基于18F-FDG PET/CT的弱监督深度学习方法判断NK/T细胞淋巴瘤预后

NK/T细胞淋巴瘤（ENKTL）异质性大，预后差，亚洲发病远高于西方。其预后判断对于最佳治疗方案选择至关重要。国际预后指数、韩国版预后指数、NK/T预后指数等模型各有不足。我们在内的众多研究提示18F-FDG PET/CT对其预后价值存诸多争议。深度学习方法可从图像中自动提取显性及隐性特征，从而提供更有价值的信息。我们前期采用深度学习中最重要的对抗神经网络方法对数据进行图像配准识别，可最大程度复原病灶。但深度学习方法的建立依赖于高质量的大数据，而临床工作中常有相关病史及影像资料不完整。建立对数据要求低、能最大限度挖掘数据价值，并可用于处理复杂任务的深度学习方法是该领域的关键问题之一。本项目拟在前期研究基础上，基于ENKTL建立一种新型正-负未标记（PNU）弱监督深度学习方法，分析高维特征显像数据，用于判断预后，并提高其在不同型号PET/CT的泛化性，为后续临床治疗方案制订提供精准依据。

Extra-nodal nature killer/T cell lymphoma (ENKTL) is heterogeneous and has a poor prognosis. The incidence in Asia is much higher than that in the West. Predicting of its prognosis is crucial for the treatment strategy. The current prognostic models such as IPI, KPI and PINK-E have their own shortcomings. Numerous studies including us suggested that glucose metabolism based 18F-FDG PET/CT has many controversies concerning its prognostic value. Deep learning methods can automatically extract dominant features and recessive features of different levels from images then provide prognostic information. We have performed image registration and recognition previously on the data based on the adversarial neural network method, which can restore the lesion to the greatest extent. However, the establishment of deep learning methods relies on high-quality big data, and it is difficult to collect completed clinical and imaging data. One of the key issues in this field is to develop methods that require low data distribution, few application restrictions, can handle complex tasks, and maximize data mining. Our project intends to develop a new type of positive-negative unlabeled (PNU) weakly supervised deep learning method based on a variety of devices, analyze high-dimensional feature imaging data, determine the prognosis, and improve its generalization, pave a way for clinical treatment strategy decision.