ABSTRACT

In brain surgery, respiration motion, outliers, and occlusions create artefacts in Infrared Thermography (IRT) imaging. In this paper, we propose a robust method to handle multiple motion, outliers, and occlusions in IRT images which consists of two phases: preprocessing and image motion estimation. In the preprocessing phase, the Region of Interest (RoI) segmentation is employed to extract the brain cortex only. Thereafter, the Phase Correlation method is employed to compensate for large motion followed by occlusion masking based on an approach applying Cellular Nonlinear Networks (CNN). Next, intensity adjustment is made with respect to the reference image. Then, a Gaussian filter is applied. In the following phase, the image motion is estimated by employing Combined Local-Global (CLG) optical flow method. In order to find the best regularization coefficient for the spatial coherency term and the number of iterations for recursive optical flow estimation, we generated ground truth and evaluated the accuracy of the estimated motion vectors based on Average Angular Error (AAE) and Average Magnitude Error (AME). The efficiency improvement of the proposed method was tested on 1024 IRT images based on different comparisons. Thereby, the proposed method shows promising results for motion estimation and correction application in brain surgery.

摘要

在脑外科手术中，呼吸运动、异常值和遮挡会在红外热成像 (IRT) 成像中产生伪影。在本文中，我们提出了一种鲁棒的方法来处理 IRT 图像中的多个运动、异常值和遮挡，该方法由两个阶段组成：预处理和图像运动估计。在预处理阶段，兴趣区域（RoI）分割仅用于提取大脑皮层。此后，基于应用蜂窝非线性网络 (CNN) 的方法，采用相位相关方法来补偿大运动，然后进行遮挡掩蔽。接下来，对参考图像进行强度调整。然后，应用高斯滤波器。在接下来的阶段，采用组合局部全局（CLG）光流方法估计图像运动。为了找到空间相干项的最佳正则化系数和递归光流估计的迭代次数，我们生成了地面实况并评估了基于平均角误差 (AAE) 和平均幅度误差 (AAE) 估计的运动矢量的准确性（ AME）。基于不同的比较，在 1024 张 IRT 图像上测试了所提出方法的效率改进。因此，所提出的方法在脑外科手术中的运动估计和校正应用中显示出有希望的结果。基于不同的比较，在 1024 张 IRT 图像上测试了所提出方法的效率改进。因此，所提出的方法在脑外科手术中的运动估计和校正应用中显示出有希望的结果。基于不同的比较，在 1024 张 IRT 图像上测试了所提出方法的效率改进。因此，所提出的方法在脑外科手术中的运动估计和校正应用中显示出有希望的结果。