基于深度学习的癫痫多模态脑影像分类研究

Deep learning has a high degree of abstraction and complexity, and is more suitable for the comparatively complex, scattered and refined neural brain imaging data of the detection model. Brain structural negative epilepsy is not associated with detectable lesions, the classification algorithm of multi-modal neuroimaging data and epileptic typing for epilepsy have important clinical significance. The project is proposed in view of the structure of the negative epilepsy classification problem, using the big data of multimodal brain imaging to develop a deep learning pattern classification algorithm based on coded matrix marking to achieve different subtypes of patients. Further, we develop an unsupervised convolution deep belief networks algorithm, which is applied to multimodal, multi-index and multivariable epileptic imaging data of structural negative epilepsy patients, so as to improve the specificity and sensitivity of individual screening. Finally, we set up the characteristics of epileptic multimodal dynamic brain network and find the key brain areas in patients' brain network reorganization, thus depict the brain activity and brain network topological structure before and after treatment and broaden the clinical evaluation system of epilepsy treatment. The aim of this project is to develop a new deep learning method for multi-modal neuroimaging data, achieve the objective and accurate classification of epilepsy involved in the theory and key technical issues.

深度学习具有高阶的抽象度和复杂度，更适用于检测模式较为复杂、分散和精细的脑影像数据。脑结构阴性癫痫未表现出明确的可探性病灶，因此，发展其多模态脑影像数据的分类算法对癫痫分类具有重要的临床意义。本项目拟针对结构阴性癫痫分类难题，利用多模态脑影像大数据，发展基于编码矩阵标记的深度学习模式分类算法，实现癫痫亚型分类；进一步，发展基于非监督的卷积深度置信网算法，用于结构阴性癫痫患者的多模态、多测度、多变量癫痫脑影像数据，提高个体化甄别的特异性、敏感性；最后，建立癫痫多模态治疗前后脑网络，探测癫痫脑网络重组的关键性脑区，从而刻画患者治疗前后的脑活动及脑网络拓扑结构的演化过程，完善癫痫诊疗临床评价体系。本项目从癫痫患者临床诊断的重要需求为出发点，发展基于多模态脑影像数据的深度学习算法，实现客观、准确的癫痫亚型脑影像分类和个体化甄别。

脑结构阴性癫痫未表现出明确的可探性病灶，针对结构阴性癫痫分型难题，因此，发展多模态多尺度脑影像数据的方法和癫痫多分类所涉及的理论、关键技术，对探测癫痫病理生理和客观影像表征具有重要的科学价值和临床意义。. 通过本项目的实施，收集多亚型癫痫，包括内侧颞叶癫痫、额叶癫痫、全面性癫痫患者的多模态脑影像数据。1）发展脑功能、结构梯度算法，发现全面性癫痫功能连接 组梯度的空间异常表征；2）建立大脑默认网络时变演变模型，发现了不同癫痫亚型普遍动态异常，以及亚型特定的改变，为癫痫亚型分类提供影像证据；3）提出脑形态学相似性网络+基因表达融合方法，发现全面性癫痫患者脑形态学异常，其基因表达主要富集在突触组织和神经递质运输的通路，从而为理解癫痫异常提供了宏观和微观尺度融合视角。. 本项目执行期间，共发表（含接收）SCI学术期刊论文13篇（11篇影响因子大于5），学术期刊包括Nature Communications, Cell Reports, IEEE Transactions on Medical Imaging, NeuroImage, Neuropathology and Applied Neurobiology, Epilepsia；获软件著作权1项；培养博士研究生3名，硕士研究生6名。