Home | Search | Help  
Home Page Università di Genova

Technical Report Details


Date 23-4-2008
Number DISI-TR-08-10
Title Multi-Cues Regularized Least-Squares applied to Brain MRI Segmentation
Authors C. Basso, E. De Vito, A. Verri
Bibtex Entry @TECHREPORT{basso08tr, author = {Basso, Curzio and De Vito, Ernesto and Verri, Alessandro}, ti
E-mail curzio.basso@disi.unige.it
Link http://www.disi.unige.it/person/BassoC/pubs/basso08tr.pdf
Abstract We present a method for using multiple image cues in a Regularized Least-Squares (RLS) regression scheme. The cues are generic continuous functions defined on the object space, such as the raw intensities or the gradient; their combinations with a Mercer kernel yield a set of cue-specific kernels that we use for regression and classification. The regression problem is cast in the direct sum space of the Reproducing Kernel Hilbert Spaces (RKHS) associated to each cue-specific kernel. This particular formulation of the problem is consistent, and can be solved by iterative or closed-form methods. Setting the problem in the direct sum space allows us to design a feature selection mechanism which operates independently on the training points and on the cue types. We show an implementation of the selection stage based on a consistent algorithm that minimizes the well known elastic-net functional. The method is applied to the automated segmentation of 3D magnetic resonance images (MRI) of the brain, approached as a voxel classification problem. The actual segmentation is performed via a number of one-versus-all least-squares classifiers, trained solving a multi-cue RLS regression problem and combined via bagging. The tests are performed on a set of publicly available, simulated T1 MRIs.
Back to Technical Reports