Classification Forests for Semantic Segmentation of Brain Lesions in Multi-channel MRI

Abstract:

Classification forests, as discussed in Chap. 4, present a series of advantages which make them a good choice for applications in medical image analysis. Classification forests are inherent multi-class classifiers (which allows for e.g. the simultaneous segmentation of different tissues), have good generalization properties (which is important as training data are often scarce in medical applications), and are able to deal with very high-dimensional feature spaces (which permits the use of long-range, context-rich features). In this chapter we demonstrate how classification forests can be used as a basic building block to develop state of the art systems for medical image analysis in two challenging applications. Given 3D multi-channel magnetic resonance images (MRI) as input we use forests for: (i) the tissue-specific segmentation of high-grade brain tumors (namely glioblastoma tumors), and (ii) the segmentation of multiple sclerosis (MS) lesions.