基于多组学大数据的鼻咽癌个体化临床智能决策算法与支持系统

Nasopharyngeal carcinoma (NPC) is endemic in southern China, which is really a serous threat to the health and lives of people in China. In primary hospitals, the low accuracy of diagnosis and staging usually leads to unreasonable treatment regimens; hence the survival rates of NPC are only 60%. In high-volume cancer centers, the survival rates have reached to more than 80%, but insufficient personalized treatment restricts the further improvement of patient outcomes. In response to the distinct demands of improving NPC diagnosis and treatment, we aim to carry out the following studies, which are based on the clinical/imaging/pathological data of NPC and artificial intelligence theories and methods. First, to develop reliable and stable imaging standardization and automatic lesion segmentation methods for multi-modal radiological images and digital pathological images, which should be of high robustness and accuracy. Second, to develop computer-aided diagnosis (CAD) models to achieve accurate qualitative diagnosis of NPC through MRI and digital pathology, and more subdevidided pathological classifications for NPC prognostication and risk prediction. CAD models can transfer the experiences of the high-volume cancer centers to the primary hospitals, in order to guide relatively precise stratified treatment of NPC. Then, to develop a dynamic risk prediction model of NPC by integrating clinical/imaging/pathological multi-omics data from multiple time points during treatment, which can dynamically evaluate the risk of tumor recurrence and metastasis of individual patients in real time. Finally, to build a clinical decision support system (CDSS) specific for NPC based on decision rules and the dynamic risk prediction model, so as to guide personalized precision treatment of NPC. Through this project, we will develop a CAD system and an intelligent CDSS for NPC, and promote them to primary hospitals and high-volume cancer centers repectively, aim to improve the overall survival rate of NPC patients.

鼻咽癌高发于中国华南地区，严重威胁我国人民生命健康。基层医院诊断和分期准确性低，造成治疗方案选择不合理，鼻咽癌生存率仅60%；大型肿瘤中心鼻咽癌生存率达80%，但个体化精准治疗开展不充分制约了生存率的进一步提升。针对不同诊疗提升需求，拟基于鼻咽癌临床/影像/病理多组学数据和人工智能理论与方法开展：①针对多模态影像和数字病理图像，研发可靠、稳定的图像标准化方法和高鲁棒性、高准确率的病灶自动分割方法；②建立计算机辅助鼻咽癌定性及分型诊断模型，将大型诊疗中心的经验传递给基层医院，指导开展相对精准的分层治疗；③融合多时间点临床/影像/病理多组学数据建立鼻咽癌动态风险预测模型，实时动态地预测个体患者复发转移风险；基于决策规则和风险预测模型构建鼻咽癌单病种的临床决策支持系统，以指导个体化精准治疗。预期研制鼻咽癌定性及分型诊断系统与个体化智能决策平台各1套并分别推广示范，以期提高鼻咽癌患者整体生存率。

鼻咽癌高发于中国华南地区，严重威胁我国人民生命健康。基层医院诊断和分期准确性低，造成治疗方案选择不合理，鼻咽癌生存率仅60%；大型肿瘤中心鼻咽癌生存率达80%，但个体化精准治疗开展不充分制约了生存率的进一步提升。针对不同诊疗提升需求，我们基于鼻咽癌临床/影像/病理多组学数据和人工智能理论与方法开展系列研究。.在医学图像标准化方法和分割方法方面，我们提出了元对比学习模型MetaCon，用于消除病理组织切片图像的差异；提出了融合专家经验、可解释的鼻咽癌MR影像自动诊断-分割方法，提高了鼻咽癌肿瘤分割准确性，原发灶和淋巴结分割准确率分别为85.5%和82.3%。.在鼻咽癌人工智能辅助病理和磁共振（MR）定性诊断研究方面，我们建立了深度学习结合弱监督学习方法，用于鼻咽癌病理定性诊断，诊断准确率达99%以上；我们基于T1WI, T2WI和T1WIC三个序列MR影像建立了3D DenseNet模型，在测试集和T早期病例子集中的诊断准确率均为100%，证实了在深度学习辅助下平扫MR影像可替代增强MR影像用于鼻咽癌定性诊断。.在基于多模态数据融合的鼻咽癌动态风险预测研究方面，我们提出了基于图正则的跨生物网络的表示学习方法，多组学数据集成稀疏表示的癌症患者风险分层模型，以及可融合不完备多模态数据、可跨中心建模的鼻咽癌预后预测方法等多模态、跨中心数据分析方法和模型；并建立了融合动态MR影像组学、临床因素和分子标志物的鼻咽癌复发转移风险预测模型，预测准确率达80%以上。