多参数MRI影像基因组学对乳腺癌早期复发的预测研究

Multiparameter MRI imaging genomics in breast cancer predictive studies of early recurrence

目的:探讨磁共振动态对比增强和高b值弥散加权成像的影像基因组学预测乳腺癌早期复发的价值。方法:回顾性分析广西医科大学第一附属医院及桂林市中医医院在2017年1月—2019年4月间收治的共100例经手术或穿刺活检病理证实为乳腺癌的患者。患者临床及影像资料,包括发病年龄、绝经状态、免疫组化(包括ER、PR、HER-2、Ki-67等)、21基因检测结果,肿瘤大小,淋巴结转移,复发,肿瘤增强模式,肿瘤边缘(无毛刺或/分叶),肿瘤形态,肿瘤增强曲线。根据21基因检测结果将乳腺癌患者分为中高危组(RS≥18)和低危组(RS<18)。从PACS系统输出b值高的患者弥散加权成像(高b值DWI)和对比增强T_(1)加权(T_(1)+C)图像,由两位经验丰富的放射科医生手工勾画出感兴趣区(ROI)。采用汇医慧影软件从选定的ROI中提取动态对比增强(DCE)和高b值DWI图像中影像组学特征,运用不同方法对MRI的特征降维,运用KNN法建立影像组学标签与21基因检测结果的预测乳腺癌早期复发的模型,将患者按7:3比例划分为训练集(70例)和测试集(30例),比较两组患者的曲线下面积(AUC)、95%CI、灵敏度及特异度;随后将提取的影像组特征学与一般临床资料进行机器学习,将患者按8:2的比例划分训练集(80例)和测试集(20例),比较两组患者临床病理资料+影像基因组学组合模型的AUC、95%CI、灵敏度及特异度,最后比较两者的预测效能。结果:在100例女性乳腺癌患者中,包括32例浸润性导管癌伴导管原位癌,55例浸润性导管癌,导管原位癌3例,浸润性导管癌伴小叶癌2例,黏液腺癌3例,浸润性小叶癌3例,髓样癌1例,浸润性导管癌伴乳头状Patcher病1例,复发5例。采用最小绝对收缩算子(LASSO)特征选择方法降维,采用KNN算法建立影像基因组学模型和临床+影像基因组学特征组合模型;两种模型预测乳腺癌早期复发的AUC、灵敏度和特异度比较,中高危险组:训练集0.78、67%、73%vs 0.91、80%、86%,测试集:0.78、73%、67%vs 0.91、86%、80%;低危险组:训练集0.59、69%、53%vs 0.68、62%、75%,测试集:0.59、53%、69%vs0.68、75%、62%。结论:基于磁共振DWI及DCE图像中提取的影像特征、21基因检测及临床特征建立的临床+影像基因组学组合模型对乳腺癌早期复发具较好的预测价值。

Objective To investigate the value of dynamic contrast enhanced magnetic resonance and high b-value diffusionweighted imaging(DWI) in imaging genomics in predicting early recurrence of breast cancer.Methods This study retrospectively analyzed 100 patients with breast cancer pathologically confirmed by surgery or puncture biopsy admitted to the First Affiliated Hospital of Guangxi Medical University and Guilin Hospital of Traditional Chinese Medicine from January 2017 to April 2019.Clinical and imaging data of patients,including age of onset,menopausal status,immunohistochemistry(including ER,PR,HER-2,Ki-67,etc.),Oncotype DX 21-gene test results,tumor size,lymph node metastasis,recurrence,tumor enhancement pattern,tumor margin(no burrs or/lobed),tumor morphology,and tumor enhancement curve.According to the results of Oncotype DX 21-gene test,breast cancer patients were divided into medium-high risk group(RS≥18) and low-risk group(RS<18).Diffuse-weighted imaging(high b-value DWI) and contrast-enhanced T_(1)-weighted(T_(1)+C) images were exported from the PACS system,and areas of interest(ROI) were manually delineated by two experienced radiologists.Huiyi Huiying software was used to extract the image omics features of dynamic contrast enhancement(DCE) and high b-value DWI images from the selected ROI.Different methods were used to reduce the dimension of MRI features.KNN method was used to establish the model of predicting early breast cancer recurrence with image omics tags and Oncotype DX 21-gene test results.They were divided into the training set(70 cases) and the test set(30 cases) to compare the AUC,95%CI,sensitivity and specificity of the two groups.Then,the extracted imaging group features and general clinical data were used for machine learning.Patients were divided into the training set(80 cases) and the test set(20 cases) according to the ratio of 8:2.The AUC,95%CI,sensitivity and specificity of the combined clinicopathological data + imaging genomics model between the two groups were compared,and finally the predictive efficacy of the two models was compared.Results A total of 100 female breast cancer patients were collected in this study.They included 32 cases of invasive ductal carcinoma with ductal carcinoma in situ,55 cases of invasive ductal carcinoma in situ,3 cases of ductal carcinoma in situ,2 cases of invasive ductal carcinoma with lobular carcinoma,3 cases of mucinous adenocarcinoma,3 cases of invasive lobular carcinoma,1 case of medullary carcinoma,1 case of invasive ductal carcinoma with papillary Patcher's disease,and 5 cases of recurrence.The minimum absolute contraction operator(LASSO) feature selection method was used to reduce dimension,and KNN algorithm was used to establish imaging genomics model and clinical + imaging genomics feature combination model.Comparison of the AUC,sensitivity and specificity of the two models for predicting the early recurrence of breast cancer:in the medium-high risk group,the training set was 0.78,67%,73% vs 0.91,80%,86%;in the test set,0.78,73%,67% vs 0.91,86%,80%;Low risk group:training set 0.59,69%,53% vs 0.68,62%,75%;test set:0.59,53%,69% vs 0.68,75%,62%.Conclusion The combined clinical and imaging genomics model based on the image features,Oncotype DX 21-gene test and clinical features extracted from magnetic resonance DWI and DCE images has good predictive value for the early recurrence of breast cancer.