A new active shape models (ASMs) was presented, which is driven by scale invariant feature transform (SIFT) local descriptor instead of normalizing first order derivative profiles in the original formulation, to segme...A new active shape models (ASMs) was presented, which is driven by scale invariant feature transform (SIFT) local descriptor instead of normalizing first order derivative profiles in the original formulation, to segment lung fields from chest radiographs. The modified SIFT local descriptor, more distinctive than the general intensity and gradient features, is used to characterize the image features in the vicinity of each pixel at each resolution level during the segmentation optimization procedure. Experimental results show that the proposed method is more robust and accurate than the original ASMs in terms of an average overlap percentage and average contour distance in segmenting the lung fields from an available public database.展开更多
A learning-based deformable registration method was presented for MR brain images. First, best geometric features are selected for each location and each resolution, in order to reduce ambiguity in matching during ima...A learning-based deformable registration method was presented for MR brain images. First, best geometric features are selected for each location and each resolution, in order to reduce ambiguity in matching during image registration. The best features are obtained by solving an energy minimization problem, which requires the features to be distinctive around the neighboring points and consistency across training samples. Secondly, the set of active points is hierarchically selected based on their saliency and consistency measurements during registration, which helps to produce accurate registration results. Finally, by incorporating those learned results into the framework of HAMMER, great improvement in both real data and simulated data is achieved.展开更多
基金The National Natural Science Foundation of China(No60271033)
文摘A new active shape models (ASMs) was presented, which is driven by scale invariant feature transform (SIFT) local descriptor instead of normalizing first order derivative profiles in the original formulation, to segment lung fields from chest radiographs. The modified SIFT local descriptor, more distinctive than the general intensity and gradient features, is used to characterize the image features in the vicinity of each pixel at each resolution level during the segmentation optimization procedure. Experimental results show that the proposed method is more robust and accurate than the original ASMs in terms of an average overlap percentage and average contour distance in segmenting the lung fields from an available public database.
基金National Natural Science Foundation of China(No.60271033)
文摘A learning-based deformable registration method was presented for MR brain images. First, best geometric features are selected for each location and each resolution, in order to reduce ambiguity in matching during image registration. The best features are obtained by solving an energy minimization problem, which requires the features to be distinctive around the neighboring points and consistency across training samples. Secondly, the set of active points is hierarchically selected based on their saliency and consistency measurements during registration, which helps to produce accurate registration results. Finally, by incorporating those learned results into the framework of HAMMER, great improvement in both real data and simulated data is achieved.
