Generalizing the mean intercept length tensor for gray-level images

Abstract

The mean intercept length tensor is the most used technique to estimate microstructure orientation and anisotropy of trabecular bone. This paper proposes an efficient extension of this technique to gray-scale images based on a closed formulation of the mean intercept length tensor and a generalization using different angular convolution kernels. First, the extended Gaussian image is computed for the binary or gray-scale image. Second, the intercepts are computed for all possible orientations through an angular convolution with the half-cosine function. Finally, the tensor is computed by means of the covariance matrix. The complexity of the method is O(n + m) in contrast with O(nm) of traditional implementations, where n is the number of voxels in the image and m is the number of orientations used in the computations. The method is generalized by applying other angular convolution kernels instead of the half-cosine function. As a result, the anisotropy of the tensor can be controlled while keeping the eigenvectors intact. The proposed extension to gray-scale yields accurate results for reliable computations of the extended Gaussian image and, unlike the traditional methodology, is not affected by artifacts generated by discretizations during the sampling of different orientations. Experiments show that the computations on both binary and gray-scale images are correlated, and that computations in gray-scale are more robust, enabling the use of the mean intercept length tensor to clinical examinations of trabecular bone. The use of kernels based on the von Mises-Fisher distribution is promising as the anisotropy can be adjusted with a parameter in order to improve its power to predict mechanical properties of trabecular bone.