文章摘要
基于多参数MRI影像组学构建机器学习模型与直肠癌Ki-67表达相关性
Correlation between Ki-67 expression and a machine learning model based on multi-parametric MRI radiomics in rectal cancer
投稿时间:2023-11-21  
DOI:10.3969/j.issn.1000-0399.2024.06.004
中文关键词: Ki-67  直肠癌  多参数磁共振  影像组学  机器学习
英文关键词: Ki-67  Rectal cancer  Multi-parametric magnetic resonance imaging  Radiomics  Machine learning
基金项目:
作者单位E-mail
孙铭洁 241002 安徽芜湖 皖南医学院研究生院  
薄娟 233030 安徽蚌埠 蚌埠医科大学研究生院  
魏龙宇 230001 安徽合肥 安徽医科大学研究生院  
付宝月 230001 安徽合肥 安徽医科大学研究生院  
李雪萌 230001 安徽合肥 安徽医科大学研究生院  
董江宁 230031 安徽合肥 中国科学技术大学附属第一医院(安徽省立医院)西区影像科  
高飞 230031 安徽合肥 中国科学技术大学附属第一医院(安徽省立医院)西区影像科 15956912758@163.com 
摘要点击次数: 89
全文下载次数: 63
中文摘要:
      目的 探讨 mp-MRI 的影像组学模型术前无创性预测直肠癌患者 Ki-67 表达水平的价值。 方法 回顾性分析 2016年 1 月至 2023 年 7 月在安徽省肿瘤医院就诊的 97 例直肠腺癌患者临床资料,术前行常规 MRI 检查及 IVIM-DWI 扫描,根据术后病理Ki-67 表达状态,分为低表达组(n=64)、高表达组(n=33)。在三个不同序列(高分辨 Ax-T2WI、动态增强平衡期 Sag LAVA-FLEX 和IVIM-DWI)上分别提取三维影像组学纹理分析特征参数。按 7∶3 比例分为训练组(n=67)和测试组(n=30),训练组用于特征筛选和建立影像组学模型,测试组用于验证所建立模型的可靠性。比较训练组和测试组患者基线资料的差异,使用 Pearson 相关性、SelectPer-centile 和 LASSO 完成最佳预测特征选择,分别基于影像组学特征采用 SGD、SVM 和 LR 机器学习算法构建模型,并进行 10 折交叉验证。应用受试者操作特征(ROC)曲线评估模型预测直肠腺癌 Ki-67 表达水平的效能,采用 DeLong 检验对曲线下面积(AUC)进行比较。 结果 从每例患者的高分辨 Ax-T2WI、动态增强平衡期 Sag LAVA-FLE 和 IVIM-DWI (b=800 s/mm2)三个序列上共提取出 5 622 个影像组学特征;筛选出 6 个影像组学特征作为最佳预测特征构建模型。SGD、SVM 和 LR 的最高 AUC 分别为 0.867、0.853 和 0.884;其中 SGD算法准确率为 76%,在 3 个模型中预测性能最佳。 结论 基于多参数 MRI 的影像组学模型在术前预测直肠腺癌 Ki-67 表达水平有一定的价值,SGD 在预测直肠腺癌 Ki-67 表达水平中效能最佳。
英文摘要:
      Objective To investigate the value of radiomics model based on mp-MRI in predicting Ki-67 expression in rectal cancer. Methods The clinical data of 97 patients with rectal adenocarcinoma in Anhui Provincial Cancer Hospital from January 2016 to July 2023 were retrospectively analyzed. Preoperative routine MRI and IVIM-DWI were performed. According to the postoperative pathological Ki-67 expression status, the patients were divided into the low expression group (n=64) and the high expression group (n=33). Three different sequences (High resolution Ax-T2WI, dynamic enhanced balanced phase Sag LAVA-FLEX and IVIM-DWI) were used to extract 3D radiomics texture analysis feature parameters. According to the ratio of 7∶3, they were divided into a training group (n=67) and a test group (n=30). The training group was used for feature screening and radiomics model establishment, and the test group was used to verify the reliability of the established radiomics model. Pearson correlation, SelectPercentile and LASSO were used to complete the selection of the best predictive features. SGD, SVM and LR machine learning algorithms were used to construct the models based on the radiomics features, respectively, and 10-fold cross-validation was performed. The receiver operating characteristic (ROC) curve was plotted to evaluate the efficacy of the model in predicting Ki-67 expression level in rectal adenocarcinoma, and the DeLong test was used to compare the area under the curve (AUC). Result A total of 5 622 radiomics features were extracted from high-resolution Ax-T2WI, Sag LAVA-FLE and IVIM-DWI (b=800 s/mm2) for each patient. Six radiomics features were selected as the best predictive features to construct the model. The highest AUCs of SGD, SVM and LR were 0.867, 0.853 and 0.884, respectively. The accuracy of SGD algorithm was 76%, which had the best prediction performance among the three models. Conclusion The multi-parametric MRI-based radiomics model has certain value in preoperative prediction of Ki-67 expression level in rectal adenocarcinoma. SGD has the best performance in predicting the expression level of Ki-67 in rectal adenocarcinoma.
查看全文   查看/发表评论  下载PDF阅读器
关闭