DCE-MRI纹理分析对乳腺癌分子分型的诊断价值

Diagnostic value of DCE-MRI texture analysis for molecular typing of breast cancer

目的探讨基于动态对比增强MRI(dynamic contrast-enhanced MRI,DCE-MRI)图像的纹理特征术前预测乳腺癌分子分型的价值。材料与方法回顾性分析宜昌市第一人民医院2021年10月至2022年10月75例经术后病理证实的乳腺癌患者的术前MRI图像及临床病理资料。采用χ2检验、方差分析对患者一般资料进行分析。对分子亚型以是与非作为二分类指标在DCE-MRI图像上提取特征参数,通过标准化、最优特征筛选器进行特征参数降维,采用独立样本t检验或Mann-Whitney U检验识别不同组间差异有统计学意义的最优纹理参数,采用ROC曲线下面积(area under the curve,AUC)评价纹理分析的诊断效能。另基于DCE-MRI纹理特征构建逻辑回归分类模型,绘制ROC曲线并评价模型对不同分子亚型的诊断效能。结果Luminal A型11例、Luminal B型36例、人表皮生长因子受体2(human epidermal growth factor receptor 2,HER-2)过表达型14例及三阴性乳腺癌(triple negative breast cancer,TNBC)14例,各亚型乳腺癌患者间年龄、绝经状态、病理学分型、MRI强化情况、淋巴结状态的差异皆无统计学意义(P>0.05)。基于MRI图像特征参数所建立的预测Luminal A型、Luminal B型、HER-2过表达型、TNBC的AUC[95%置信区间(confidence interval,CI)]值分别为0.92(0.77~1.00)、0.83(0.62~1.00)、0.83(0.55~1.00)、0.72(0.43~1.00)。Luminal A型与非Luminal A型组间3个纹理参数差异有统计学意义(P<0.05),三者AUC值分别为0.73、0.70和0.75,以纹理特征三维灰度共生矩阵-聚类阴影(3D grey level co-occurrence matrix cluster shadow,3D_glcm_CS)>0.439时诊断Luminal A型效能最高。Luminal B型与非Luminal B型组间2个纹理特征差异具有统计学意义(P<0.05),当原始灰度共生矩阵-聚类阴影(original gray level co-occurrence matrix cluster shadow,o_glcm_CS)>0.169时诊断Luminal B型效能最佳。HER-2过表达型与非HER-2过表达型组间5个纹理特征差异均有统计学意义,其AUC值分别为0.76、0.81、0.79、0.80和0.82,以三维灰度区域大小矩阵-小区域低灰度优势(3D grey level size zone matrix small area low gray level emphasis,3D_glszm_SALGLE)≤-0.460时诊断HER-2过表达型的效能最高(AUC=0.82,P<0.001);TNBC与非TNBC组间仅纹理特征高通滤波器-相邻灰度色差矩阵(wavelet LH neighbouring gray tone difference matrix busyness,w-LH_ngtdm_B)的差异有统计学意义,其AUC值为0.65。结论动态对比增强MRI纹理分析可以无创有效地预测乳腺癌分子分型,对术前乳腺癌分子亚型的分类具有重要的指导价值。

Objective:To explore the value of texture features based on dynamic contrast-enhanced MRI(DCE-MRI)images in preoperative prediction of molecular typing of breast cancer.Materials and Methods:The preoperative MRI images and clinicopathological data of 75 patients with breast cancer confirmed by postoperative pathology in the First People's Hospital of Yichang from October 2021 to October 2022 were retrospectively analyzed.The general data of patients were analyzed by chi-square test and variance analysis.Feature parameters were extracted from DCE-MRI images for molecular subtypes with yes and no as binary classification indicators.Dimension reduction of feature parameters was performed by standardized and optimal feature filters.Independent sample t-test or Mann-Whitney U test was used to identify the optimal texture parameters with statistically significant differences between different groups.The area under the ROC curve(AUC)was used to evaluate the diagnostic efficacy of texture analysis.In addition,a logistic regression classification model was constructed based on dynamic enhanced MRI texture features,and the ROC curve was drawn to evaluate the diagnostic efficacy of the model for different molecular subtypes.Results:There were 11 cases of Luminal A type,36 cases of Luminal B type,14 cases of human epidermal growth factor receptor 2(HER-2)overexpression type and 14 cases of triple negative breast cancer(TNBC).There was no significant difference in age,menopausal status,pathological classification,MRI enhancement and lymph node status among patients with different subtypes of breast cancer(P>0.05).The AUC[95%confidence interval(CI)]values of Luminal A,Luminal B,HER-2 overexpression and TNBC were 0.92(0.77-1.00),0.83(0.62-1.00),0.83(0.55-1.00)and 0.72(0.43-1.00),respectively.There were statistically significant differences in the three texture parameters between Luminal A and non-Luminal A groups(P<0.05).The AUC values of the three were 0.73,0.70 and 0.75,respectively.When the texture feature 3D grey level co-occurrence matrix cluster shadow(3D_glcm_CS)>0.439,the diagnostic efficiency of Luminal A type was the highest.There were significant differences in the two texture features between Luminal B group and non-Luminal B group(P<0.05).When original gray level co-occurrence matrix cluster shadow(o_glcm_CS)>0.169,the diagnostic efficiency of Luminal B type was the best.There were statistically significant differences in the five texture features between the HER-2 overexpression group and the non-HER-2 overexpression group.The AUC values were 0.76,0.81,0.79,0.80 and 0.82,respectively.When 3D grey level size zone matrix small area low gray level emphasis(3D_glszm_SALGLE)≤-0.460,the diagnostic efficiency of HER-2 overexpression was the highest(AUC=0.82,P<0.001).Only the difference of texture feature wavelet LH neighbouring gray tone difference matrix busyness(w-LH_ngtdm_B)between TNBC and non-TNBC was statistically significant,and the AUC value was 0.65.Conclusions:DCE-MRI texture analysis can noninvasively and effectively predict the molecular subtypes of breast cancer,which has important guiding value for the classification of preoperative molecular subtypes of breast cancer.