Abstract:

Objective　To establish a model for predicting the expression of epidermal growth factor receptor (EGFR) in breast cancer by ultrasound signs/parameters combined with non-morphological (functional and hemodynamic) magnetic resonance imaging (MRI) parameters and evaluate the diagnostic efficiency of the model. Methods　A retrospective study was conducted on 167 female patients with breast cancer aged 21-79 (45.23±11.72) years who underwent surgical resection after neoadjuvant chemotherapy (NAC) in Lianyungang First People's Hospital from January 2015 to December 2021. Before NAC treatment, all patients underwent breast mass biopsy, and their histopathological and EGFR analysis results were obtained. All patients underwent ultrasound and MRI examination before puncture (the interval between the two examinations was not more than 5 days). The parameters recorded were EGFR, apparent diffusion coefficient (ADC) value, Max Slope, time to peak (TTP), signal enhancement rate (SER), early enhancement rate (EER), time-signal strength curve (TIC), tumor maximum diameter, tumor margin and boundary, aspect ratio, microcalcification, color Doppler flow imaging (CDFI) grading, resistance index (RI), axillary lymph node metastasis (ALNM). Statistical analysis was conducted in R language 4.0.3 (https://www.r-project.org). All data were randomly divided into a modeling group and a verification group by the ratio of 7:3, and the random seeds were set. The comparison between the modeling group and the verification group was conducted according to the above variable types by independent sample t test, Mann-Whitney U test, and χ² test, respectively, to analyze the differences in the ultrasound parameters/signs, MRI non-morphological parameters, and EGFR expression. The random forest (RF) model was used to select variables in the EGFR modeling group. Generalized linear modeling (GLM) was used to obtain regression coefficient, odds ratio (OR), and 95% confidence interval for each modeling variable in the modeling group, and a colored nomogram was drawn. The model discrimination was evaluated by receiver operating characteristic curve (ROC), and the model effectiveness was evaluated by decision curve method (DCA). Results　ADC, maximum diameter, EER, and TTP were selected as the variables to construct the model for predicting EGFR. The model discrimination evaluation: ROC analysis showed that the area under the curve (AUC) of the modeling group was 0.815 (95%CI 0.726-0.905); in the verification group, the AUC was 0.805 (95%CI 0.660-0.949); AUC >0.80 indicated high discrimination. Model effectiveness evaluation: DCA analysis showed that the diagnosis probability in the patients of the modeling group was in the range of 0%-50%, and that of the verification group was in the range of 15%-55%. Conclusions　Based on machine learning, DCA, and other methods, the prediction model of EGFR expression by ultrasound signs/parameters combined with MRI non-morphological parameters before breast cancer treatment is proved to have good diagnostic efficacy. It is expected to produce higher clinical benefit and better clinical application in improving preoperative diagnosis and curative effect evaluation of immune factors for breast cancer.

Key words:

Breast cancer, Prediction model, Ultrasound, MRI, EGFR

摘要：

目的　构建乳腺癌治疗前超声征象联合磁共振成像（MRI）非形态学（功能学及血流动力学）参数预测表皮生长因子受体（EGFR）表达模型并评价模型的诊断效能。方法　回顾性分析2015年1月至2021年12月连云港市第一人民医院经新辅助化疗（NAC）后手术治疗的乳腺癌患者167例，年龄21～79（45.23±11.72）岁。所有患者在NAC治疗前，均行乳腺肿块穿刺活检并取得组织病理学和EGFR分析结果，患者穿刺前均行超声及MRI检查（2项检查间隔时间不超过5 d），记录变量为：EGFR、表观扩散系数（ADC）值、最大倾斜率（Max Slope）、达峰时间（TTP）、信号增强率（SER）、早期强化率（EER）、时间-信号强度曲线（TIC）、肿瘤长径、肿瘤边缘及边界、纵横比、微钙化、彩色多普勒血流成像（CDFI）分级、阻力指数（RI）、腋窝淋巴结转移（ALNM）。统计学分析采用R语言4.0.3（https：//www.r-project.org）。全部数据按照7∶3的比例随机分为建模组和验证组（设置随机种子）。建模组和验证组之间比较按照上述变量类型分别采用独立样本t检验、Mann-Whitney U检验和χ²检验分析超声参数/征象及MRI非形态参数与EGFR表达的组间差异。EGFR建模组筛选变量方法采用随机森林模型（RF），在建模组中采用广义线性建模（GLM）得到每个建模变量的回归系数、比值比（OR）和95%置信区间，并绘制彩色列线图。模型区分度评价采用受试者工作特征曲线（ROC），模型有效性评价采用决策曲线法（DCA）。结果　ADC、长径、EER和TTP筛选为构建预测EGFR表达模型的变量。该模型区分度评价：ROC分析显示建模组曲线下面积（AUC）=0.815，95%CI 0.726～0.905；验证组AUC=0.805，95%CI 0.660～0.949，AUC>0.80区分度较高。模型有效性评价：DCA分析显示，建模组患者诊断概率在0%～50%范围内，验证组诊断概率在15%～55%范围内。结论　基于机器学习和DCA等方法构建的乳腺癌治疗前超声征象/参数联合MRI非形态学参数预测EGFR表达的模型，有较好的诊断效能，可获得更高临床效益，对提高乳腺癌免疫因子的术前诊断、疗效评估有较好的临床应用价值。

关键词:

乳腺癌, 预测模型, 超声, MRI, EGFR