目的 探讨基于多参数MRI的影像组学融合模型术前预测直肠癌脉管侵袭（LVI）的应用价值。 方法 回顾性队列研究。纳入2016年1月—2019年12月在山西省肿瘤医院行直肠癌根治性切除术的224例直肠癌患者的临床病理资料和多参数MRI数据，其中男129例、女95例，年龄28~83（61.3±9.7）岁。按7∶3的比例随机分为训练集157例和验证集67例。使用ITK-SNAP图像分割软件，在高分辨率T2加权像（T2WI）、弥散加权成像（DWI）及增强T1加权像（cT1WI）上，逐层手动勾画肿瘤感兴趣区得到全容积感兴趣区（VOI），将DWI的勾画信息复制到表观弥散系数（ADC）图像（由2个不同b值的DWI图像自动计算生成）。采用最小冗余最大相关（mRMR）、最小绝对收缩与选择算子（LASSO）算法及多因素logistic回归的三步降维法，筛选影像组学特征并构建影像组学标签。通过多因素logistic回归分析筛选出临床病理特征和MRI表现特征中的独立预测因子。分别构建基于T2WI、ADC、cT1WI单一序列和联合序列的影像组学模型及纳入临床病理特征后的融合模型并制作列线图。采用受试者操作特征曲线下面积（AUC）、校准曲线、决策曲线（DCA）评估模型的效能及临床效益。 结果 224例直肠癌患者经术后病理证实LVI阳性70例、阴性154例。训练集和验证集的临床病理特征及MRI表现特征比较，差异均无统计学意义（P值均>0.05）。三步降维法筛选后得到6个与直肠癌LVI相关的关键特征（P值均<0.05）。血清癌胚抗原（CEA）是直肠癌LVI的独立预测因子［比值比（95%可信区间）2.071（1.038~4.131），P=0.039］。基于T2WI、ADC、cT1WI单一序列及联合序列的影像组学模型在训练集中的AUC分别为0.765、0.772、0.776、0.878，在验证集中的AUC分别为0.741、0.739、0.764、0.846；纳入CEA构建的融合模型训练集和验证集的AUC分别为0.899、0.876，预测效能最佳。校准曲线显示融合模型有良好的校正性能。验证集的DCA曲线显示，阈值概率范围在0.10~0.20和0.35~0.90时融合模型的净收益最大。 结论 基于多参数MRI的影像组学特征和CEA构建的融合模型在直肠癌LVI的术前预测中有较高的诊断效能，其可视化列线图可以作为术前预测LVI的有效工具。 Objective This study aims to investigate the application value of a fusion model based on multiparameter magnetic resonance imaging (MRI) for the preoperative prediction of lymphovascular invasion (LVI) in rectal cancer. Methods Retropective cohort study was conduted. The clinicopathological data and multi-parameter MRI data of 224 patients with rectal cancer who underwent radical resection for rectal cancer in Shanxi Province Tumor Hospital from January 2016 to December 2019 were analyzed, including 129 males and 95 females, aged 28 to 83 (61.3±9.7) years old. The patients were randomly divided into two groups, namely, the training group (n=157) and the validation group (n=67), According to a ratio of 7∶3. ITK-SNAP image segmentation was used to manually delineate the ROI of tumor slice by slice on the images of T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and enhanced T1-weighted imaging (cT1WI) sequences to obtain the volume of interest. The delineation information of DWI was copied onto the apparent diffusion coefficient (ADC) map. A three-step dimensionality reduction method based on the maximum relevance minimum redundancy, least absolute shrinkage and selection operator regression, and multiple logistic regression was used for feature selection and radiomics signature building. Independent predictors of clinicopathologic features and MRI features were screened by multivariate logistic regression analysis. A radiomic model based on single and combined sequences of T2WI, ADC, and cT1WI and fusion models with clinicopathological features were constructed, and the corresponding nomogram was made. The area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve were used to evaluate the efficacy and clinical benefit of the model. Results Postoperative pathological examination confirmed LVI in 70 patients and negative in 154 patients. No significant differences in clinicopathologic features and MRI findings were observed between the training and validation groups (all P values > 0.05). Six key features related to LVI of rectal cancer were obtained after three-step screening (all P values < 0.05). Carcinoembryonic antigen (CEA) was an independent predictor of colorectal cancer (odds ratio[95% confidence interval] 2.071 [1.038~4.131], P = 0.039). The AUC of the radiomics model based on single and combined sequences of T2WI, ADC, and cT1WI were 0.765, 0.772, 0.776, and 0.878 in the training group and 0.741, 0.739, 0.764, and 0.846 in the validation group,respectively. The AUC of the fusion model training group and validation group constructed by CEA were 0.899 and 0.876, respectively, which showed the best prediction efficiency. The calibration curve showed that the fusion model had a good calibration performance. The decision curve of the verification group showed that the fusion model had the maximum net benefit when the threshold probability ranged from 0.10 to 0.20 and from 0.35 to 0.90. Conclusion The fusion model constructed based on the radiomic features of multi-parameter MRI and CEA has high diagnostic efficacy in predicting LVI of rectal cancer before surgery. The visual nomogram of this model can be used as an effective tool for predicting LVI before surgery.
