Toth Robert, Chappelow Jonathan, Rosen Mark, Pungavkar Sona, Kalyanpur Arjun, Madabhushi Anant
Rutgers, The State University of New Jersey, New Brunswick, NJ, USA.
Med Image Comput Comput Assist Interv. 2008;11(Pt 1):653-61. doi: 10.1007/978-3-540-85988-8_78.
In this paper we present MANTRA (Multi-Attribute, Non-Initializing, Texture Reconstruction Based Active Shape Model) which incorporates a number of features that improve on the the popular Active Shape Model (ASM) algorithm. MANTRA has the following advantages over the traditional ASM model. (1) It does not rely on image intensity information alone, as it incorporates multiple statistical texture features for boundary detection. (2) Unlike traditional ASMs, MANTRA finds the border by maximizing a higher dimensional version of mutual information (MI) called combined MI (CMI), which is estimated from kNN entropic graphs. The use of CMI helps to overcome limitations of the Mahalanobis distance, and allows multiple texture features to be intelligently combined. (3) MANTRA does not rely on the mean pixel intensity values to find the border; instead, it reconstructs potential image patches, and the image patch with the best reconstruction based on CMI is considered the object border. Our algorithm was quantitatively evaluated against expert ground truth on almost 230 clinical images (128 1.5 Tesla (T) T2 weighted in vivo prostate magnetic resonance (MR) images, 78 dynamic contrast enhanced breast MR images, and 21 3T in vivo T1-weighted prostate MR images) via 6 different quantitative metrics. Results from the more difficult prostate segmentation task (in which a second expert only had a 0.850 mean overlap with the first expert) show that the traditional ASM method had a mean overlap of 0.668, while the MANTRA model had a mean overlap of 0.840.
在本文中，我们提出了MANTRA（基于多属性、非初始化、纹理重建的主动形状模型），它融合了许多改进流行主动形状模型（ASM）算法的特性。与传统的ASM模型相比，MANTRA具有以下优势。（1）它不仅仅依赖于图像强度信息，因为它融合了多个用于边界检测的统计纹理特征。（2）与传统ASM不同，MANTRA通过最大化一种称为组合互信息（CMI）的更高维互信息版本来找到边界，该互信息是从kNN熵图估计得到的。CMI的使用有助于克服马氏距离的局限性，并允许智能地组合多个纹理特征。（3）MANTRA不依赖于平均像素强度值来找到边界；相反，它重建潜在的图像块，并且基于CMI具有最佳重建效果的图像块被视为物体边界。我们的算法通过6种不同的定量指标，针对近230张临床图像（128张1.5特斯拉（T）T2加权的体内前列腺磁共振（MR）图像、78张动态对比增强乳腺MR图像以及21张3T体内T1加权前列腺MR图像）的专家标注真值进行了定量评估。在更具挑战性的前列腺分割任务中（其中第二位专家与第一位专家的平均重叠率仅为0.850），结果表明传统ASM方法的平均重叠率为0.668，而MANTRA模型的平均重叠率为0.840。
