基于增强CT影像组学及形态学征象对非小细胞肺癌脏层胸膜侵犯的预测价值研究

Study on the value of enhanced CT-based radiomics and morphological features in predicting visceral pleural invasion in non-small cell lung cancer

目的探讨基于增强CT影像组学方法及形态学征象在术前预测非小细胞肺癌患者脏层胸膜侵犯(VPI)的效能。方法回顾性研究。纳入2019年1月—2021年1月蚌埠医学院第一附属医院收治的肺癌患者220例,其中男145例、女75例,年龄43~89(62.6±10.0)岁,均行根治性手术治疗。按照术后病理检查确诊有VPI 90例、无VPI 130例。将入组患者按照4∶1的比例随机分配到训练组(176例)与验证组(44例)。基于术前增强CT图像提取影像组学特征,采用LASSO-logistic回归模型选择动静脉期相关性最高的影像组学特征建立VPI预测模型。利用独立样本t检验和χ^(2)检验筛选临床资料及CT形态学征象等相关变量,结合最终选择的相关性最高的影像组学特征构建联合模型;绘制受试者工作特征(ROC)曲线,采用曲线下面积(AUC)评价模型在训练组和验证组中对VPI的预测效能,DeLong检验用于比较模型间AUC的差异。结果从提取出的1878个影像组学特征中筛选出动脉期及静脉期各10个最具相关性的影像组学特征,分别用于建立动脉期和静脉期影像组学VPI预测模型。在训练组和验证组中,静脉期影像组学模型AUC值分别为0.867(95%CI 0.815~0.920)和0.855(95%CI 0.746~0.964),均大于动脉期的0.844(95%CI 0.784~0.904)和0.814(95%CI 0.677~0.951),差异均有统计学意义(Z=2.20、2.07,P值均<0.05)。有、无VPI的患者在空洞征、毛刺征与胸膜凹陷征3种CT形态学特征的差异均有统计学意义(χ^(2)=8.30、7.87、10.32,P值均<0.05)。训练组与验证组患者基线资料比较,差异无统计学意义(P值均>0.05)。联合模型由最终选择的10个相关性最高的静脉期影像组学特征及上述3种CT形态学征象共同构建,其在训练组和验证组中AUC分别为0.914(95%CI 0.875~0.953)和0.884(95%CI 0.785~0.984),均大于静脉期影像组学模型,差异均有统计学意义(Z=3.09、2.21,P值均<0.05),即联合模型对VPI的预测效能更高。结论基于增强CT静脉期图像的影像组学特征联合空洞征、毛刺征与胸膜凹陷征3种CT形态学征象构建的联合模型,对于术前非小细胞肺癌患者是否发生VPI的判定具有很好的预测效能,可以协助临床决策。

Objective This study aimed to explore the efficacy of preoperative prediction of visceral pleural invasion(VPI)in patients with non-small cell lung cancer(NSCLC)based on enhanced CT radiomics and morphological features before operation.Methods A total of 220 patients with lung cancer treated in the First Affiliated Hospital of Bengbu Medical College from January 2019 to January 2021 were analyzed retrospectively,including 145 males and 75 females aged 43-89(62.6±10.0)years old.According to the postoperative pathological examination,90 cases were diagnosed with VPI,and 130 cases did not have VPI.According to the proportion of 4:1,the patients were randomly divided into training group(176 cases)and verification group(44 cases).The radiomics features were extracted based on the preoperative enhanced CT images.The LASSO-Logistic regression model was used to select the radiomics features with the highest correlation between arteriovenous phase and venous phase to establish the VPI prediction model.Independent sample t-test andχ^(2) test were used to screen clinical data,CT morphological features,and other related variables,which were combined with the final selection of the most relevant radiomics features to build a joint model.The working characteristic(ROC)curve of the subjects was drawn,and the area under the curve(AUC)was used to evaluate the prediction efficiency of the model for VPI in the training and verification groups.The Delong test was used to compare the AUC differences between models.Results Among the 1878 radiomics features extracted,the 10 most relevant radiomics features in the arterial and venous phases were selected and used to establish VPI prediction models.In the training and validation groups,the AUC values of the venous imaging group model were 0.867(95%CI 0.815-0.920)and 0.855(95%CI 0.746-0.964),respectively,which were greater than those of 0.844(95%CI 0.784-0.904)and 0.814(95%CI 0.677-0.951)in the arterial phase,and the differences between the groups were statistically significant(Z=2.20,2.07,all P values<0.05).Significant differences were found in three CT morphological features,namely,cavity sign,spiculation sign,and pleural indentation sign,between patients with VPI and those without VPI(χ^(2)=8.30,7.87,10.32,all P values<0.05).There was no significant difference in baseline data between the training group and the validation group(all P values>0.05).The combined model was constructed by using the final 10 highest venous phase radiomics features that were correlated and the above three CT morphological signs.The AUCs in the training and validation groups were 0.914(95%CI 0.875-0.953)and 0.884(95%CI 0.785-0.984),respectively,which were greater than those in the venous phase imaging model,and the AUC difference between the two models in the training and validation groups was statistically significant(Z=3.09,2.21,all P values<0.05).The joint model had higher prediction efficiency for VPI.Conclusion Based on the radiomics features of enhanced CT venous phase images combined with cavity sign,spiculation sign and pleural indentation sign,the joint model of CT morphological signs can predict the occurrence of VPI in patients with non-small cell lung cancer before operation and assist clinical decision-making.