利用深度学习挖掘多模态磁共振成像特征构建鼻咽癌远处转移预测模型

Distant relapse is the main cause of treatment failure and death for nasopharyngeal Carcinoma (NPC). Currently, the American Joint Committee on Cancer TNM staging is the most commonly used risk-stratified tool, but it is criticized for inadequate predictive accuracy. The clinical TNM staging of NPC is mainly on the basis of nuclear magnetic resonance imaging (MRI). Previous studies have shown that the digital characteristic of MRI might be predictive of risk of distant metastasis for radically treated NPC. However, only scarce data reported on deep mining of the MRI data for individualized risk-stratification.Deep learning is a novel machine learning technique, of which the self-learning property is helpful for understanding the nature of MRI in both essential and abstract ways. In this study, we plan to apply the deep-learning algorithm to MRI of NPC patients to explore the association between the multi-parametric digital features of MRI and the presence of distant relapse, and to identify the digital features that are significantly predictive of risk of distant metastasis. Additionally, we intend to incorporate the identified digital features of MRI along with previously discovered metastasis-related biomarkers into a compound predictive model as well as an accompanying nomogram, which is clinically useful for individualized prediction of the 5-year distant metastasis probabilities. The present study will provide an intuitive and practical tool for optimization of the risk-stratification and treatment planning among NPC patients by improving the survival among high-risk patients and avoiding overtreatment in other selected individuals .

远处转移是鼻咽癌治疗失败和死亡的主要原因，当前主要依据TNM临床分期分层评估，尚不能准确预测。MR图像是现行TNM临床分期的主要依据，现有研究表明挖掘其内在特征预测鼻咽癌远处转移具有可行性。目前对MR图像进行深度挖掘并预测个体化转移风险的研究较少报道。深度学习是一种新型机器学习方法，其自学习特性能够自动挖掘MR图像特征，获取更抽象、更本质的理解。因此，本研究拟利用深度学习算法挖掘鼻咽癌多模态MR图像的内在特征，构建鼻咽癌病变区域的自动识别、分类模型，探索图像特征与远处转移之间的相关性，明确与远处转移密切相关的图像特征及其对远处转移的预测价值。在此基础上，综合多模态MR图像数字化特征和课题组前期发现的、与远处转移相关的分子标志物建立复合预测模型，并通过列线图转为便捷的评估工具。预期可为临床提供更准确的远处转移预测方法，指导临床针对高失败风险病例制定个体化治疗方案，提高患者生存率。

远处转移是鼻咽癌治疗失败和死亡的主要原因，当前主要依据TNM临床分期分层评估，尚不能准确预测。MR图像是现行TNM临床分期的主要依据，现有研究表明挖掘其内在特征预测鼻咽癌远处转移具有可行性。目前对MR图像进行深度挖掘并预测个体化转移风险的研究较少报道。深度学习是一种新型机器学习方法，其自学习特性能够自动挖掘MR图像特征，获取更抽象、更本质的理解。本研究利用深度学习方法挖掘鼻咽癌多模态MR图像的内在特征，构建鼻咽癌病变区域的自动识别、分类模型，探索图像特征与远处转移之间的相关性，明确与远处转移密切相关的图像特征及其对远处转移的预测价值。在此基础上，综合多模态MR图像数字化特征和与远处转移相关的分子标志物建立复合预测模型，并通过列线图转为便捷的评估工具。本研究成果能为临床提供更准确的远处转移预测方法，指导临床针对高风险病例制定个体化治疗方案，提高患者生存率。