面积(AUC)最大(0.9278),K近邻AUC也较高(0.907),logistic回归最小(0.717);F1值最高的为贝叶斯网络,其次依序为支持向量机、决策树模型、K近邻及Logistic回归模型。通过相关属性值分析各特征变量重要性,前列腺内外腺分界是否清晰与前列腺癌危险度相关性最高,其次为前列腺包膜完整度、前列腺对称性、前列腺外腺腺体血流情况,结节灶回声对模型分类贡献最低。结论:基于超声图像特征的贝叶斯网络模型预测前列腺癌危险度的性能最优。

Objective:Five machine learning models including decision tree,K-nearest neighbor,Bayesian network,logistic regression,and support vector machine combined with prostate transrectal ultrasound image characteristics and clinical data were established to evaluate the value of these models in predicting the risk of prostate cancer.Materials and Methods:A retrospective analysis of 198 patients who underwent transrectal ultrasound examination of the prostate in our hospital to obtain pathological results and diagnosed prostate cancer.Characteristics of transrectal prostate ultrasound image,age,total prostate specific antigen,free prostate specific antigen,free prostate specific antigen ratio,volume,and prostate specific antigen density were entered into relevant attribute value analysis,and the characteristics through correlation analysis of individual attributes and categories were analyzed.Later,these ultrasound image features and clinical data were entered into five machine learning models for training and verification,and the predictive value of the machine learning models was evaluated through the F1 value and ROC curve.Results:Bayesian network model predicts prostate cancer risk ROC with the largest area under the curve(0.9278),K-nearest neighbor AUC was also higher(0.907),logistic regression was the smallest(0.717);Bayesian network had the highest F1 value,followed by Support vector machine,decision tree model,K nearest neighbor and logistic regression model.The importance of each characteristic variable was analyzed by the relevant attribute values.Whether the boundary between the inner and outer glands of the prostate is clear was the highest correlation with the risk of prostate cancer,followed by the integrity of the prostate capsule,the symmetry of the prostate,the blood flow of the glands outside the prostate,and the nodules.The echo contributed the least to the model classification.Conclusion:The Bayesian network model based on ultrasonic image features is the best model for predicting PCa risk.