Tissue segmentation and classification of MRSI data using canonical correlation analysis

Abstract

Nowadays, MRSI represents a powerful non-invasive diagnostic tool. The ability of Magnetic Resonance Spectroscopy to detect metabolites is already very useful in daily radiologic practice since it provides significant biochemical information on the molecules of the organism under investigation. MRSI data can also be exploited in tissue segmentation techniques, which play a crucial role in many biomedical applications, such as the quantification of tissue volumes, localization of possible pathologies, improvement of pre-surgical diagnosis, and optimization of the surgical approach, therapy planning, etc. A variety of methods are available in the literature. They are often used in combination in order to solve different segmentation problems.They can be divided into several categories: thresholding techniques (1), region growing techniques (2), clustering techniques (3), Markov random field models (4), classifiers (5), artificial neural networks (6), etc. In this article, a fast and reliable tissue segmentation technique, based on a statistical method called Canonical Correlation Analysis (CCA), is proposed. This method is the multivariate variant of the ordinary correlation analysis and has already been successfully applied to functional Magnetic Resonance Imaging data in order to map sensor, motor, and cognitive functions to specific areas in the brain (7). Here CCA is adapted for MRSI data processing in order to detect possible homogeneous tissue regions, such as tumor regions, characterizing the considered sample. The goal is achieved by combining the spectral-spatial information provided by the MRSI data set and a signal subspace that models the spectrum of a characteristic tissue type, which may be present in the organ under investigation and, therefore, needs to be detected. More precisely, CCA quantifies the relationship between two sets of variables, magnitude spectra of the measured data and signal subspace, by means of correlation coefficients. These coefficients are afterward exploited in order to construct nosologic images (8) in which all the detected tissues are visualized. Such images can be easily interpreted by radiologists and physicians and, along with clinical and radiologic information, can improve the accuracy of the diagnosis.Extensive studies, performed on simulated as well as in vivo prostate MRSI data, were carried out in order to explore the properties of the proposed method. Moreover, the performance of CCA and ordinary correlation analysis was compared. The aforementioned studies show that CCA significantly outperforms ordinary correlation analysis in terms of accuracy and robustness.The article is organized as follows. In the Theory section, the basic definition of CCA is introduced and 3 possible implementations are outlined. In the Methods section, the application of CCA to MRSI data is described and the set up for the simulation studies and the acquisition environment of the in vivo studies is defined. In the Results and Discussion section, the results of the simulation and in viv...

The metabolic signature related to high plant growth rate in Arabidopsis thaliana

“…Canonical correlation analysis (CCA) is a multivariate technique often used in psychological, climate and ecological studies to quantify the associations between two separate data sets measured on the same experimental units (22)(23)(24)(25). In contrast to the aforementioned pairwise correlation analysis, CCA yielded a much stronger correlation of 0.73.…”

A hierarchical spectral clustering and nonlinear dimensionality reduction scheme for detection of prostate cancer from magnetic resonance spectroscopy (MRS)

“…20 and 23, z score was employed for MRS based detection of glioma and CaP, respectively. Another statistical technique, canonical correlation analysis ͑CCA͒, 24 based on calculating the canonical coefficients to obtain correlated linear relationships between two multidimensional variables was shown to be useful in successfully classifying prostate MRSI datasets into four classes: Aggressive tumor, tumor, mixed tissue, and healthy tissue.…”

NMF in MR Spectroscopy

“…Indeed, as described in the Introduction, MRSI provides biochemical information as well as spatial information about the sample under investigation and a natural assumption is that neighbour voxels contain similar tissue types. Previous studies carried out on other algorithms [41,42] already showed that the simultaneous exploitation of the two types of information significantly improves the performance of the considered methods. Hence, it is reasonable to expect similar improvements by introducing spatial information in the NMF-based algorithms as well.…”