基于CT图像深度学习模型在结直肠癌旁肿瘤沉积中的诊断价值  

Diagnostic value of deep learning model based on CT images for adjacent tumor deposits in colorectal cancer

作　　者：刘燕[1] 罗锦文[1] 刘艳丽[1] 唐亚霞[1] LIU Yan;LUO Jinwen;LIU Yanli;TANG Yaxia(Department of Medical Imaging,the Fifth Affiliated Hospital of Guangzhou Medical University,Guangdong Guangzhou 510700,China)

机构地区：[1]广州医科大学附属第五医院医学影像科,广东广州510700

出　　处：《现代肿瘤医学》2025年第4期620-627,共8页Journal of Modern Oncology

基　　金：广东省医学科学技术研究基金(编号:A2023485);广东省广州市教育局高校研究生科研项目(编号:2024312248)。

摘　　要：目的:探讨基于CT图像深度学习模型在术前预测结直肠癌(colorectal cancer,CRC)旁肿瘤沉积(tumor deposits,TDs)的价值。方法:回顾性分析经术后病理证实的300例CRC患者CT图像及临床资料,对横轴位静脉期增强的CT图像进行图像归一化及重采样,并采用ITK-SNAP软件在图像上对全肿瘤区域进行全瘤标注,根据病理结果分为TDs组和无TDs组,并按照7∶3比例随机分为训练集(210例)和验证集(90例)。对全瘤标注图像进行深度学习特征提取及影像组学特征提取,采用最小绝对收缩与选择算子(Least Absolute Shrinkage and Selection Operator,LASSO)回归进行影像组学特征筛选并分别建立影像组学模型、深度学习模型及影像组学+深度学习模型,使用受试者工作特征曲线下面积(area under the ROC curve,AUC)评估各模型的诊断效能,使用决策曲线分析(decision curve analysis,DCA)评估模型的临床价值。结果:影像组学模型、深度学习模型、影像组学+深度学习模型在训练集中的AUC值(95%CI)分别为0.810(0.776~0.872)、0.846(0.776~0.917)、0.868(0.812~0.924),在验证集中的AUC值(95%CI)分别为0.800(0.736~0.864)、0.826(0.761~0.891)、0.855(0.795~0.916),敏感性分别为73.22%、56.56%、67.67%,特异性分别为64.01%、85.93%、87.30%,Delong测试验证集中三组模型间均有统计学差异(P<0.05)。三组模型在训练集和验证集中均有良好的校准和区分能力,影像组学联合深度学习模型诊断效能最高,DCA曲线显示影像组学联合深度学习模型对预测TDs净获益最佳。结论:基于CT图像深度学习模型对术前预测CRC旁TDs具有很好的诊断价值,影像组学联合深度学习模型诊断效能最高,预测TDs净获益最佳,能有助于更好帮助临床医师治疗决策。Objective:To explore the value of deep learning model based on CT images in preoperative prediction of tumor deposits(TDs)adjacent to colorectal cancer(CRC).Methods:A retrospective analysis was conducted on CT images and clinical data of 300 CRC patients confirmed by postoperative pathology.The transverse venous phase enhanced CT images were normalized and resampled,and the ITK-SNAP software was used to annotate the entire tumor area on the images.The patients were divided into TDs group and non-TDs group based on pathological results,and randomly divided into a training set(210 cases)and a validation set(90 cases)in a 7∶3 ratio.Perform deep learning feature extraction and radiomics feature extraction on whole tumor annotated images,use the Least Absolute Shrinkage and Selection Operator(LASSO)regression to screen radiomics features,and establish radiomics model,deep learning model,and radiomics+deep learning model.Use area under the receiver operating characteristic curve(AUC)to evaluate the diagnostic performance of each model,and use decision curve analysis(DCA)to evaluate the clinical value of the model.Results:The AUC values(95%CI)of the radiomics model,deep learning model,and radiomics+deep learning model in the training set were 0.810(0.776~0.872),0.846(0.776~0.917),and 0.868(0.812~0.924),respectively.The AUC values(95%CI)in the validation set were 0.800(0.736~0.864),0.826(0.761~0.891),and 0.855(0.795~0.916),respectively.The sensitivities were 73.22%,56.56%,and 67.67%,and the specificities were 64.01%,85.93%,and 87.30%,respectively.There were statistically significant differences among the three models in the validation set by Delong test(P<0.05).The three models in the training and validation sets had good calibration and discrimination abilities.The radiomics+deep learning model had the highest diagnostic performance.The DCA showed that the radiomics+deep learning model had the best net benefit for predicting TDs.Conclusion:The deep learning model based on CT images has good diagnostic value for preope

关 键 词：计算机断层扫描 深度学习 结直肠癌 肿瘤沉积 

分 类 号：R735.3[医药卫生—肿瘤]