基于多模态磁共振放射组学与临床指标的前列腺癌智能检测及风险预测模型建立

Establishment of models for the intelligent detection and risk prediction of prostate cancer based on the combination of multi-modality magnetic resonance imaging radiomics and clinical indicators

目的:利用多模态磁共振放射组学开发前列腺癌自动检测模型,并使用列线图构建多因素回归模型,将前列腺MRI放射组学特征与临床多个检测指标进行整合,从而对患前列腺癌风险性进行预测。方法:回顾性研究于2019年2月~2021年10月病理证实为前列腺癌和其他前列腺良性肿瘤的患者133例。所有病例均行前列腺直肠指检(DRE)、前列腺特异性抗原(PSA)、游离前列腺特异性抗原(F-PSA)、FPSA/PSA检测。治疗前多模态前列腺MRI图像(DWI+DCE+T2WI)用于提取放射特征,最大相关最小冗余(m RMR)算法用于消除混杂变量,使用最小绝对收缩和选择算子(LASSO)逻辑回归进行放射特征选择。通过曲线下面积(AUC)、准确性、特异性、敏感性评估放射特征的诊断性能;通过多元logistic回归选择临床指标和放射组学特征模型来制定放射组学列线图,并使用校准曲线和Hosmer-lemeshow试验验证其可靠性。结果:两名观察者测量的所有数据ICC均在0.80以上。所有前列腺MRI图像随机分为训练组和验证组(7:3)。在训练组中,DWI、DCE和T2WI的AUC分别为0.882、0.821、0.848,在验证组中,DWI、DCE和T2WI的AUC分别为0.861、0.810、0.838;三模态联合模型的训练组和验证组AUC分别为0.912、0.898。Delong检验结果显示DWI模型性能优于DCE和T2WI模型性能,DCE和T2WI模型检测性能相仿,三模态联合模型性能优于任意一种模型性能。使用ROC曲线评估列线图、影像组学和临床指标的预测性能,结果显示列线图的AUC值为0.941,准确率、敏感性、特异性分别为0.929、0.891、0.893。列线图前列腺癌预测性能最好,临床指标的预测性能较差,校准曲线和Hosmer-Lemeshow检验结果也验证了上述观点。结论:多模态前列腺MRI放射组学模型能准确鉴别前列腺肿瘤的良恶性,放射组学列线图在前列腺癌风险预测中表现出令人满意的效果。

Objective To develop an automatic detection model of prostate cancer using multi-modality magnetic resonance imaging (MRI) radiomics,and to predict the risk of prostate cancer by a multifactor regression model constructed by nomogram based on the combination of prostate MRI radiomics and clinical indicators.Methods A retrospective analysis was conducted on 133 patients with prostate cancer and other benign prostatic lesions confirmed by pathology from February 2019 to October 2021.All patients underwent directeral rectum examination (DRE),and were tested for prostate specific antigen (PSA),free-prostate specific antigen (F-PSA) and F-PSA/PSA.After extracting radiological features from multi-modality prostate MRI images (DWI+DCE+T2WI) before treatment,the minimal redundancy maximal relevance (mRMR)algorithm was used for eliminating hybrid variables,and the least absolute shrinkage and selection operator (LASSO) for radiological feature selection.The diagnostic performances of radiological features were evaluated by area under ROC curve(AUC),accuracy,specificity and sensitivity.Multiple logistic regression analysis was used to select clinical indicators which were then combined with radiomics feature model to formulate radiomics nomogram.The model reliability was verified by calibration curve and Hosmer-lemeshow test.Results The ICC of all data measured by two observers was above 0.80.All MRI images of the prostate were randomly divided into training group and verification group at a ratio of 7:3.The AUC of DWI,DCE and T2WI were 0.882,0.821,0.848 in training group,and 0.861,0.810,0.838 in verification group,while the combination model of triple-modality MRI achieved AUC of 0.912 and 0.898 in training group and validation group,respectively.The Delong test results show that DWI model outperformed DCE and T2WI models (the latter two had similar performances),and that the performance of combination model of triple-modality MRI was superior to that of any other model.ROC curve was used to evaluate the predictive performance of nomogram,radicomics and clinical indicators,and the results revealed that the AUC,accuracy,sensitivity and specificity of nomogram were 0.941,0.929,0.891 and 0.893,respectively.Nomogram had the best predictive performance for prostate cancer,and the predictive performance of clinical factors was poor.Both calibration curve and Hosmer-lemeshow test results verified the above findings.Conclusion Multi-modality prostate MRI radiomics model can accurately identify benign and malignant prostate tumors.Radiomics nomogram shows a satisfactory performance in the prediction of prostate cancer risk.