School of Automation, Guangdong University of Technology
采用原始的蒙皮区域卷积神经网络(Mask Region-based Convolutional Neural Network, Mask R-CNN) 获取周围神经MicroCT图像中的神经束轮廓时存在收敛慢、精度低等问题。首先构建两个数据子集。然后提出了一种密集连接型网络结构,提取神经束区域特征。此外改进了目标检测部分候选框的得分评价规则,并结合迁移学习策略改进原始算法的训练方式。采用准确率和交并比指标评价算法的准确度,精细度阈值指标评价轮廓获取的精度,并首次确定了精细度阈值的最佳值。实验结果表明,改进后算法在两个数据子集中的准确率和交并比均在83%和87%以上。在精细度阈值为0.85时,获得的神经束轮廓最佳。由此可见,改进后算法能够良好的实现从周围神经MicroCT图像中获取神经束轮廓的目标,为周围神经内部结构的三维可视化奠定基础。
An improved Mask Region-Convolutional Neural Network (Mask R-CNN) algorithm is proposed to conquer the shortcoming such as the slow convergence rate, low accuracy in the original Mask R-CNN algorithm to obtain contours of fascicular groups from MicroCT images of peripheral nerve. First, the dataset of images was constructed and divided into two subsets. Second, the network architecture with dense connection is proposed to abstract the feature of fascicular groups. Third, the regulation of proposal box scores in object detection part was improved. Additionally, the transfer learning strategy was combined with Mask R-CNN in training process. The average precision (AP) and the Intersection over Union (IoU) are adopted as evaluation indices of algorithm accurate, and the precision threshold is adopted as the evaluation index of algorithm precision. It was the first time identifies the best values of the precision threshold. Experiment results show that the AP and the IoU of the improved approach exceeds 83% and 87% in two peripheral nerve MicroCT image subsets. The improved algorithm has the best contours of fascicular groups at the threshold of 0.85. Experiments show that the improved algorithm can extract the contours of fascicular groups exactly and lay the foundation for the three dimensional visualization of the internal structure of peripheral nerve.