针对广东高发癌症鼻咽癌的筛查产生了大量CT图像，但因缺少大规模专家标注数据限制了监督式深度神经网络的应用。此外，学习复杂模型需要高性能计算设施，进一步阻碍了计算机辅助诊断的研究和推广。因此在医学图像分析的深度学习方案中，如何充分利用海量未标注数据，同时有效提高计算效率是亟待解决的关键问题。本项目采用半监督学习和深度强化学习，通过对半监督识别模型、高性能深度学习算法和病灶识别的高效率计算方法三项内容的研究，实现低标注数据依赖、高计算效率的病灶区域识别总体目标。本项目创新性地采用基于主动学习和自学习方法的半监督模型充分挖掘未标注数据信息；通过引入层次性范畴分级概念，构建高泛化能力的特征探索机制，实现高准确率的多征象类型联合识别；并利用人工智能体实现高效率的病灶区域检测。本项目提出的高效率低标注数据依赖的病灶区域识别方法，有望为实现高计算性能、低标注数据依赖的医学影像智能分析提供关键技术支撑。

The screening of nasopharyngeal carcinoma (NPC), high incidence of cancer in Guangdong, has produced a large number of CT images, but the lack of large-scale expert annotated data limits the application of supervised deep neural network. In addition, learning complex models requires high-performance computing facilities, which further hinders the research and popularization of computer-aided diagnosis system. Therefore, in the deep learning scheme of medical image analysis, the urgent key problem is how to improve the computational efficiency effectively while making full use of the massive unannotated data. The overall target of this project is to realize lesion area recognition with low annotated data dependence and high computational efficiency. This project adopts semi-supervised learning and deep reinforcement learning to research the semi-supervised recognition model, high-performance deep learning algorithm and efficient calculation method of lesion recognition. This project innovatively uses the semi-supervised recognition model based on active learning and self-learning methods to mine the unannotated data information, constructs the feature exploration mechanism of high generalization ability by introducing the concept of hierarchical category classification to achieve high accuracy of multi-sign type joint recognition, and uses artificial intelligence agent to detect lesion area efficiently. The proposed method of lesion recognition with high efficiency and low annotated data dependence may provide key technical support for medical image intelligent analysis with high computing performance and low annotated data dependence.