A joint model based on multi-sequence MRI imaging radiomics for prediction of benign and malignant ovarian tumors
张晓红 1王晋君 1武志峰 2全帅 3王翔 1王鹏飞 1王煊 1崔碧 1秦粽园 1赵丹 1郑昭
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 山西医科大学附属运城市中心医院影像科，运城 044000
2. 山西白求恩医院CT室，太原 030001
3. 通用电气医疗（上海）有限公司，上海 210000
折叠
摘要
目的 探讨基于多序列MRI影像组学构建联合模型术前预测卵巢肿瘤良恶性的价值。 方法 回顾性队列研究。纳入2019年1月—2023年5月山西医科大学附属运城市中心医院经术后病理确诊的卵巢肿瘤患者237例。患者年龄13~79（52.9±12.7）岁，其中卵巢良性肿瘤90例、恶性肿瘤147例，术前均行多序列盆腔MR扫描。以7∶3的比例将患者随机分为训练组166例和验证组71例。在T2加权像（T2WI）、T1WI、弥散加权成像（DWI）、表观弥散系数（ADC）图像中，对肿瘤进行逐层手动分割勾画感兴趣区，得到三维感兴趣区，并提取影像组学特征。以卵巢肿瘤的良恶性为研究标签，对训练组患者采用最小冗余最大相关（mRMR）算法进行影像组学特征去冗除杂，继而采用最小绝对收缩与选择算子回归为主的三步降维法筛选特征，并构建基于上述4个序列MRI的单一序列以及4个序列联合的多序列影像组学模型。多因素logistic回归用于筛选卵巢肿瘤良恶性的独立预测因子，结合多序列影像学模型，使用R语言建立联合模型并绘制列线图。采用受试者操作特征（ROC）曲线、校正曲线、决策分析曲线评估列线图的预测效能及临床效益。 结果 训练组和验证组良恶性肿瘤间比较，肿瘤的性状、侧向性特征差异均有统计学意义（P值均<0.05）。基于ADC、DWI、T1WI、T2WI序列的MRI上提取影像组学特征，经特征筛选后得到10个与卵巢肿瘤良恶性相关的关键特征（P值均<0.05）构建影像组学模型。训练组单一序列构建的影像组学模型预测卵巢肿瘤良恶性的ROC曲线下面积（AUC）分别为0.798、0.802、0.819、0.818，验证组AUC分别为0.792、0.798、0.803、0.806。4个序列联合的多序列影像组学模型AUC（训练组：0.839，验证组：0.833）大于4个单一序列的影像组学模型。单因素与多因素logistic回归分析显示，肿瘤性状（OR=0.421，95%CI：0.293~0.605）和肿瘤侧向性（OR=0.229，95%CI：0.104~0.503）是患者良恶性的独立预测因子（P值均<0.001）。多序列影像组学模型与肿瘤性状、侧向性特征构建的联合模型在训练组与验证组中的AUC分别为0.864、0.855，校正曲线显示出列线图有良好的校正性能，决策曲线表明当风险阈值概率范围在0.1~0.8时，采用联合模型预测卵巢肿瘤良恶性的净收益优于单一序列影像组学模型。 结论 基于多序列MRI影像组学特征与临床特征的联合模型对卵巢肿瘤良恶性具有良好的预测价值。 Objective This study aimed to explore the value of a joint model based on multisequence MRI imaging radiomics for the prediction of benign and malignant ovarian tumors. Methods This retrospective cohort study included 237 patients with ovarian tumors confirmed by postoperative pathology at the Yuncheng Central Hospital Affiliated to Shanxi Medical University from January 2019 to May 2023. All patients were female with ages ranging 13-79 (52.9±12.7) years. A total of 147 malignant and 90 benign ovarian tumors were included. All the patients underwent a multisequence pelvic MRI scan before their surgery and were randomly divided into the training group (n=166) and validation group (n=71) at a ratio of 7∶3. Volume regions of interest were manually delineated layer by layer in T2-weighted images (T2WI), T1-weighted images (T1WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps of each patient. Radiomic features were extracted from each patient. With benign and malignant patients as the research label, minimum redundancy maximum correlation was used to eliminate redundancy and clutter in the training set of patients. A three-step dimension reduction method based on the least absolute shrinkage and selection operator regression was used for feature selection and radiomics signature building. Multivariate logistic regression was applied to screen the independent predictive factor for benign and malignant ovarian tumors. R language was employed to establish the model and draw a column graph. The prediction performance was determined by calibration, discrimination, and clinical usefulness. Results Significant differences in the morphology and lateral characteristics of benign and malignant tumors were found between the training and validation groups (all P values <0.05). Ten key radiomic features associated with benign and malignant ovarian tumors were obtained after feature selection based on ADC, and DWI, T 2WI and T1WI (all P values <0.05) to build a radiomic model. The area under the receiver operating characteristic curve (AUC) values of the single sequence in the training group were 0.798, 0.802, 0.819, and 0.818, and those in the validation group were 0.792, 0.798, 0.803, and 0.806. The corresponding AUCs of the multisequence MRI radiomic model based on the combination of the four sequences in the training (0.839) and validation groups (0.833) were higher than those of the single radiomic model. Univariate and multivariate logistic analysis showed that tumor morphology and lateral characteristics were independent predictive factors of benign and malignant tumors. Multivariate logistic regression analysis showed that tumor characteristics ( OR=0.421, 95%CI:0.293-0.605) and tumor laterality (OR=0.229, 95%CI:0.104-0.503) were independent