MRI纹理特征评估乳腺癌脉管浸润及其对病理信息缺失的填充

MRI texture feature on breast cancer vascular invasion assessment and missing histological information imputation

目的 探讨MRI纹理特征评估浸润性乳腺癌脉管浸润的价值,及其对缺失的乳腺癌分子分型和细胞角蛋白5/6基因表达信息的填充。方法 将114例接受MRI检查并经病理确诊为乳腺癌的患者分为训练组(n=51)、内部验证组(n=30)、外部验证组(n=33)。根据临床病理特征与MRI征象观察脉管浸润独立危险因素,构建MRI征象模型和纹理特征模型,评估2种模型的诊断效能估。采取交叉验证支持向量机递归特征消除法进行纹理特征选择,结合乳腺癌患者病理信息,建立非负矩阵分解(NMF)模型与协同过滤模型,采取受试者工作特征曲线进行模型填充性能评价。结果 3组患者瘤周胸肌前水肿、淋巴结转移状态差异具有统计学意义(P<0.05),是影响脉管浸润的独立危险因素。纹理特征模型诊断乳腺癌脉管浸润的曲线下面积高于MRI征象模型(P<0.05)。在缺失率为20%~40%时,NMF法的曲线下面积高于协同过滤法(P<0.05);在特征数为140时,NMF模型的曲线下面积高于协同过滤模型(P<0.05)。使用影像组学特征NMF模型的曲线下面积高于不使用影像组学特征NMF模型(P<0.05)。结论 MRI纹理特征可在术前有效预测乳腺癌脉管浸润,填充缺失的分子分型以及细胞角蛋白5/6基因表达信息。

Objective To explore the value of MRI texture features in evaluating the vascular infiltration of invasive breast cancer,as well as the missing molecular typing and cytokeratin 5/6 gene expression information imputation. Methods A total of 114 patients with breast cancer diagnosed by MRI and pathology were divided into training group(n=51), internal validation group(n=30) and external validation group(n=33). According to the clinicopathological features and MRI texture features, the independent risk factors of vascular infiltration were observed, the MRI sign model and texture feature model were constructed and their diagnostic efficacy was compared. Non-negative matrix decomposition(NMF) and collaborative filtering imputation models were established with the texture features selected by cross-validation support vector machine recursive feature elimination, and pathological information of breast cancer patients. The receiver operating characteristic curve was used to evaluate their imputation performance. Results There were statistically significant differences in peritumor prethoracic edema and lymph node metastasis status, which were independent risk factors for vascular invasion(P<0.05). The area under the curve(AUC) of the texture feature model in diagnosis of breast cancer vascular invasion was higher than that of MRI model(P<0.05). When the loss rate was 20%-40%, the AUC of NMF was higher than that of collaborative filtering(P<0.05). When the feature number was 140, the AUC of NMF was higher than that of collaborative filtering(P<0.05). The AUC of NMF with image omics was higher than that of NMF without image omics(P<0.05). Conclusion MRI texture features can effectively predict breast cancer vascular invasion before surgery, and fill in missing molecular typing and cytokeratin 5/6 gene expression information.