Hyperspectral Data Processing: Algorithm Design and Analysis

32.7 Conclusions

Spectral processing of multispectral MR images for partial volume estimation (PVE) is a new approach in MR image analysis where tissue substances are characterized by spectral information provided by MR image pulse sequences rather than spatial domain-based processing, which relies on spatial information provided by spatial correlation among sample pixels. As a result, spectral processing performs soft decisions on every pixel vector based on estimated abundance fractions of tissue substances to produce their partial volumes as opposed to spatial domain-based processing that makes hard decisions on every pixel vector based on class-labeling assignment. In order to materialize the concept of using spectral processing LSMA was first introduced by Wang et al. in a series of papers (Wang et al., 2000, 2001, 2003) in MR image classification where only unconstrained LSMA methods were investigated. This chapter extends their methods to two constrained LSMA methods and explores their utility in MR tissue characterization. Experimental results demonstrate that constrained LSMA generally performs better than unconstrained LSMA methods in terms of using unmixed results as PVE, specifically, tissue quantification, a task that cannot be accomplished by spatial domain-based processing techniques. However, it is the tissue quantification that can be used to calculate partial volumes of tissues, which are crucial in diagnosis of many diseases in progressive stages such as Alzheimer's ...