predictors of malignancy (all P values <0.001). The AUCs of the joint model based on multisequence MRI radiomic model, tumor morphology, and lateral characteristics were 0.864 and 0.855 in the training and validation groups, respectively. The calibration curves showed good calibration performance. When the risk threshold was 0.10-0.80, the net benefit of using a joint model to predict benign and malignant ovarian tumors was better than that of using a simple radiomic model. Conclusion The joint model based on multisequence MRI radiomics and clinical features has good predictive value for benign and malignant ovarian tumors.
Abstract
Objective This study aimed to explore the value of a joint model based on multisequence MRI imaging radiomics for the prediction of benign and malignant ovarian tumors. Methods This retrospective cohort study included 237 patients with ovarian tumors confirmed by postoperative pathology at the Yuncheng Central Hospital Affiliated to Shanxi Medical University from January 2019 to May 2023. All patients were female with ages ranging 13-79 (52.9±12.7) years. A total of 147 malignant and 90 benign ovarian tumors were included. All the patients underwent a multisequence pelvic MRI scan before their surgery and were randomly divided into the training group (n=166) and validation group (n=71) at a ratio of 7∶3. Volume regions of interest were manually delineated layer by layer in T2-weighted images (T2WI), T1-weighted images (T1WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps of each patient. Radiomic features were extracted from each patient. With benign and malignant patients as the research label, minimum redundancy maximum correlation was used to eliminate redundancy and clutter in the training set of patients. A three-step dimension reduction method based on the least absolute shrinkage and selection operator regression was used for feature selection and radiomics signature building. Multivariate logistic regression was applied to screen the independent predictive factor for benign and malignant ovarian tumors. R language was employed to establish the model and draw a column graph. The prediction performance was determined by calibration, discrimination, and clinical usefulness. Results Significant differences in the morphology and lateral characteristics of benign and malignant tumors were found between the training and validation groups (all P values <0.05). Ten key radiomic features associated with benign and malignant ovarian tumors were obtained after feature selection based on ADC, and DWI, T 2WI and T1WI (all P values <0.05) to build a radiomic model. The area under the receiver operating characteristic curve (AUC) values of the single sequence in the training group were 0.798, 0.802, 0.819, and 0.818, and those in the validation group were 0.792, 0.798, 0.803, and 0.806. The corresponding AUCs of the multisequence MRI radiomic model based on the combination of the four sequences in the training (0.839) and validation groups (0.833) were higher than those of the single radiomic model. Univariate and multivariate logistic analysis showed that tumor morphology and lateral characteristics were independent predictive factors of benign and malignant tumors. Multivariate logistic regression analysis showed that tumor characteristics ( OR=0.421, 95%CI:0.293-0.605) and tumor laterality (OR=0.229, 95%CI:0.104-0.503) were independent predictors of malignancy (all P values <0.001). The AUCs of the joint model based on multisequence MRI radiomic model, tumor morphology, and lateral characteristics were 0.864 and 0.855 in the training and validation groups, respectively. The calibration curves showed good calibration performance. When the risk threshold was 0.10-0.80, the net benefit of using a joint model to predict benign and malignant ovarian tumors was better than that of using a simple radiomic model. Conclusion The joint model based on multisequence MRI radiomics and clinical features has good predictive value for benign and malignant ovarian tumors.
关键词
卵巢肿瘤/磁共振成像/影像组学/预测模型
Key words
Ovarian tumors/Magnetic resonance imaging/Radiomics/Prediction model
万方数据