Spectral Clustering and Label Fusion For 3D Tissue Classification: Sensitivity and Consistency Analysis

Clustering algorithms have found application in tissue classification in MRI. Standard techniques such as k-means iteratively define intensity clusters based on the distribution of voxels in intensity space. Spectral clustering is potentially more powerful as it models voxel-to-voxel relationships rather than voxel-to-cluster relationships. Unfortunately, for images of n-voxels naive application leads to an n(n− 1)/2 voxel comparison problem and an order n× n eigenvalue problem which has prevented these techniques being widely investigated in 3D medical imaging. In this paper we report an empirical evaluation of a stochastic sampling approach to modelling voxel-to-voxel relationships for spectral clustering. Stochastic sampling captures sufficient intensity structure to give plausible tissue classification in 3D brain MRI. We test the stability of our approach to similarity parameter choice, sample size and stochastic effects in simulated and real 3D MR images.