本项目以对疾病诊疗和人类认知了解具有重要作用的MRI脑影像为对象，以新一代人工智能中突破性的深度学习方法为依托，针对脑肿瘤精细标注样本数量不足与深度神经网络模型训练的大量标签数据需求问题，开展面向MRI脑肿瘤分割的深度神经网络新模型与方法研究。首先，从加强网络特征表达能力出发，引入具有高判别能力的高阶统计特征建模方法，构建基于高阶统计特征建模的MRI脑肿瘤深度分割网络；其次，从提高网络定位能力着手，对高性能目标检测网络进行任务迁移，探索基于区域卷积神经网络的MRI脑肿瘤分割新网络；然后，从脑肿瘤精细标注样本生成着手，开展基于生成对抗网络的MRI脑肿瘤分割网络研究；最后，基于研究方法来开发实用性的MRI脑肿瘤自动分割系统。项目研究有助于提升MRI脑肿瘤分割的研究水平，并可为相关研究提供一定方法借鉴，对于辅助脑疾病临床诊疗、开展大脑结构和功能分析也具有重要意义。

This project takes MRI brain imaging that plays an important role in disease diagnosis and human cognition as research object. Based on the breakthrough deep learning methods in the new generation artificial intelligence, it focuses on new models and methods of deep neural networks for MRI brain tumor segmentation task, aiming at resolving the problem caused by the insufficient fine-labeled brain tumor images and large data requirements of deep learning methods. Firstly, starting from the enhancement of network feature expression ability, high-order statistical feature modeling methods with high discriminative ability are introduced to construct the novel MRI brain tumor depth segmentation networks. Then, to improve the locating capability of the networks, the project tries to transfer the high-precision target detection networks, and constructs new MRI brain tumor segmentation networks based on regional convolutional neural networks. Moreover, aiming at generating more fine-labeled samples of brain tumors, we further explore MRI brain tumor segmentation network based on Generative Adversarial Networks. Finally, a relatively practical automatic MRI brain tumor segmentation system will be developed by using the presented networks. The research results can not only help to improve the research level of MRI brain tumor segmentation, but also provide method reference for the related research. Meanwhile, it is of great significance for the clinical diagnosis of brain diseases, and the function and structure analysis of human brain.