基于U-Net的多图谱标签融合算法

Multi-atlaslabel fusion based on U-Net

为有效提高海马体多图谱分割算法的精度,将U-Net卷积神经网络用于海马体MR多图谱分割的标签融合。算法在图谱选择阶段,计算互信息、梯度相似性选择图谱,避免周围组织结构对图谱选择的干扰,选择与目标图谱更贴合的浮动图像组。在预处理阶段,提取以海马体为中心的感兴趣区域有效降低数据规模。在配准过程中,利用重采样代替粗配准环节,减少了“粗”配准环节所需时间,再采用具有良好的平滑性、连续性和拓扑保持性的微分同胚Demons精配准算法。在标签融合阶段,提出基于深度学习理论改进的U-Net多图谱MRI海马体分割算法。实验结果表明,改进的算法分割精度较于传统算法平均提高了5%,算法时间缩短了50%左右。改进后的基于U-Net的多图谱海马体分割算法对目标图像海马体的分割具有高精度、高效率的特点。

In order to effectively improve the accuracy of multi-atlas segmentation algorithm of hippocampus,U-Net convolutional neural network is applied to label fusion of multi-atlas.The algorithm in the atlases selection performs mutual information and gradient similarity calculation,avoids the interference of the surrounding tissue structure on the atlas selection,selects the floating image group which is more suitable for the target map.In the pre-processing stage,extracting the region of interest centered on hippocampus can effectively reduce the size of data.In the registration process,re-sampling is used instead of the coarse registration,which reduces the time,and then uses the diffeomorphic demons algorithm,which has good smoothness,continuity and topological retentiveness.In the label fusion stage,an improved U-Net network based on deep learning theory for multi-atlas MRI hippocampal segmentation algorithm is proposed.The experimental results show that the segmentation accuracy of the improved algorithm is about 5%higher than that of the traditional algorithm,the algorithm time is reduced by about 50%.The improved U-Net network based multi-atlas hippocampal segmentation algorithm has the characteristics of high precision and high efficiency for segmentation of the hippocampus in target image.