【摘要】 目的 观察静脉期CT影像组学诊断结直肠癌（CRC）伴肿瘤沉积（TD）和/或淋巴结转移（LNM）的价值。方法 回顾性纳入263例CRC患者;根据病理结果分为TD(+)LNM（+）组（A组，n=103）、TD(+)LNM（-）组（B组，n=63）及TD(-)LNM（+）组（C组，n=97）并按8∶2比例随机归入训练集（n=210）及测试集（n=53）。基于静脉期CT筛选CRC影像组学特征，分别基于支持向量机（SVM）、多层感知器（MLP）及K邻近法（KNN）构建影像组学模型，分析其诊断效能。结果 训练集及测试集中，总体均以SVM模型诊断CRC伴TD和/或LNM的效能最高。因3组样本量分布不均，遂选取对处理类别不平衡问题更为稳健的微平均曲线下面积（Micro AUC）进行DeLong检验;结果显示，训练集中SVM模型与MLP及KNN模型差异均有统计学意义（P均<0.05），测试集中两两比较差异均有统计学意义（P均<0.05）。结论 静脉期CT影像组学模型、尤其SVM模型有助于诊断CRC伴TD和/或LNM。更多还原

【Abstract】 Objective To observe the value of venous phase CT radiomics for diagnosing colorectal cancer(CRC) complicated with tumor deposit(TD) and/or lymph node metastasis(LNM). Methods Totally 263 CRC patients were retrospectively enrolled and divided into TD(+)LNM(+) group(group A,n=103), TD(+)LNM(-) group(group B,n=63) and TD(-)LNM(+) group(group C,n=97) based on pathological results, also randomly divided into training set(n=210) and test set(n=53) at the ratio of 8∶2. Radiomics features of CRC were screened based on venous phase CT images. Radiomics models were constructed with support vector machine(SVM), multilayer perceptron(MLP) and K-nearest neighbor(KNN), respectively, and their diagnostic efficacy were analyzed. Results In both training set and test set, SVM model had the highest overall efficacy for diagnosing CRC complicated with TD and/or LNM. Due to the uneven distribution of sample sizes among 3 groups, DeLong test was performed on the micro-averaged area under the curve(Micro AUC), which was more robust for handling class imbalance problems. In training set, significant differences were found between SVM and MLP models, as well as between SVM and KNN models(both P<0. 05), while in test set, significant differences were observed between each 2 models(all P<0. 05). Conclusion Venous phase CT radiomics models, especially SVM model was helpful for diagnosing CRC complicated with TD and/or LNM.更多还原