摘要
目的构建基于MRI影像组学特征的机器学习模型,以预测PI-RADS 3前列腺病变的良恶性及侵袭性。方法回顾性分析296例PI-RADS 3前列腺病变病人的MRI影像资料。其中,PCa病人141例,非PCa病人155例。并将病人以7∶3的比例随机划分为训练集和独立验证集。由2名医师使用3D Slicer软件在T_(2)WI、DWI、DCET_(1)WI上手动勾画病变,并分别提取影像组学特征。采用组内相关系数(ICC)评估2名医师病变勾画的稳定性。采用t检验和最小绝对值收敛和选择算子(LASSO)算法进行特征筛选。使用支持向量机(SVM)算法分别构建前列腺病变良恶性预测模型及PCa侵袭性预测模型。采用Mann-Whitney U检验比较2组间前列腺特异性抗原(PSA)水平。采用受试者操作特征(ROC)曲线下面积(AUC)、准确度、敏感度和特异度评估模型的预测效能。结果141例PCa病人的PSA水平高于非PCa病人(P<0.05)。141例PCa病人中,临床有意义PCa(csPCa)100例,临床无意义PCa(ciPCa)41例。csPCa病人的PSA水平也高于ciPCa病人(P<0.05)。共分析296个病灶,每个病灶均提取2553个影像组学特征。2名医师对病变勾画均具有良好的一致性(ICC:组间0.81,组内0.84)。在前列腺病变的良恶性预测中,训练集207个病灶,验证集89个病灶,最终筛选出14个特征(9个DWI、3个T_(2)WI、2个DCE特征),构建的预测模型在训练集中的AUC、准确度、敏感度和特异度分别为0.93(95%CI:0.91~0.95)、0.82、0.78、0.85,在独立验证集中分别为0.89(95%CI:0.86~0.92)、0.81、0.86、0.77;在PCa的侵袭性预测中,训练集98个病灶,验证集43个病灶,最终筛选出12个特征(5个DWI,4个T_(2)WI,3个DCE特征),构建的预测模型在训练集中的AUC、准确度、敏感度和特异度分别为0.92(95%CI:0.89~0.94)、0.85、0.87、0.84,在独立验证集中分别为0.85(95%CI:0.81~0.89)0.72、0.73、0.70。结论基于MRI影像组学特征的机器学习模型能有效预测PI-RADS 3前列腺病变的良恶性及侵袭性。
Objective To construct a machine learning model based on MRI radiomics features to predict malignancy and aggressiveness of PI-RADS 3 prostate lesions.Methods The MRI data of 296 patients with PI-RADS 3 prostate lesions were retrospectively analyzed.Among them,141 were PCa,and 155 were non-PCa.These patients were randomly divided into training and independent validation group at a ratio of 7∶3.Lesions were manually segmented by 2 radiologists using 3 D Slicer software on T_(2)WI,DWI,and DCE-T_(1)WI,and radiomics features were extracted.The intra group correlation coefficient(ICC)was used to evaluate the stability of segmentations by the 2 radiologists.The t test and the least absolute shrinkage and selection operator(LASSO)algorithm were used for feature selection.The SVM algorithm was used to develop a machine learning prediction model for differentiating PCa from non-PCa and high-grade from low-grade PCa.The Mann-Whitney U test was ued to compare the levels of prostate specific antigen(PSA)between the two groups.The area under the receiver operating characteristic(ROC)curve(AUC),accuracy,sensitivity,and specificity were used to evaluate the predictive power of the model.Results The PSA level of 141 PCa patients was higher than that of 155 non-PCa patients(P<0.05).Among141 PCa patients,100 were clinically significant PCa(csPCa)and 41 were clinically insignificant PCa(ciPCa).The PSA level of csPCa patients was also higher than that of ciPCa patients(P<0.05).A total of 296 lesions were analyzed,and 2553 radiomic features were extracted from each lesion.The segmentations by the 2 radiologists had good agreement(ICC:inter-observer 0.81,intra-observer 0.84).For PCa versus non-PCa,207 lesions were asigned into training group,and 89 lesions were asigned into validation group.Ultimately,14 features(9 from DWI,3 from T_(2)WI,and 2 from DCE)were selected,the prediction model had an AUC of 0.93(95%CI:0.91-0.95),accuracy of 0.82,sensitivity of 0.78,and specificity of 0.85 in the training group,and0.89(95%CI:0.86-0.92),0.81,0.86,and 0.77 in the validation group,respectively.For low-grade versus high-grade PCa,98 lesions were asigned into training group,and 43 lesions were asigned into validation group.Ultimately,12 features were selected(5 from DWI,4 from T_(2)WI,and 3 from DCE),the prediction model had an AUC,accuracy,sensitivity,specificity of0.92(95%CI:0.89-0.94),0.85,0.87,and 0.84 in the training group,and 0.85(95%CI:0.81-0.89),0.72,0.73,and 0.70 in the validation group,respectively.Conclusion The machine learning model based on MRI radiomics features has high diagnostic efficacy in distinguishing cancerous vs.noncancerous PI-RADS 3 prostate lesions and high-grade vs.low-grade PI-RADS 3 PCa.
作者
李天平
骆训容
罗明芳
张涵
谢海柱
王培源
LI Tianping;LUO Xunrong;LUO Mingfang;ZHANG Han;XIE Haizhu;WANG Peiyuan(Department of Radiology,Yantai Affiliated Hospital of Binzhou Medical University,Yantai 264100,China;School of Medical Imaging,Binzhou Medical University;Department of Radiology,Yantai Yuhuangding Hospital)
出处
《国际医学放射学杂志》
北大核心
2021年第6期638-643,共6页
International Journal of Medical Radiology
关键词
机器学习
影像组学
前列腺癌
磁共振成像
Machine learning
Radiomics
Prostate cancer
Magnetic resonance imaging
