近年来，脑疾病和脑科学成为研究热点，脑深部电刺激疗法作为重要的脑疾病疗法，其适应症快速增长，为越来越多的患者减轻病痛。本文从脑疾病治疗中至关重要的电极植入位置出发，探讨传统电极位置识别方法的误差来源与修正。针对基于磁共振图像的电极位置精准识别这一问题，本文结合神经网络算法进行了研究。脑深部电刺激的疗效依赖于准确的电极植入，最佳刺激靶点的研究也依赖于对电极植入位置的准确评估。传统的计算机断层扫描和磁共振图像融合的方法，假定了术前与术后的脑组织位置相同，忽视了可能存在的脑漂现象，导致术后电极位置识别出现偏差。本文基于临床试验数据，提出了全脑位移向量场分析方法，证实了患者术后脑漂长期存在，发现脑漂与偏侧化、方向性、时间效应相关，并提出了修正脑漂的方法。提高传统融合方法的位置识别精度，需要用到术后磁共振图像修正脑漂。而术后磁共振图像可以直接用于电极位置识别，这是更加理想的范式。该范式无电离辐射，无需限制扫查频次，并可以结合磁共振相容脑深部电刺激系统，有着巨大的应用潜力。然而磁共振成像的复杂特性，导致图像中出现比电极尺寸更大的低幅值伪影区域，影响电极位置的识别，且多数情况下伪影中心与电极触点中心并不重合。本文通过理论分析与体模实验研究了伪影中心偏移，发现磁共振图像直接识别电极的偏差主要位于电极轴向，影响因素主要为磁场角，且偏差值近似与磁场角二倍的正弦值成正比，由此提出电极位置识别的粗修正公式。在对于磁共振图像电极伪影规律的深入认识基础上，本文提出基于磁共振图像进行电极位置识别的新方法。考虑到传统方法难以兼顾磁共振伪影的非线性特征与伪影的动态复杂特性，而机器学习方法可以捕捉到复杂的特征。因此本文基于磁共振电极伪影模板，提出了神经网络位置识别算法。通过数据增强方法对有限的患者伪影模板进行扩增，训练出精准位置识别神经网络，最终位置识别误差平均值可达0.35 mm。本文提出的机器学习算法，相对于传统术后电极位置识别方法具有精度更高、鲁棒性更强的技术特征。最后通过疗效预测、最佳触点选择以及术后长期电极位移评估几个方面，验证了精准电极位置识别技术在临床中的应用价值。
In recent years, brain diseases and brain science have become a research hotspot. As an important brain disease therapy, deep brain electrical stimulation therapy has a rapid growth in its indications, reducing the pain for more and more patients. This paper discusses the error source and correction of the traditional electrode position recognition method from the electrode implantation position which is very important in the treatment of brain diseases; Then, aiming at the problem of accurate location of magnetic resonance, combined with neural network algorithm, this paper makes an exploration and research.Firstly, the efficacy of deep brain electrical stimulation depends on accurate electrode implantation, and the study of the best stimulation target also depends on the accurate evaluation of the electrode implantation position. The traditional fusion method postoperative electrode position recognition method depends on the fusion of preoperative and postoperative structural image data. This method assumes that the brain tissue positions at two different times are the same, and introduces brain drift deviation. This paper evaluates this deviation in clinical trial data, and puts forward a method to correct brain drift. In traditional fusion methods, brain drift correction depends on postoperative magnetic resonance images. If the brain drift correction can be performed directly using postoperative magnetic resonance data, it will be the most ideal electrode position recognition paradigm, and the scanning frequency will no longer be limited by the safe dose of radiation.Magnetic resonance imaging (MRI) is used more and more widely. It has the characteristics of no radiation. However, the RF magnetic field in the imaging process can cause the heating of the electrode tip of brain pacemaker and bring potential safety hazards. After the emergence of magnetic resonance compatible deep brain electrical stimulation system, the potential of magnetic resonance imaging in clinical application has been released. However, the complex characteristics of magnetic resonance imaging and the difference of magnetic sensitivity between the electrode and brain tissue make the electrode present a low amplitude artifact region larger than the size of the electrode in the image, and in most cases, the artifact center does not coincide with the contact center. This paper studies the influencing factors of artifact center offset through theoretical analysis and phantom experiment, and improves it through theoretical formula.Based on the in-depth understanding of the law of electrode artifacts in magnetic resonance images, this paper attempts to propose a new method based on magnetic resonance electrode localization. Considering that traditional methods are difficult to take into account the nonlinear characteristics of MRI artifacts and the dynamic complexity of artifacts, however machine learning methods can capture complex features. Therefore, a convolution neural network localization algorithm based on magnetic resonance electrode artifact is proposed in this paper. The limited patient artifact data are amplified by data enhancement method. The neural network is accurately located at the training place, and the average localization error can reach 0.35 mm. The machine learning algorithm proposed in this paper has the technical characteristics of high precision and strong robustness compared with the traditional postoperative electrode position calibration method.Finally, the application value of accurate electrode positioning technology in clinic is verified through the prediction of clinical outcome, the selection of the best contact and the long-term electrode displacement after operation.
