Label fusion is a powerful image segmentation strategy that is becoming increasingly popular in medical imaging. However, satisfying the requirements of higher accuracy and less running time is always a great challeng...Label fusion is a powerful image segmentation strategy that is becoming increasingly popular in medical imaging. However, satisfying the requirements of higher accuracy and less running time is always a great challenge. In this paper we propose a novel patch-based segmentation method combining a local weighted voting strategy with Bayesian inference. Multiple atlases are registered to a target image by an advanced normalization tools(ANTs) algorithm. To obtain a segmentation of the target, labels of the atlas images are propagated to the target image. We first adopt intensity prior and label prior as two key metrics when implementing the local weighted voting scheme, and then compute the two priors at the patch level. Further, we analyze the label fusion procedure concerning the image background and take the image background as an isolated label when estimating the label prior. Finally, by taking the Dice score as a criterion to quantitatively assess the accuracy of segmentations, we compare the results with those of other methods, including joint fusion, majority voting, local weighted voting, majority voting based on patch, and the widely used Free Surfer whole-brain segmentation tool. It can be clearly seen that the proposed algorithm provides better results than the other methods. During the experiments, we make explorations about the influence of different parameters(including patch size, patch area, and the number of training subjects) on segmentation accuracy.展开更多
基金Project supported by the National Natural Science Foundation of China(No.61203224)the Science and Technology Innovation Foundation of Shanghai Municipal Education Commission,China(No.13YZ101)
文摘Label fusion is a powerful image segmentation strategy that is becoming increasingly popular in medical imaging. However, satisfying the requirements of higher accuracy and less running time is always a great challenge. In this paper we propose a novel patch-based segmentation method combining a local weighted voting strategy with Bayesian inference. Multiple atlases are registered to a target image by an advanced normalization tools(ANTs) algorithm. To obtain a segmentation of the target, labels of the atlas images are propagated to the target image. We first adopt intensity prior and label prior as two key metrics when implementing the local weighted voting scheme, and then compute the two priors at the patch level. Further, we analyze the label fusion procedure concerning the image background and take the image background as an isolated label when estimating the label prior. Finally, by taking the Dice score as a criterion to quantitatively assess the accuracy of segmentations, we compare the results with those of other methods, including joint fusion, majority voting, local weighted voting, majority voting based on patch, and the widely used Free Surfer whole-brain segmentation tool. It can be clearly seen that the proposed algorithm provides better results than the other methods. During the experiments, we make explorations about the influence of different parameters(including patch size, patch area, and the number of training subjects) on segmentation accuracy.
文摘部分多标签学习(partial multi-label learning,PML)旨在训练实例的标签集中同时处理真实标签和噪声标签。当前大多数PML只考虑单视图场景,却忽略了现实中多视图场景的情况。虽然部分研究利用了多视图特征信息,但多依赖于获取多视图的子空间,对多视图特征信息的学习不够充分。针对以上挑战,提出了一种新颖的多视图融合的PML预测模型PMLMF(partial multi-label learning based on multi-view fusion)。利用矩阵非负分解获得共享子空间,以捕获多视图的共享信息,并将其与原始的多视图数据集相融合,构建新的多视图数据集。利用低秩表示获得标签相关性系数矩阵,从而有效去除噪声并恢复相关标签。将模型推广到非线性版本,以有效地处理线性不可分割的问题。在7个多视图部分多标签数据集中进行了大量的实验,充分验证了该方法的有效性。
