基于腹腔镜超声的影像组学机器学习模型预测肾脏小肿块良、恶性的价值

Machine Learning Based on Laparoscopic Ultrasound Radiomics for Discriminating Benign from Malignant Small Renal Masses

目的探讨基于腹腔镜超声的不同影像组学机器学习模型预测最大直径≤4 cm的肾脏肿块良、恶性的价值。方法回顾性分析2012年12月至2019年12月在华中科技大学同济医学院附属同济医院行腹腔镜手术肾脏实质性肿块患者的术中腹腔镜超声检查资料,根据病理结果分为恶性组(n=80)和良性组(n=62),随机选取60%的病例为训练组(n=84),40%的病例为验证组(n=58)。采用Pyradiomics包提取良性组和恶性组病灶的107个影像组学特征。采用独立样本t检验或Mann-Whitney U检验对训练组中的特征进行筛选,基于支持向量机(SVM)、随机森林(RF)、逻辑回归(LR)、极限学习机(ELM)和K最近邻(KNN)这5种机器学习方法建模,采用ROC曲线评估模型诊断效能,并通过验证组数据验证模型的稳定性。结果一共142个病灶纳入研究,影像组学方法初步提取107个特征,筛选出在肾脏肿块良、恶性分类中起主要作用的10个高度重复且非冗余的稳定特征。5种机器学习模型的预测性能分别为:SVM(ROC下面积、敏感度、特异度、准确度分别为0.816、0.882、0.750、0.828)、RF(ROC下面积、敏感度、特异度、准确度分别为0.881、0.971、0.792、0.897)、LR(ROC下面积、敏感度、特异度、准确度分别为0.866、0.794、0.708、0.759)、ELM(ROC下面积、敏感度、特异度、准确度分别为0.808、0.824、0.792、0.810)和KNN(ROC下面积、敏感度、特异度、准确度分别为0.831、0.912、0.750、0.845)。结论基于腹腔镜超声影像组学的机器学习模型可以区分肾脏小肿块的良、恶性。

Objective To investigate the feasibility of machine learning models based on laparoscopic ultrasound radiomics in discriminating benign and malignant renal masses(≤4 cm).Methods From December 2012 to December 2019,142 patients with laparoscopic surgery of renal masses performed in Tongji Hospital of Tongji Medical College of Huazhong University of Science and Technology were included in the study.There were 80 malignant cases and 62 benign cases according to pathological results of renal masses.The 142 cases were randomly divided into training group(n=84)and validation group(n=58)with a ratio of 6∶4.Radiomics features of renal masses on laparoscopic ultrasound two-dimensional images were extracted using pyradiomics software.The independent sample T-test or Mann-Whitney U test was used to compare features between benign and malignant cases in the training group,and features with statistical differences were selected.Based on the significant features,five machine learning models were established:Support Vector Machine(SVM),Random Forest(RF),Logistic Regression(LR),Extreme Learning Machine(ELM)and K nearest neighbor(KNN).Receptor operator of characteristic(ROC)curve was constructed to evaluate the performance of each model.Models were verified in the validation group for the stability.Results A total of 142 renal masses were analyzed in our study.107 radiomic features were initially extracted from laparoscopic ultrasound images.10 significant features were selected to construct prediction models because of high reproducibility and little redundancy.The performance of machine learning models in discriminating benign and malignant masses was as follows:SVM(AUC,sensitivity,specificity,accuracy:0.816,0.882,0.750,0.828),RF(AUC,sensitivity,specificity,accuracy:0.881,0.971,0.792,0.897),LR(AUC,sensitivity,specificity,accuracy:0.866,0.794,0.708,0.759),ELM(AUC,sensitivity,specificity,accuracy:0.808,0.824,0.792,0.810)and KNN(AUC,sensitivity,specificity,accuracy:0.831,0.912,0.750,0.845).Conclusion Machine learning models based on laparoscopic ultrasound radiomics are feasible in discriminating malignant masses from benign masses in kidney.