Multiparameter magnetic resonance imaging-based radiomics model for the preoperative prediction of lymphovascular invasion in rectal cancer
Objective This study aims to investigate the application value of a fusion model based on multiparameter magnetic resonance imaging (MRI) for the preoperative prediction of lymphovascular invasion (LVI) in rectal cancer. Methods Retropective cohort study was conduted. The clinicopathological data and multi-parameter MRI data of 224 patients with rectal cancer who underwent radical resection for rectal cancer in Shanxi Province Tumor Hospital from January 2016 to December 2019 were analyzed, including 129 males and 95 females, aged 28 to 83 (61.3±9.7) years old. The patients were randomly divided into two groups, namely, the training group (n=157) and the validation group (n=67), According to a ratio of 7∶3. ITK-SNAP image segmentation was used to manually delineate the ROI of tumor slice by slice on the images of T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and enhanced T1-weighted imaging (cT1WI) sequences to obtain the volume of interest. The delineation information of DWI was copied onto the apparent diffusion coefficient (ADC) map. A three-step dimensionality reduction method based on the maximum relevance minimum redundancy, least absolute shrinkage and selection operator regression, and multiple logistic regression was used for feature selection and radiomics signature building. Independent predictors of clinicopathologic features and MRI features were screened by multivariate logistic regression analysis. A radiomic model based on single and combined sequences of T2WI, ADC, and cT1WI and fusion models with clinicopathological features were constructed, and the corresponding nomogram was made. The area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve were used to evaluate the efficacy and clinical benefit of the model. Results Postoperative pathological examination confirmed LVI in 70 patients and negative in 154 patients. No significant differences in clinicopathologic features and MRI findings were observed between the training and validation groups (all P values > 0.05). Six key features related to LVI of rectal cancer were obtained after three-step screening (all P values < 0.05). Carcinoembryonic antigen (CEA) was an independent predictor of colorectal cancer (odds ratio[95% confidence interval] 2.071 [1.038~4.131], P = 0.039). The AUC of the radiomics model based on single and combined sequences of T2WI, ADC, and cT1WI were 0.765, 0.772, 0.776, and 0.878 in the training group and 0.741, 0.739, 0.764, and 0.846 in the validation group,respectively. The AUC of the fusion model training group and validation group constructed by CEA were 0.899 and 0.876, respectively, which showed the best prediction efficiency. The calibration curve showed that the fusion model had a good calibration performance. The decision curve of the verification group showed that the fusion model had the maximum net benefit when the threshold probability ranged from 0.10 to 0.20 and from 0.35 to 0.90. Conclusion The fusion model constructed based on the radiomic features of multi-parameter MRI and CEA has high diagnostic efficacy in predicting LVI of rectal cancer before surgery. The visual nomogram of this model can be used as an effective tool for predicting LVI before surgery.
Rectal neoplasmsNeoplasm invasivenessForecastingLymphovascular invasionMagnetic resonance imagingRadiomicsJudging model
万方数据
