基于18F-FDG PET/CT图像建立的多元影像组学模型对乳腺癌原发灶HER-2表达状态的预测价值

目的 评估基于18F-氟脱氧葡萄糖（FDG） PET/CT图像建立的多元影像组学模型对乳腺癌原发灶人表皮生长因子受体2（HER-2）表达状态的预测价值。 方法 回顾性分析2010年1月1日至2019年12月31日于天津医科大学肿瘤医院行18F-FDG PET/CT检查的273例女性乳腺癌患者的临床和影像学资料，年龄26~78（51.8±10.8）岁。根据乳腺癌原发灶HER-2表达状态的不同将患者分为HER-2阳性组和HER-2阴性组，比较2组患者的临床特征和PET/CT代谢参数的差异。勾画原发灶的感兴趣区，提取所有影像组学特征并建立基于PET/CT图像的影像组学特征模型，将样本中70%的患者作为训练集，剩余30%的患者作为测试集，采用受试者工作特征曲线比较PET/CT代谢参数及影像组学模型对乳腺癌原发灶HER-2表达状态的预测能力，采用十折交叉验证计算预测模型的平均性能。采用Wilcoxon秩和检验比较组间各PET代谢参数是否存在差异。计数资料的比较采用χ2检验，计量资料的比较采用两独立样本t检验和Mann-Whitney U秩和检验。 结果 HER-2阳性组患者106例、阴性组患者167例。HER-2阴性组患者合并腋下淋巴结转移的比例较HER-2阳性组患者高85.03%（80/106）对75.47%（142/167）]，且差异有统计学意义（χ2=3.900，P<0.05）。除腋下淋巴结转移情况外，2组患者的年龄、病理学类型及肿瘤分期间的差异均无统计学意义（t=?0.028，χ2=5.429、1.891，均P>0.05）。2组患者的PET代谢参数最大标准化摄取值、平均标准化摄取值、标准化摄取值峰值、肿瘤代谢体积、病灶糖酵解总量之间的差异均无统计学意义（Z=?1.583~?0.064，均P>0.05）。最终筛选出具有较好预测价值的37个影像组学特征建立组学特征模型，其中，PET影像组学参数8个、CT影像组学参数29个。在训练集中，影像组学特征模型曲线下面积（AUC）、准确率、灵敏度和特异度分别为0.913（95% CI:0.871~0.954）、0.882（95%CI:0.832~0.922）、0.849（95%CI:0.759~0.910）和0.910（95%CI:0.841~0.952）；在测试集中，影像组学特征模型AUC、准确率、灵敏度和特异度分别为0.820（95%CI:0.723~0.918）、0.830（95%CI:0.738~0.900）、0.875（95%CI:0.701~0.959）和0.807（95%CI:0.683~0.892）；经十折交叉验证后，影像组学模型的AUC、准确率、灵敏度、特异度的均值分别为0.818、0.847、0.908、0.764。 结论 相较于传统的PET代谢参数，基于18F-FDG PET/CT图像建立的多元影像组学模型对乳腺癌原发灶HER-2表达状态有较好的预测价值，有助于临床医师筛选曲妥珠单抗受试人群，改善患者预后。

The predictive value of 18F-FDG PET/CT derived multivariate radiomic mdodel in HER-2 status for primary breast cancer

Objective To evaluate the predictive value of 18F-FDG PET/CT derived multivariate radiomic model in human epidermal growth factor 2 (HER-2) status for primary breast cancer (BC). Methods A total of 273 BC patients aged 26?78(51.8±10.8) years with complete clinical data and imaging data who underwent 18F-FDG PET/CT imaging before any treatment from January 1, 2010, to December 31, 2019, were included in the retrospective study. According to HER-2 status in primary BC lesion, the BC patients were classified into HER-2 positive group and HER-2 negative group. The differences in clinical characteristics and PET/CT metabolic parameters between the two groups were compared. For radiomic analysis, a multivariate radiomic model based on PET/CT was established after lesion segmentation and radiomic feature extraction. Furthermore, all the candidates were randomly divided into the training set and testing set at a ratio of 7∶3. Receiver operator characteristic curve analysis was used to determine the predictive power of PET metabolic parameters and develop a radiomic model in HER-2 status. Furthermore, the average performance of the radiomic model in the prediction of HER-2 status was determined after tenfold cross-validation. The Wilcoxon rank sum test was performed to compare the differences in PET metabolic parameters between the two groups. Chi-square test was used for qualitative data, whereas two independent sample t test was used for quantitative data with normal distribution. Mann-Whitney U rank sum test was employed for quantitative data that did not obey normal distribution. Results A total of 106 patients were classified in HER-2 positive group, and 167 patients were in the negative group. The proportion of patients with axillary lymph node metastasis in the HER-2 negative group was higher than that in the HER-2 positive group (85.03%(80/106) vs. 75.47%(142/167)), and the difference was statistically significant (χ2=3.900, P<0.05). By contrast, no significant difference was found in age, pathological type, and tumor stage between the two groups (t=?0.028, χ2=5.429, 1.891; all P>0.05). For the five PET metabolic parameters between the two groups, namely, maximum standard uptake value, mean standard uptake value, peak of standard uptake value, metabolic tumor volume, and total lesion glycolysis, no statistically significant difference was found in the study (Z=?1.583 to ?0.064, all P>0.05). In the training set, the area under the curve (AUC), accuracy, sensitivity, and specificity of the radiomic model were 0.913(95%CI: 0.871–0.954), 0.882(95%CI: 0.832–0.922), 0.849(95%CI: 0.759–0.910), and 0.910(95%CI: 0.841–0.952), respectively. In the testing set, the AUC, accuracy, sensitivity, and specificity of the radiomic model were 0.820(95%CI: 0.723–0.918), 0.830(95%CI: 0.738–0.900), 0.875(95%CI: 0.701–0.959), and 0.807(95%CI: 0.683–0.892), respectively. After tenfold cross-validation, the average AUC, accuracy, sensitivity, and specificity of the imaging omics model were 0.818, 0.847, 0.908, and 0.764, respectively. Conclusions The established multivariate radiomic model based on 18F-FDG PET/CT images outperformed the traditional PET metabolic parameters in the prediction of HER-2 status for primary BC. This model can contribute to the clinical screening of a potential sensitive population for trastuzumab monoclonal antibody treatment and finally improve the prognosis for BC.