摘要: |
【摘要】 目的 运用机器学习技术构建基于超声图像形态定量特征的乳腺癌风险在线预测模型。 方法 回顾性收集2019年1月至2020年10月于皖南医学院第一附属医院行超声检查的乳腺肿块患者1046例,按随机数字以7:3比例分为训练集732例,验证集314例。所有肿块均有明确的病理结果。对训练集和验证集肿块分别按病理结果分为良性组和恶性组,使用ImageJ图像分析软件提取肿块超声图像的形态学定量特征。应用单因素分析训练集的良性组和恶性组定量特征,并运用机器学习技术进行多因素分析,确定影响癌症风险的独立危险因素,构建列线图预测模型。最后,使用验证集数据对模型效能进行测试,建立ROC曲线和校准曲线分析评价模型的诊断效能,并设计开发在线应用程序。结果 训练集732例患者中,430例(58.7%)病理为恶性;验证集314例患者中,199例(63.4%)病理为恶性。单因素分析显示形态学定量指标Aspect ratio (AR)、Circularity(C)、Modified Feret Angle (MFA)、Solidity(S)以及患者年龄在良性组和恶性组差异有统计学意义(P<0.05)。多因素分析显示AR、Circularity和年龄是预测恶性的独立危险因素(P<0.05)。基于上述独立危险因素构建乳腺癌风险预测列线图模型,并以互联网在线应用程序的方式呈现,模型内部验证AUC为0.931,敏感性88.1%,特异性85.4%,外部验证AUC为0.901,敏感性84.2%,特异性85.8%;校准曲线显示模型校准度良好,预测风险与实际风险未出现明显偏离。 结论 使用机器学习技术构建的乳腺癌风险列线图预测模型具有良好的诊断效能,而以互联网在线应用程序的方式呈现模型使其更加具有可操作性和实用性,将有助于对患者进行个体化预测和治疗决策。 |
关键词: 超声检查,乳腺肿瘤 机器学习 临床预测模型 列线图 互联网医疗 |
DOI: |
投稿时间:2021-10-12修订日期:2021-11-09 |
基金项目: |
|
The establishment of an online nomogram for predicting breast cancer based on quantitative ultrasound morphological features |
houyin,zhangqingling |
() |
Abstract: |
【Abstract】 Objective We aim to use machine learning methods to construct an online nomogram for predicting breast cancer based on quantitative ultrasound morphological features. Methods A total of 1046 patients with breast lesions who underwent ultrasound examination were retrospectively collected in the study from January 2019 and October 2020 were divided into 732 patients in the training set, and 314 patients in the validation set according to a 7:3 ratio of random numbers. The lesions of the training set were classified into benign and malignant groups according to pathology results, then the ultrasound morphological features were quantified using the ImageJ software. All features were evaluated for the association with the cancer risk using univariate and multivariate analysis by the machine learning methods, then a nomogram was established for the determined risk factors and was presented in the form of an online application. The diagnostic performance of the nomogram was evaluated by receiver operating characteristic (ROC) curve analysis and calibration plot. Results Pathological examination revealed a malignant lesion in 430 patients (58.7 %) for the training set and 199 patients (63.4 %) for the validation set. Univariate analysis showed that the morphological quantitative features AR, Circularity, MFA, and Solidity were statistically different between the benign and malignant groups (P<0.05). Multivariate analysis showed that AR, Circularity, and age were independent risk factors for malignancy (P<0.05). Based on the independent risk factors, a nomogram model for predicting breast cancer was constructed, and an online application was released. The area under the curve(AUC) of the nomogram was 0.931,0.901, the sensitivity was 88.1%, 84.2%, and the specificity was 85.4%,85.8% in the internal and external validation, respectively. The calibration plot showed good agreement between predicted and actual risk. Conclusions The nomogram for predicting breast cancer based on quantitative ultrasound morphological features has good performance, and the presentation of the nomogram in the form of an internet online application makes it more practical, which may help tailor appropriate treatment decisions. |
Key words: Ultrasonography, Breast neoplasms Machine learning Clinical prediction models Nomogram Internet medicine |