【摘要】 目的 探讨基于乳腺X线及临床特征构建机器学习模型预测乳腺导管内癌(DCIS)分子亚型并评估其预测效能。方法回顾性分析经手术病理证实的239例DCIS患者资料。根据免疫组化结果分为激素受体(HR)阳性型(139例)、人表皮生长因子受体2(HER2)阳性型(91例)、三阴性型(9例)。采集10个临床特征及17个X线影像特征,筛选出有统计学差异的特征构建多项式朴素贝叶斯(MNB)学习模型。通过受试者工作特征(ROC)曲线分析曲线下面积(AUC)评价学习模型的预测效能,并对模型进行可解释性分析。结果 HR阳性型、HER2阳性型、三阴性型DCIS学习模型的AUC分别为0.786、0.821、0.725,敏感度分别为0.714、0.741、0.333,特异度分别为0.700、0.711、0.797,准确度分别为0.708、0.722、0.778。结论 基于乳腺X线及临床特征预测DCIS不同分子亚型的学习模型效能较好,HER2阳性型学习模型的预测效能最优。更多还原

【Abstract】 Objective To construct machine learning models based on mammography and clinical features to predict molecular subtypes of ductal carcinoma in situ(DCIS),and to evaluate its predictive efficacy.Methods A total of 239 patients with DCIS confirmed by surgical pathology were analyzed retrospectively.According to the immunohistochemical results,all patients were classified into three groups,including hormone receptor(HR) positive type group(n=139),human epidermal growth factor receptor 2(HER2)positive type group(n=91) and triple negative type group(n=9).A total of 10 clinical features and 17 X-ray image features were collected,and the statistically different features were screened to construct the multinomial naive Bayes(MNB) learning model.The area under the curve(AUC) of receiver operating characteristic(ROC) curve was used to evaluate the predictive efficacy of learning model and the interpretability of the model was analyzed.Results The AUC of learning model in predicting HR positive type group,HER2 positive type group,and triple negative type group were 0.786,0.821,0.725;the sensitivity were 0.714,0.741,0.333;the specificity were 0.700,0.711,0.797;the accuracy were 0.708,0.722,0.778,respectively.Conclusion The learning model based on mammography and clinical features in predicting different molecular subtypes of DCIS have good performance,especially for the learning model of HER2 positive type.更多还原