前列腺增强磁共振成像肿瘤的影像组学良恶性自动判别算法

Auto Identifying Algorithm of Benign and Malignant Tumors of Enhanced MRI Radiomics of Prostate

目的:运用放射组学从前列腺肿瘤磁共振增强图像中提取高通量的病灶纹理特征,结合套索算法(LASSO)回归算法进行特征筛选,利用筛选的纹理特征建立前列腺增强磁共振分类模型来自动检测前列腺癌。方法:回顾性分析经3.0 T磁共振成像(MRI)增强扫描的前列腺肿瘤患者118例,经病理证实前列腺癌63例,前列腺良性肿瘤55例,增强图像在ITK-SNAP 3.6.0上进行手工分割。使用A.K.软件对前列腺良恶性肿瘤总共118个纹理特征进行定量分析,去除相关系数大于0.9的特征以消除重复冗余。用最大相关和最小冗余(m RMR)以及LASSO两种特征选择方法来选择训练队列中最有用的预测特征,使用LASSO回归建立基于前列腺增强MRI肿瘤内基质和肿瘤内基质包含外周组织的两个良恶性自动检测模型。通过受试者工作特征曲线(ROC)和Delong检验曲线分析和评价模型的性能。结果:测试组的肿瘤内基质组模型曲线下面积(AUC)为0.819,肿瘤基质及外周组模型AUC为0.865。当预测得分阈值为0.5时,肿瘤内基质组模型、肿瘤基质及外周组模型的灵敏度分别为0.750和0.776,特异度分别为0.875和0.912,准确度为0.823和0.862。对两组ROC曲线进行Delong检验,P=0.0134。结论:基于磁共振增强图像的放射组学模型结合LASSO算法自动检测前列腺癌,表现出较高的性能。前列腺肿瘤基质及外周组的性能显著优于肿瘤内基质组。

Objective Radiomics was used to extract high-throughput texture features from the enhanced magnetic resonance imaging(MRI)of prostate tumors,combined with least absolute shrinkage and selection operator(LASSO)regression algorithm for feature selection.The selected texture features were used to establish a prostate enhanced MRI classification model to automatically detect prostate cancer.Methods A total of 118 patients with prostate cancer who underwent enhanced 3.0T MRI scan were retrospectively analyzed,including 63 cases of prostate cancer and 55 cases of benign prostate tumors confirmed by pathology.The enhanced images were manually segmented on ITK-SNAP 3.6.0.A total of 118 texture features of benign and malignant prostate tumors were quantitatively analyzed by A.K.software,and features with correlation coefficients greater than 0.9 were removed to eliminate repetitive redundancy.Two feature selection methods,max-relevance and min-redundancy(mRMR)and LASSO were used to select the most useful predictive features in the training cohort.LASSO regression was used to establish two automatic detection models for benign and malignant lesions based on tumor stroma and tumor stroma containing peripheral tissues.The performance of the model was analyzed and evaluated by receiver operating characteristic(ROC)curve and Delong test curve.Results The area under curve(AUC)of the test group was 0.819 in the tumor stroma group,and the AUC of the tumor stroma and peripheral group was 0.865.When the threshold of prediction score was 0.5,the sensitivity,specificity and accuracy of the tumor stroma group model,thetumor stroma and peripheral group model were 0.750 and 0.776,0.875 and 0.912,0.823 and 0.862,respectively.Delong test was performed on the ROC curve of the two groups,P=0.0134.Conclusion The radiomics model based on magnetic resonance enhanced image combined with LASSO algorithm shows high performance in automatically detecting prostate cancer.The performance of intratumoral stroma and peripheral group was significantly better than that of the intratumoral stroma group.