近年来，脑起搏器（Deep Brain Stimulation, DBS）在脑疾病的治疗上发挥越来越重要的作用，其适应症逐渐增多，越来越多的患者因脑起搏器植入手术而获益。激活组织体积（Volume of Tissue Activated, VTA）是评估脑起搏器的电极对周围脑组织影响程度的主流方法，也可以辅助医生进行术后程控快速确定个体化的最佳刺激参数。然而生物神经纤维建模的标准VTA计算方法时间成本过高限制了其在临床的应用，已有的多种VTA快速计算方法在方向性电极逐渐应用的背景下变得不再适用，建立一种能够适用于环形电极和方向性电极的通用的高精度实时预测模型变得越来越迫切。为此，本文针对VTA计算的临床需求开展了如下研究：激活组织体积的计算依赖于电极周围电场的准确计算以及判断神经元在某区域是否激活。本文首先通过脑组织的有限元建模计算，探究了不同脑组织下电场的差别，得到了非均匀脑组织对电场计算结果影响较小的结论，而电极植入后不同时期的水肿和结缔组织会使电场发生一些明显地变化，同时也对各向异性对电场的影响做了初步探究；本文建立了神经元的轴突模型，分析了轴突模型的具体结构及其离子通道的控制方程，探究了其在典型DBS刺激下的响应特点，并总结了轴突模型的响应规律，即轴突模型的响应由其胞外电势二阶导数的分布所主导，且轴突的激活阈值近似与刺激脉宽成反比。而后，本文在上述搭建计算模型的基础上进行了大规模计算，利用轴突在外部刺激下的响应数据复现了AF-Max阈值计算方法，并进一步训练了高效的基于卷积神经网络（Convolutional Neural Network, CNN）的轴突激活预测器，全面评估了CNN计算VTA的效果，与AF-Max方法及其他方法做了比较，证明了卷积神经网络可以快速、准确且稳健地预测DBS刺激下的神经激活反应，从而提高DBS编程的效率。最后，本文探究了轴突预测模型在移动设备上的部署，设计了相应的交互接口并成功地在安卓平台模拟器和实机中运行，在此基础上做了算力分析，并对比了不同场景下的运行时长。而后采取参数缩减和量化的方式将原有CNN模型进行了优化，在VTA计算性能无明显降低的情况下大幅降低了模型运行时间和存储空间，实现了在移动设备中实时计算VTA的优化目标。
In recent years, DBS (deep brain stimulation) have played an increasingly significant role in treating brain disorders, with expanding indications and a growing number of patients benefiting from implantation surgeries. The calculation of the volume of tissue activated (VTA) is the mainstream method for assessing the impact of pacemaker electrodes on surrounding brain tissues and assists physicians in quickly determining optimal individualized stimulation parameters post-surgery. However, the standard VTA calculation method, based on biophysical neural fiber modeling, is time-consuming and its clinical application is limited. Existing rapid VTA calculation methods have become inadequate with the increasing use of directional electrodes, making the development of a universal, high-precision, real-time predictive model for both ring and directional electrodes more urgent. This thesis addresses the clinical needs for VTA computation as follows:The calculation of activated tissue volume relies on accurate computation of the electric field around the electrode and determination of neuronal activation in specific regions. Initially, through finite element modeling of brain tissues, this study explores the variations in electric fields across different brain tissues, concluding that non-homogeneous brain tissues minimally affect the results of electric field calculations. However, post-implantation edema and cicatricial tissue significantly alter the electric fields, and the study also preliminarily investigates the impact of anisotropy on electric fields. Additionally, an axonal model for neurons was established, analyzing its structure and ion channel control equations, examining its response under typical DBS stimulation, and summarizing that the axonal model's response is dominated by the distribution of the second derivative of the extracellular potential, with the axonal activation threshold inversely proportional to the stimulus pulse width.Subsequently, based on these computational models, extensive calculations were performed, utilizing axonal response data under external stimulation to implement the AF-Max threshold calculation method. A highly efficient CNN-based axonal activation predictor was trained and its effectiveness in computing VTA was comprehensively evaluated. The CNN demonstrated rapid, accurate, and robust predictions of neural activation under DBS stimulation, thereby enhancing the efficiency of DBS programming.Finally, this study explored the deployment of the axon prediction model on mobile devices, designed a corresponding interactive interface, and successfully executed it on both Android emulators and physical devices. Subsequently, computational performance was analyzed by comparing runtimes across various scenarios. By employing parameter reduction and quantization techniques, the original CNN model was optimized, significantly decreasing model runtime and storage requirements without notably compromising VTA computational performance, thus achieving the optimization goal of real-time VTA computation on mobile devices.
