基于深度卷积神经网络MRI特征对ER阳性HER2阴性乳腺癌复发风险预测研究

Breast cancer is one of the major malignant tumors which threaten women’s lives and health. Different molecular subtypes of breast cancers have different prognosis and recurrence risks. Endocrinotherapy were needed in all breast cancer patients with ER positive, HER2 negative, while the chemotherapy decision making was based on the recurrence risks. Currently, the most commonly used method to evaluate recurrence risk is 21-genes test to detect the gene expression in the specimen. And a new in vivo, non invasive method could be found. The results based on our previous study showed that the imaging features correlated with the expression of 21-genes, and could be a new method for the evaluation of recurrence risks. The deep convolutional Neural Networks could excavate the high-latitude and high-level abstract features of imaging data and made the foundation of the specification of predictive recurrence risk model used MRI data. This project try to using deep convolutional neural networks to analyze the imaging features of ER positive, HER2 negative breast cancer, and also using the comparison of imaging-genes to demonstrate the imaging features of breast cancer heterogeneity, based on 21-genes test scores to evaluate the value of convolutional features, aiming at a new method of breast cancer recurrence risk in vivo and noninvasive prediction.

乳腺癌是威胁女性生命健康的主要恶性肿瘤之一。不同分子分型乳腺癌具有不同的预后及复发风险。ER阳性HER阴性乳腺癌患者术后均需行内分泌治疗，而辅助化疗决策制定需要通过复发风险评估决定。目前，最常用的复发风险评估方式为检测乳腺癌标本中21个相关基因表达，现阶段尚缺乏在体无创的评估方法。我们前期的研究表明影像学特征与21基因复发分数具有相关性，能为预测复发风险提供一种新的模式。基于深度学习的卷积神经网络能挖掘复杂影像数据高纬度高层次抽象特征，能更好的构建基于MRI数据的复发风险模型。本项目拟应用基于深度学习的卷积神经网络提取ER阳性HER2阴性乳腺癌患者MRI特征，通过影像-基因对照阐示乳腺癌肿瘤异质性的影像学特征，通过与21基因表达的相关性分析，探究卷积特征在乳腺癌复发风险预测中的价值，以期为在体、无创评估乳腺癌复发风险提供新的手段。

乳腺癌是威胁女性生命健康的主要恶性肿瘤之一。ER阳性HER阴性乳腺癌患者术后均需行内分泌治疗，而辅助化疗决策制定需要通过21基因复发风险评估决定。本项目希望发现影像学特征与21基因复发分数之间的内在关联，通过深度学习构建ER阳性HER2阴性乳腺癌复发风险预测模型。我们的研究通过对220例ER阳性HER2阴性乳腺癌的动态增强MRI图像进行三维病灶分割，然后进行特征提取（临床病理特征，形态学特征，纹理特征，小波变幻特征），共计788个特征，通过特征出现频率选取频率出现大于70%以上的特征，纳入随机森林分类器，进行分类预测，构建三个复发风险预测模型（临床病理预测模型；影像组学模型；临床病理影像联合模型），我们的研究显示三个模型中，影像组学模型能更好的预测低中高三分类的复发风险 (AUC:0.78)。影像组学模型及临床病理影像联合模型能很好的预测二分类的复发风险问题，后两个模型诊断效能都高于临床病理预测模型。我们此次研究初步探索了在体无创探索乳腺癌复发风险预测的可能性，为进一步ER阳性HER2阴性乳腺癌预后提供了一种新的方式。