【摘要】 基于多图谱的分割方法能有效解决脑MR图像中特定对象如海马体、杏仁核等组织的自动精确分割问题.为加快处理速度,该方法需要从大型图谱数据集中挑选与目标分割图像比较近似的若干图谱作为分割参考.传统的多图谱分割方法在选择图谱时通常只依据图谱图像与目标图像在灰度上的相似性,没有考虑到两者在分割对象局部形态上的相似性,使得所选择的图谱对分割的参考价值难以保证.针对这一问题,文中提出一种图谱图像与目标图像的以分割对象为中心的加权相似性度量方法,首先通过将图谱图像向目标图像作全局配准得到分割对象在目标图像中的估计位置,然后根据分割对象在两幅图像间的局部扭曲形变来度量它们的相似性.文中将这种加权相似度应用到多图谱分割方法中,在图谱选择时只挑选与目标图像具有较高加权相似度的部分图谱,在融合分割标记时将加权相似度作为权重.对IBSR脑MR图像中豆状核壳核组织的分割实验结果表明,基于该加权相似性度量的多图谱分割方法可以得到较高的分割精度.更多还原

【Abstract】 The multi-atlas based segmentation method provides an effective solution for automatically and accurately segmenting specific tissues such as the hippocampus and amygdala from brain MR images.In order to speed up processing,this method needs to pick out those atlases which are similar to the target segmenting image to be the references of segmentation.Traditional multi-atlas methods generally select atlases in accordance with the intensity similarity between atlas image and target image,without considering the morphology similarity of both images in local of the segmenting subject,making the reference value of selected atlases be unguaranteed.Toaddress this shortage,this paper proposes a segmenting subject centric weighted image similarity measurement.First,the atlas image is registered to the target image globally to obtain the estimated localization of the segmenting subject in target image;and then,the similarity is measured by the local distortion of the segmenting subject between two images.We apply this weighted similarity into multi-atlas based segmentation method,where only those atlases which have high weighted similarities with the target image are picked for segmenting,and the segmentation labels are also fused by taking the weighted similarity as the weight.Experiments on segmenting the putamen tissue of brain MR images from IBSR demonstrate that this weighted similarity based multi-atlas segmentation method can achieve high accuracy.更多还原