基于T_(2)WI纹理分析鉴别宫颈鳞癌和腺癌的价值研究

Application of T_(2)-weighted texture analysis to differentiate cervical squamous cell carcinoma from adenocarcinoma

目的探讨T_(2)WI纹理分析在术前鉴别宫颈鳞癌和宫颈腺癌中的价值。方法选取92例患者的术前T_(2)WI图像,从A·K软件中提取纹理参数共1117个,进一步进行纹理分析。通过Mann-Whitney u检验、单变量logistic回归分析和最小冗余最大相关性算法,选择纹理特征来区分宫颈鳞癌和宫颈腺癌。基于随机森林算法构建区分鳞癌和腺癌的预测模型,并通过受试者工作特征(ROC)曲线分析,以评价模型的诊断性能。最后,通过10次留组交叉验证法(LGOCV)评估预测模型的稳健性和可重复性。结果本文中在特征选择后,最终保留了5个鉴别宫颈鳞癌和宫颈腺癌的纹理特征用于构建RF模型。预测模型在区分宫颈鳞癌和宫颈腺癌方面具有良好的分类性能:准确率为81.4%、敏感性为81.33%,特异性为81.8%。此外,使用10次LGOCV算法证明了预测模型的稳健性和可重现性(平均AUC,0.83)。结论磁共振T_(2)WI纹理分析在鉴别宫颈鳞癌和宫颈腺癌方面有一定的鉴别诊断价值。

objective To investigate the value of T_(2)WI texture analysis in differentiating cervical squamous cell carcinoma from cervical adenocarcinoma before surgery.Methods The preoperative T_(2)WI images of 92 patients(78 cases of cervical squamous cell carcinoma and 14 of adenocarcinoma)were analyzed retrospectively.A total of 1133 texture parameters were ex⁃tracted from A·K software for further texture analysis.Textural features were selected to distinguish cervical squamous cell car⁃cinoma from cervical adenocarcinoma by Mann-Whitney u test,univariate logistic regression analysis,and least redundant maxi⁃mum correlation algorithm.In addition,a predictive model for distinguishing squamous cell carcinoma from adenocarcinoma was constructed based on a random forest algorithm and analyzed by receiver operating characteristic(ROC)curves to evaluate the diagnostic performance of the model.Finally,the robustness and reproducibility of the prediction model were assessed by the 10-leave-group cross-validation(LGOCV)method.Results In this study,after feature selection,five texture features differentiat⁃ing cervical squamous cell carcinoma from cervical adenocarcinoma were finally retained for constructing the RF model.The pre⁃dictive model had good classification performance in distinguishing cervical squamous cell carcinoma from cervical adenocarci⁃noma,with accuracy of 81.4%,sensitivity of 81.33%,and specificity of 81.8%.In addition,robustness and reproducibility of the prediction model was demonstrated using a 10-fold LGOCV algorithm(mean AUC,0.83).Conclusion MR T_(2)WI texture analysis has certain value in differentiating cervical squamous cell carcinoma from cervical adenocarcinoma.