Structured Learning and Prediction in Computer Vision
Foundations and Trends® in Computer Graphics and Vision
| Hi-index | 0.00 |
In cancer, pathological tissue often exhibits abnormal perfusion and vascular permeability. These can be estimated by monitoring the abundance of an injected contrast medium over time, using Dynamic Contrast-Enhanced (DCE) MR Imaging. The resulting spatially resolved time curves are usually interpreted in terms of a pharmacokinetic model which is fitted by maximum likelihood. However, the resulting nonlinear least squares (NLLS) problem may exhibit spurious local optima leading to false parameter estimates at individual voxels in the generated parameter map. We propose the application of a spatial prior model in form of a generalized Gaussian Markov random field. By using information from parameter estimates at neighboring voxels and computing a maximum a posteriori solution for the whole parameter map at once, false local optima at individual voxels can be avoided. Since the number of variables gets very big for common image resolutions, standard NLLS solvers cannot be employed anymore. We therefore propose a generalized iterated conditional modes (ICM) approach operating on blocks instead of sites. Results on DCE-MR images of the prostate show less speckle noise in the resulting parameter maps. Furthermore, the mean square error (MSE) in the affected voxels is significantly smaller, thus reflecting a better fit.