基于卷积神经网络和自适应卡尔曼滤波的脑功能FMT/PAT联合动态成像方法

The main feature of brain activity is the continuous change of the underlying brain states, even in a constant environment. It is always a complex phenomenon in a highly coordinated manner in different areas across the whole brain. Observation on a single scale of space or time is difficult to predict and understand the brain activity process accurately. Therefore, the in vivo visualization of the dynamic function across the whole brain is of great significance for analyzing the principle and mechanism of the brain, and for understanding the complex neural activities related to the brain diseases. However, due to their limitations, single imaging modality is difficult to achieve continuous and non-ionized dynamic measurement of brain function with high sensitivity, high temporal and spatial resolution and high detection depth. Therefore, the purpose of this study is to establish a combined fluorescence/photoacoustic dynamic measurement method and imaging theory of the brain function with a high sensitivity, high temporal and spatial resolution in different brain diseases models. This study will set up a combined fluorescence / photoacoustic imaging system and develop a dynamic imaging theory of brain function based on the convolutional neural network and the adaptive extended Kalman filtering. Finally, the proposed system and theory will be validated with the phantom and the small animal experiments using different fluorescent probes (indocyanine green, calcium fluorescence probe, tumor targeting probe etc.).

大脑的状态即使在恒定的环境下也在不断变化，并表现出不同脑区域高度协同合作的复杂现象，单一空间或时间尺度上的观测是很难准确预测和理解大脑活动的进程。因此，实现全脑范围动态功能在体可视化对分析大脑的工作原理、运行机制以及对理解脑疾病相关的复杂神经活动具有重要意义。然而，单一成像方式由于其各自的局限性，难以实现全脑范围高灵敏度、高时空分辨率、高探测深度、无电离损伤的脑功能连续动态测量。因此，本研究旨在建立荧光/光声联合脑功能动态测量方法和成像理论，实现不同脑疾病模型下高灵敏度、高时空分辨率的全脑范围无创动态脑功能多参数成像研究。本研究将建立荧光/光声一体化联合成像系统，发展基于卷积神经网络与自适应扩展卡尔曼滤波的动态荧光/光声脑功能成像理论，并选取不同荧光探针（吲哚菁绿、钙离子荧光探针、肿瘤靶向探针等）开展有效的仿体和活体小动物验证研究。

传统的静态荧光分子层析成像（fluorescence molecular tomography，FMT）仅能反映特定时刻组织体内荧光剂聚集状态。而在一些疾病模型中，尤其是脑功能成像中，单一空间或时间尺度上的观测是很难准确地预测和理解大脑活动的进程的。特别地，肿瘤组织相对正常组织有很大的空间异质性，包括特异性分子标记物的表达、非特异性试剂的代谢过程以及试剂的渗透性等。动态FMT成像可以提供静态FMT成像无法比拟的肿瘤的丰富特征。除此之外，使用特定的靶向分子探针、非靶向荧光试剂和载药纳米粒子等，动态FMT成像还可以研究相关试剂的药代动力学特征，是发展新型药物和成像试剂的必要步骤，是评估药物治疗效果非常重要的手段。为实现高灵敏度、高时空分辨率的全脑范围脑功能动态成像，单一成像方式难以满足全部成像要求，因此多模态融合的方式势在必行。.首先，我们设计并搭建了基于光子计数动态FMT成像系统，并进行了相应的仿体和在体实验验证了系统的有效性，测试了系统的灵敏度和分辨率；其次，建立了基于自适应扩展卡尔曼滤波（Extend Kalman filtering, EKF）的动态FMT算法，采用广义回归神经网络(generalized regression neural network, GRNN) 对不同测量条件下的测量误差协方差矩阵进行预测，模拟实验验证了方法的有效性；最后，设计并进行了超声/荧光双模态系统和相关仿体实验，仿体实验中以超声结果为先验信息引导FMT成像实现了更优的成像结果。