本研究在今年的研究进展主要包括：基于神经网络模型的级联配准策略和基于深度学习模型的疾病预测研究方法。研究组针对脑图像大形变问题和大脑局部复杂结构配准问题进行研究，提出了一种级联配准策略解决大形变问题，在级联配准后期的精配准阶段，通过损失函数的约束调整解决局部细微结构的配准问题。此外研究组还结合访学合作项目的背景，对基于深度学习模型提取脑MRI图像特征进行疾病预测的潜力进行了初步探索。The research progress of this study this year mainly includes cascade registration strategy based on neural network models and disease prediction research methods based on deep learning models.The research group researched the problem of large deformation in brain images and the registration of complex local brain structures and proposed a cascaded registration strategy to solve the problem of large deformation.In the precision registration stage of the later stage of cascaded registration,the registration problem of local fine structures is solved by adjusting the constraints of the loss function.In addition,the research group also explored the potential of extracting brain MRI image features for disease prediction based on deep learning models,taking into account the background of the visiting study cooperation project.

本研究主要基于大脑MRI结构图像进行图像特征学习并进行图像配准和疾病预测相关的研究。本研究的方法主要基于深度学习模型进行图像特征学习与提取，尤其是深度卷积网络。利用深度学习模型，研究组前期在脑MRI图像分割、脑MRI图像去噪和去伪影、PET到CT图像合成、脑MRI图像配准等方面进行了多方位的初步探索。经过前期的研究，研究组发现利用深度学习模型提取更有效的MRI特征是改善图像配准和疾病预测性能的关键。因此，研究组后期主要聚焦于如何基于深度学习模型更为有效的利用MRI图像特征来改善图像配准和疾病预测性能并提出了一种能够灵活融合深度学习特征与传统特征的框架，该框架能够明显提高深度学习模型的疾病预测性能。该项研究成果已在医学图像研究领域顶级期刊《Medical Image Analysis》上发表。This study is mainly based on brain MRI structural images for image feature learning and related research on image registration and disease prediction.The method of this study is mainly based on deep learning models for image feature learning and extraction,especially deep convolutional networks.Using deep learning models,the research group conducted preliminary explorations in various aspects such as brain MRI image segmentation,brain MRI image denoising and artifact removal,PET to CT image synthesis,and brain MRI image registration in the early stage.After preliminary research,the research team found that using deep learning models to extract more effective MRI features is key to improving image registration and disease prediction performance.Therefore,the research group mainly focuses on how to more effectively utilize MRI image features based on deep learning models to improve image registration and disease prediction performance and proposes a framework that can flexibly integrate deep learning features with traditional features.This framework can significantly improve the disease prediction performance of deep learning models.The research findings have been published in'Medical Image Analysis',which is the top journal in medical image research.

本研究在今年的研究进展主要包括:(1)通过对脑图像分割进行文献调研，为利用分割图进行复杂结构的配准研究提供启发;(2)针对MRI图像噪声和网格状伪影的去除进行研究，改善MRI图像信噪比;(3)改进利用分割图进行MRI图像配准模型训练进行弱监督的方法，从而改善复杂结构的配准精度;(4)对PET图像到CT图像的转换进行探索;(5)初步尝试利用对抗生成模型进行MRI图像配准。这些研究为研究组后续的研究聚集提供了一定的理论基础和丰富的实践经验。。The research progress of this study this year mainly includes:(1)conducting literature research on brain image segmentation,providing inspiration for the registration research of complex structures using segmentation images;(2)Research on the removal of MRI image noise and grid artifacts to improve the signal-to-noise ratio of MRI images;(3)Improve the method of training MRI image registration models using segmentation maps for weak supervision,thereby improving the registration accuracy of complex structures;(4)Explore the conversion of PET images to CT images;(5)Preliminary attempt to use adversarial generative models for MRI image registration.These studies provide a certain theoretical basis and rich practical experience for the subsequent research aggregation of the research group.