基于磁共振影像组学的列线图对宫颈癌淋巴血管间隙浸润的评估研究

Assessment research of MRI radiomics-based nomogram on LVSI of cervical cancer

目的:构建基于多序列磁共振成像(MRI)为基础的影像组学列线图模型,对宫颈癌是否存在淋巴血管间隙浸润(LVSI)进行准确预测。方法:选取在医院就诊并进行手术治疗的369例宫颈癌患者MRI的影像和病理资料,采用随机数表法按照0.7∶0.3比例将病例资料分为机器学习的训练集(258例)和测试集(111例)。应用ITK-SNAP软件勾勒并测量肿瘤体积,并对肿瘤宫旁侵犯情况进行评估,应用Logistic回归将肿瘤体积与宫旁侵犯情况联合建立模型(Model1)。从T2WI、表观扩散系数(ADC)图、动脉期、静脉期及延迟期图像中分别进行组学参数的提取,通过LASSO回归筛选与肿瘤LVSI最为相关的组学参数,计算患者的影像组学评分。将影像组学评分、肿瘤体积和宫旁浸润联合建立影像组学列线图模型(Model^(2)),并使用德龙检验与Model1进行比较。使用霍斯默-莱姆肖(HL)检验评估Model^(2)的拟合优度。结果:训练集和测试集中Model1受试者工作特征(ROC)曲线下面积(AUC)分别为0.864和0.889。Model^(2)在训练集和测试集中对宫颈癌LVSI的AUC分别为0.913和0.942,Model^(2)与Model1相比,差异均有统计学意义(HL检验,P<0.05),且Model^(2)具有良好的拟合优度。结论:多序列MRI为基础的列线图模型能够预测宫颈癌的LVSI。

Objective:To construct a nomogram model of radiomics based on multi sequence magnetic resonance imaging(MRI),so as to conduct accuracy prediction for whether there is lymph-vascular space invasion(LVSI)in cervical cancer.Methods:The MRI images and pathological data of 369 patients with cervical cancer who received surgical treatment in hospital were selected,and random number table method was adopted to divided the case data into machine learning training cohort(258 cases)and test cohort(111 cases)as a ratio of 0.7:0.3.ITK-SNAP software was used to outline and measure tumor volume,and to assess the parametrial involvement of tumor.Logistic regression was used to establish Model1 through combined tumor volume and parametrial involvement.Radiomics parameters were extracted from T2WI images,apparent diffusion coefficient(ADC)images,arterial phase images,venous phase images and delay phase images,respectively.Lasso regression was used to screen the radiomics parameters which were the most correlation with tumor LVSI so as to calculate the score of radiomics.The combination of radiomics score,tumor volume and parametrial involvement were used to construct nomogram model of radiomics(Model^(2)),which was compared with Model1 by using Delong test.Hosmer-Lemeshow(HL)test were used to assess the goodness of fit of Model^(2).Results:The area under curve(AUC)of receiver operating characteristics(ROC)of Model1 were 0.864 and 0.889 in training cohort and test cohort,respectively.AUCs of Model^(2) on LVSI of cervical cancer were 0.913 and 0.942in training cohort and test cohort,respectively.The difference of Model^(2) and Model1 was significant(HL test,P<0.05),and Model^(2) has favorable goodness of fit.Conclusion:Multi-sequence MRI-based nomogram model can predict LVSI of cervical cancer.