Stacked Sequential Multi-class Discriminative Dictionary Learning for Brain MRI Segmentation

Abstract

The segmentation of brain structures in Magnetic Resonance Imaging is a challenging problem due to the low contrast and resolution of the structures and the noisy images. The Discriminative Dictionary Learning Segmentation is a classification technique which has been applied for different image processing problems such as compression, image denoising and recently in Magnetic Resonance Imaging segmentation. We consider the segmentation problem as a classification problem and apply Discriminative Dictionary Learning Segmentation to solve it using a patch-based representation and minimising the reconstruction error. The main limitation of this method is that the classification is performed independently for each voxel. We propose to add contextual information for the classification of the image voxels using Stacked Sequential Learning as a second stage. We define a feature vector from the classification results of Multi-class Discriminative Dictionary Learning and apply a decision tree classifier. We validate the proposal using a public database presented in the SATA Challenge. Using the two stages Stacked Sequential Multi-class Discriminative Dictionary Learning Segmentation method, we obtain an improvement of X% with respect to Multi-class Discriminative Dictionary Learning.