MRI特征辅助分类乳腺癌分子亚型的临床研究

The Study of MRI Features Assisted Classification of Molecular Subtypes of Breast Cancer

目的 应用磁共振(MRI)图像中提取的特征和机器学习方法来帮助区分乳腺癌分子亚型,以期为临床诊治提供参考。方法 回顾性分析我院于2021年10月-2023年10月间确诊的178例乳腺癌患者的临床资料,每个患者肿瘤的形状、MRI特征和基于直方图的特征是使用内部软件从增强前和三次增强后的MRI图像上提取的。同时收集临床和病理资料。基于机器学习模型识别重要的成像特征并建立预测IDC亚型的模型。采用留一法交叉验证(LOOCV)避免模型过度拟合,采用Kruskal-Wallis检验确定统计学意义。结果 LOOCV过程生成一个具有不同特征的模型,在排名前20位的特征中,有11项在IDC亚型之间存在显著差异(P<0.05)。综合前九种病理和影像特征,预测模型对IDC亚型的识别准确率为83,4%。病理和影像联合模型对各亚型的准确率分别为89.2%(ERPR1)、63.6%(ERPR-/HER21)和8 2.5%(TN)。当仅结合前9个成像特征时,预测模型在LOOCV上识别IDC亚型的总体准确率为71.2%。病理和影像联合模型对各亚型的准确率分别为69.9%(ERPR1)、62.9%(ERPR-/HER21)和81.0%(TN)。结论 我们开发了一个基于机器学习的预测模型,该模型使用从M RI提取的特征来区分具有显著预测能力的IDC亚型。

Objective To use the features extracted from magnetic resonance(MRI) images and machine learning methods to help distinguish the molecular subtypes of breast cancer.Methods the clinical data of 178 patients with breast cancer diagnosed in our hospital from October 2021 to October 2023 were retrospectively analyzed.The shape,MRI features and histogram based features of each patient's tumor were extracted from the MRI images before and after three times of enhancement using internal software.Simultaneously collect clinical and pathological data.Identify important imaging features based on machine learning models and establish models for predicting IDC subtypes.Using the left one method cross validation(LOOCV) to avoid overfitting of the model,Kruskal Wallis test was used to determine statistical significance.Results The LOOCV process generated a model with different features,with 11 out of the top 20 features showing significant differences between IDC subtypes(P<0.05).Combining the first nine pathological and imaging features,the prediction model has an accuracy of 83.4% in identifying IDC subtypes.The accuracy of the combined pathological and imaging models for each subtype was 89.2%(ERPR1),63.6%(ERPR-/HER21),and 82.5%(TN),respectively.When only the first nine imaging features are combined,the overall accuracy of the prediction model in identifying IDC subtypes on LOOCV is 71.2%.The accuracy of the combined pathological and imaging models for each subtype was 69.9%(ERPR1),62.9%(ERPR-/HER21),and 81.0%(TN),respectively.Conclusion We have developed a machine learning based prediction model that uses features extracted from MRI to distinguish IDC subtypes with significant predictive ability.